Product Management for Engineers · Page 2

From Request to Problem: Framing, Jobs-to-be-Done and Opportunity Trees

A feature request is a proposed solution dressed as a requirement. This page covers the step before any solution: how a request becomes a one-story problem with a boundary, how a user segment is described by the job it hires a product for rather than by demographics, and how an opportunity solution tree organizes what interviews turn up. Afterwards you can rewrite a spec as a problem, state the job behind a segment, and lay out an outcome-opportunity-solution tree whose leaves are testable assumptions.

Deep dive 60 min read 12 sources As of 2026-10-26
The framing flip: a request becomes a problem story, then a cheaper solution, in Basecamp's permissions example Request as received What is going wrong? Problem story Solution, fitted Customer asks for more complex permission rules. It could easily take six weeks. Not "What could we build?" but "What is really going wrong?" Dig below face value. Someone archived a file, not knowing it would vanish for everyone else on the system. A warning on the archive action explains the impact. A one-day change. Cost, team-days (bar length proportional; five-day weeks assumed) As requested 6 weeks = 30 As framed 1 day = 1 Bounded by: a narrow problem, an appetite, explicit no-gos (Sections 2 and 3)
Figure 1. The flip from solution space to problem space, using Basecamp's own permissions example: a six-week build became a one-day change once the team asked what was going wrong1. The 30:1 bar ratio converts "six weeks" and "one day" to team-days by assuming five-day weeks; the source gives only the two durations and is a single anecdote, not a measured rate1.

1. A request is a solution in disguise

1.1 The flip, in Basecamp's own example

Shape Up, the method book of 37signals (Basecamp), states the move more cleanly than any other fetched source1. A customer asked for more complex permission rules, which "could easily have taken six weeks to build the change she wanted." Instead of taking the request at face value the team dug deeper and found that someone had archived a file without knowing it would disappear for everyone else on the system. A rule restricting who may archive was one answer; a warning on the archive action that explains the consequence was another. The book's verdict: "That's a one-day change instead of a six-week project"1 (Ch. 3, "Narrow down the problem").

The book names the operation: "we flip from asking 'What could we build?' to 'What's really going wrong?'" and follows it with the question that makes the flip operational: "At what point specifically does someone's current workflow break down without this thing they're asking for?"1

1.2 Why requests arrive as solutions

Two SVPG papers, twenty years apart, describe the same pattern. The 2005 PRD paper says it is "challenging to untangle the underlying problem to be solved, when all around you people are telling you essentially the solution they think they want"; the CEO, sales reps, engineers and customers are "all too happy to 'help'," asking for a button in a particular place, a feature because a competitor has it, or different colors2 (Step 3, p. 5). The cost it names is that "there are often much better ways of solving a problem" if the team is given the time and freedom to find them2.

The 2025 stakeholder paper turns this into a working rule: teams "reframe the many requests for features and projects (no matter where they originate) as problems to solve, along with a definition of success," because "our initial thoughts on how to best solve a problem are very often not able to deliver the necessary results. This is true no matter where the idea comes from, and no matter how smart the person may be"3 (Foundation, "Framing Work As Problems To Solve"). Stakeholders are asked to supply "the specific problem you need solved, who specifically you need that problem solved for, and how you will measure success"3, three fields that Section 2.3 reuses.

The engineer is on the other side of the same pattern. Gergely Orosz's first trait of product-minded engineers is that they "don't settle for getting a specification and jumping to implement it. They think about other ideas and approach the product manager with these"4. His list is observational, carries no data, and does not address infrastructure or platform teams4. The mechanism, in my reading, is positional: the requester sees a symptom and the existing product surface, so the nearest edit to that surface (a rule, a button, a column) is what gets proposed, while the implementer sees the solution space and is often the first able to say "there are three other ways to do this."

1.3 A test for a solution in disguise

Teresa Torres gives a crisp test, written for the nodes of an opportunity tree (Section 8) but equally applicable to requests: ask "Is there more than one way to address this opportunity?"5 Her example: "I want to go out to eat" admits only one way to satisfy it, so it is a solution; asking "Why do you want to go out to eat?" yields "I don't have time to cook" and "I want to eat right now," each of which has many solutions5.

Applied to engineering requests, the test sorts quickly. A hypothetical, with no figures and no source covering this runtime: a request reads "Add Q4 quantization support to the Hexagon backend." It names one mechanism (a quantization format) on one backend. Candidate underlying problems include that the target model does not fit in memory on a class of devices, that decode is too slow for an interactive feature, or that the customer's app ships its model in a format the runtime cannot load. A new quantization format is one resolution of the first two and irrelevant to the third. Until the story is known, the request cannot be estimated, ranked, or declined on its merits.

2. The problem as one story

2.1 A story is a test of fitness

Shape Up's pitch format has five ingredients (problem, appetite, solution, rabbit holes, no-gos), and the first rule about the first ingredient is that a problem and a solution travel together: "It's critical to always present both a problem and a solution together"1 (Ch. 6, "Ingredient 1. Problem"). The reason given is evaluative. Their example request, "Add tabs to the iPad app," might appeal to designers, but without a problem "there's no test of fitness to judge whether one solution is better than the other," and the discussion devolves into a debate about UI approaches1.

The format the book then prescribes for the problem is narrow: "The best problem definition consists of a single specific story that shows why the status quo doesn't work." The story is a baseline against which "people will be able to weigh the solution against this specific problem—or other solutions if a debate ensues—and judge whether or not that story has a better outcome with the new solution swapped in"1. A second benefit is separation of two arguments. A solution can be excellent while "the problem only happens to customers who are known to be a poor fit to the product"; the book's example is six weeks spent on "an ingenious solution that only benefits a small percentage of customers known to have low retention." A stated problem lets a team argue about demand separately from the quality of the fix1.

For an engineer the mapping is a test case. A requirements list states properties; a story is a scenario you can replay against each candidate to see whether the ending changes. This is my reading, not the book's, but it explains the insistence on one story rather than a summary of many: a scenario with specifics can be executed, an aggregate cannot. Two designs that both satisfy "make permissions more flexible" are indistinguishable until each is run against "someone archived a file and it vanished for everyone."

2.2 Ask when, not why: the calendar case

The most instructive fetched example of how a story is obtained is Basecamp's calendar request1 (Ch. 3 case study; Ch. 2 "Case study: The Dot Grid Calendar"). Customers kept asking for "a calendar." A proper one "can easily take six months or more," and earlier calendars had been used by "only about 10% of customers," so the appetite was one six-week cycle: "we could only build about a tenth of what people think of when they say 'calendar.' The question became: which tenth?" A requirements list cannot answer that. A story can.

So they called a customer who had asked and, "instead of asking her why she wants a calendar and what it should look like, we asked her when she wanted a calendar. What was she doing when the thought occurred to ask for it?"1 Reconstructed from the book: she worked in an office with a large calendar drawn on a chalkboard wall where officemates marked client meetings (the status quo); one day, working from home, a client called to schedule a meeting (the trigger); she drove to the office to read the wall, hit terrible traffic, and found no slot that worked (the break); she "could have saved an hour in traffic and a lot of frustration" had she been able to check open spots from home (the cost)1.

The insight the book draws: it "wasn't 'computerize the calendar'"; "'see free spaces' was the important thing for this use case, not 'do everything a calendar does.'" Basecamp's agenda view could list deadlines but "wasn't good for resource scheduling because you couldn't see empty spaces on it." The narrowed need, "help me see free spaces so I can figure out when to schedule something," led to the Dot Grid, a two-month read-only grid with a dot per event, no dragging, no multi-day spans, no categories1. The book is explicit about what the story did not give them: "We didn't have a solution yet. But now we felt like we had a problem that was specific enough to spark an idea that could fit within our appetite."

Teresa Torres reaches for the same unit of evidence from another direction. Her prompt is "Tell me about a time when…", because stories "naturally expose unmet needs, pain points, and desires" and show "when the opportunity occurred, what the customer was trying to do, and much more"5. Two practitioners with no shared organization agree that a specific past episode is the unit to collect, which is mild corroboration of the technique and not evidence of its effect: both are prescriptions without outcome data15. Interview technique itself is outside what the fetched sources let this page teach.

2.3 Anatomy of a framed problem

Two sources specify fields. SVPG's stakeholder paper asks for the problem, who has it, and the success measure3; Shape Up adds the story and the boundary (appetite and no-gos)1. Splitting the story into trigger, status quo, break and cost is my reading of the chalkboard narrative, not a field list the book prescribes.

Table 1. Fields of a framed problem, where each comes from, and one hypothetical on-device instance. The last column describes no real customer or measurement.
FieldQuestion it answersSource of the fieldHypothetical instance (illustration only)
TriggerWhen did the need appear, and what was the person doing?Calendar story, "when" question1; "when the opportunity occurred"5An app developer runs a feature on a device class for the first time and the model loads but responses are unusably late.
WhoWho specifically has it?"who specifically you need that problem solved for"3; fit of the customer1Developers integrating a local assistant into a shipping app, not hobbyists running a model once.
Status quoWhat do they do today, and where does it break?"why the status quo doesn't work"1They time the pipeline by hand, guess which stage dominates, and change settings by trial.
CostWhat did the break cost them?"an hour in traffic and a lot of frustration"1A feature slips a release because latency cannot be explained.
Success measureHow will we know it is fixed?"how you will measure success"3A developer in that situation can name the dominant stage without an external profiler.
BoundaryHow much will we spend, and what is excluded?Appetite and no-gos1One small batch; no new quantization formats; no changes to the model file format.

A success measure named at framing time is a hypothesis about what "fixed" looks like, not a number to defend. Section 8.2 gives Torres's way of separating an outcome's direction from its target; Page 4 covers how measures are chosen and trusted.

2.4 Grab-bags and the "2.0" label

When no single story anchors the idea, the framing has failed recognizably. Shape Up: "redesigns" or "refactorings" that "aren't driven by a single problem or use case" are "the worst offenders" among unclear ideas, and "Files 2.0" was the team's own mistake: "we didn't ask ourselves what specifically we were going to do. The project turned out to be a mess because we didn't know what 'done' looked like." They recovered by splitting it into "Better file previews" and "Custom folder colors," each with an appetite, and shipped those1 (Ch. 3).

The productive rewrite shows the repair. "Redesign the Files section" becomes "We need to rethink the Files section because sharing multiple files takes too many steps," which generates the questions that bound a project: "What's not working? In what context are there too many steps? What parts of the existing design can stay the same and what parts need to change?"1 The engineer's version of the test: if you cannot describe the finished state as a story ending differently, the work item is a grab-bag, however well the tickets under it are sized.

3. Boundaries: "good" is relative to constraints

3.1 What bounds a problem

Shape Up ends its framing chapter with a readiness condition: "a raw idea, an appetite, and a narrow problem definition" are the three things needed before a team defines the elements of a solution1 (Ch. 3, "Boundaries in place"). The appetite belongs to framing because it changes which conversation is possible: "The conversations we have are going to be entirely different if people think we're talking about a small improvement or a major redesign." It is "a time budget for a standard team size," in two sizes, a small batch (one designer and one or two programmers, one or two weeks) and a big batch (a full six-week cycle), and the pitch chapter calls it "another part of the problem definition"1 (Ch. 3; Ch. 6 "Ingredient 2. Appetite").

How an appetite differs from an estimate ("Estimates start with a design and end with a number. Appetites start with a number and end with a design"), how it compares with scoring methods, and what happens to unfunded ideas are Page 3's subject. For framing, what matters is the direction of dependency: the budget is an input to the design, and the problem story supplies the test the design must pass inside it.

3.2 The best solution is argmax under a constraint

The book's statement is blunt: "There's no absolute definition of 'the best' solution. The best is relative to your constraints. Without a time limit, there's always a better version. The ultimate meal might be a ten course dinner. But when you're hungry and in a hurry, a hot dog is perfect"1 (Ch. 3).

An engineer can read this as an optimization statement. Writing \( s \) for a candidate solution, \( \text{fit}(s \mid \text{story}) \) for how well it changes the ending of the problem story (Section 2.1), \( \text{cost}(s) \) for what it takes to build, and \( T \) for the appetite,

$$ s^{*}(T) \;=\; \arg\max_{s \,:\, \text{cost}(s) \le T} \ \text{fit}(s \mid \text{story}). $$

Change \( T \) and the argmax moves while the problem stays fixed. This is my formalization of the book's claim. It does not say a small budget produces a good solution; it says "good" is undefined without a budget, so a request without one is under-specified the way a benchmark without a hardware target is. The same budget governs how hard to look for the story: "Our appetite can also tell us how much research is worthwhile. If it's not critical now and we can't get our hands around the problem, we'll walk away from it and work on something else"1.

3.3 When is framing work worth its cost? A derived break-even

"How much research is worthwhile" has a simple expected-value form. The derivation is mine; the sources supply only the two durations in Figure 1 and the principle that the appetite sets the research budget1. Let \( B \) be the team-days to build the request as stated, \( s \) the team-days for the narrowed fix, \( c \) the team-days spent investigating, and \( p \) the probability that the investigation finds a narrower problem the narrow fix resolves; otherwise you build the original at \( B \). Then

$$ E[\text{cost with framing}] = c + p\,s + (1-p)\,B, \qquad \text{saving} = p\,(B-s) - c, \qquad p^{*} = \frac{c}{B-s}. $$

With Basecamp's durations as \( B = 30 \) and \( s = 1 \) team-days (six weeks and one day at five-day weeks, an assumption the source does not state), an investigation of \( c = 1 \) team-day breaks even at \( p^{*} = 1/29 \approx 3.4\% \); at \( c = 3 \) it needs \( p^{*} \approx 10.3\% \). For a two-week project (\( B = 10 \)) a one-day investigation needs \( p^{*} = 1/9 \approx 11.1\% \). Because \( c \) does not grow with \( B \) here, the larger the request the smaller the success probability the investigation needs.

Break-even probability that framing work pays for itself, as a function of investigation cost, for three build sizes 0% 10% 20% 30% 40% 50% 60% 0 1 2 3 4 5 investigation cost c (team-days) break-even probability p* (percent) B=10 B=30 B=60 B = 10: p* = c / 9 B = 30: p* = c / 29 B = 60: p* = c / 59 B = 30, c = 1: p* = 3.4% Above a line, framing saves time on average.
Figure 2. Derived, not measured: \( p^{*} = c/(B-s) \) with \( s = 1 \) team-day and no failed narrowing. \( B \) is the build cost of the request as stated, in team-days; \( B = 30 \) converts Basecamp's "six weeks" at five-day weeks1. The orange point is the worked example in the text. \( c \) and \( p \) are free parameters; no source supplies them.

Framing break-even (derived; all inputs are knobs, none is data)

Break-even p*
Expected saving
Saving as share of B

Formula: \( p^{*} = c / ((1-q)B - s) \) and saving \( = p\,((1-q)B - s) - c \). With \( q = 0 \) this reduces to the text above; a negative saving means the investigation does not pay at that \( p \).

Derivation with a narrowing that ships and fails

If the narrow fix ships but fails to resolve the problem with probability \( q \), after which you build the original at \( B \), the expected cost is \( c + p(s+qB) + (1-p)B \). Subtracting from \( B \) gives the saving \( p((1-q)B - s) - c \) and \( p^{*} = c/((1-q)B - s) \). With \( B = 30,\ s = 1,\ c = 1,\ q = 25\% \): \( p^{*} = 1/21.5 \approx 4.7\% \); at \( p = 0.2 \) the saving is \( 0.2 \times 21.5 - 1 = 3.3 \) team-days. A shipped miss also costs trust, which the model omits.

3.4 Two framing tests from the 2005 PRD

The ten-step PRD paper has two checks that apply at framing time2 (Step 2, pp. 2-3). The first is the elevator test: "If you had a chance to ride the elevator with the CEO of your company, and she asked what the point of your product was, could you answer that question before the ride is up?" One cause of failure it names is that "your product is trying to solve a non-problem; maybe you have a technology that you're still trying to find an application for"2. An engineer who has built a capable runtime should read that case twice: a technology looking for an application fails the framing test regardless of its merit.

The second check is that objectives be "prioritized" and come with "how they will be measured." The paper's example, "1) ease of use, 2) retail price under $100, and 3) compatibility with previous release," makes "ease of use" specific by the usability level the target user needs2. In engineering terms a prioritized objective list is a tie-breaking order for when two good designs conflict. Both tests are 2005 prescriptions without outcome data2.

4. Where problem inputs come from, and how each channel distorts them

4.1 Eight channels and what each does to the signal

A problem statement is only as good as the observations it was built from, and every channel transforms them on the way in. No fetched source ranks channels by reliability or measures a channel's error; they name specific distortions. Table 2 collects them and adds one column of my own, grouping the distortions into context loss (the why is stripped before the observation reaches you), solution substitution (the channel returns an answer in the solution space) and projection (the observer's attributes stand in for the user's).

Table 2. Problem-input channels, what each is useful for, the distortion the source names, and a distortion class assigned by the writer (the class column is not from the sources).
ChannelUseful forDistortion the source namesClass (mine)Source
Story-based customer interviewsA specific past episode: when the need appeared, what the person was doingTeams "overreact to the most recent customer interview"; Torres waits for three to four stories before mapping so one story does not dominatesampling (n is small)5
Sales conversations, support ticketsPointers to where to look next"Missing context"; "puts us at risk of addressing the wrong opportunity"; Torres uses them as inspiration for upcoming interviewscontext loss5
Stakeholders (executives, other functions)Business constraints a solution must satisfy: go-to-market, regulation, financials, partnershipsArrive as feature requests; first-guess solutions are "very often not able to deliver the necessary results," whoever proposes themsolution substitution3
CompetitorsHomework on alternatives and the team's own capabilities (Step 1)A feature requested "because a competitor has it" is a solution with its problem removedsolution substitution2
Asking people to specify the product (surveys, focus groups)Not claimed by the sourceA milkshake chain asked target-demographic customers to list an ideal milkshake's characteristics and responded to the feedback; "sales did not improve"solution substitution6
The team's own usage and intuitionFast hypotheses"You'll tend to assume that your customer is more like you than he really is"; opportunities generated "off the top of our heads" carry "biases and half-truths"projection25
Behavioral analytics and logsWhat users did, at scaleAmbiguous about cause: a rise in page views may mean a popular feature or a confusing interface; metrics "are not a substitute for early or formative user research"context loss7
Raw customer feature requestsWhere the pain is locatedName a mechanism, not a failure: "customers are asking for group notifications"; the default reply should be "Interesting. Maybe some day."solution substitution1

4.2 Three of the distortions in detail

Analytics: the "what" without the "why." The HEART paper, by Google user-experience researchers, is the clearest fetched statement of why a count is not an observation of a problem7 (§ "Related Work," "PULSE Metrics," "Adoption and Retention," "Task Success"). A rise in page views "may occur because the feature is genuinely popular, or because a confusing interface leads users to get lost in it, clicking around to figure out how to escape." In its worked case Google Finance saw "a surge in both page views and seven-day active users" during the September 2008 market meltdown, but the metrics "did not indicate whether the surge was driven by new users interested in the crisis, or existing users panic-checking their investments"; the team separated the groups with adoption and retention metrics. A second limit matters for problem discovery: "With web server log file data, it can be difficult to know which task the user was trying to accomplish."7 Page 4 covers the metrics themselves.

Surveys and focus groups: people answer the question you ask. In the HBS account the chain's marketing department "asked people who fit the demographic to list the characteristics of an ideal milkshake (thick, thin, chunky, smooth, fruity, chocolaty, etc.). The would-be customers answered as honestly as they could, and the company responded to the feedback. But alas, milkshake sales did not improve"6. The failure, as told, is not dishonesty: the question asked for attributes of the product, so the answers were attributes of the product, and none referred to the circumstance in which a product is chosen. It is one secondhand case; Section 6 treats how much weight it bears.

Your own usage: the projection problem and its inversion. The 2005 PRD paper writes profiles down as a projection defense: without a precisely characterized target user, "you will have this abstract notion and find it difficult to understand how your customers will react. You'll tend to assume that your customer is more like you than he really is," and "most of us are not anything like our target customers"2 (Step 3, pp. 4-6). For an engineer at a company whose users are engineers the claim inverts partially, because the team is closer to the user than a consumer-product team is. Whether the projection error shrinks, stays or merely changes form (the team knows internals the user never sees) is not established by any source I could fetch.

4.3 Broad channels point, narrow channels explain

Two sources already pair a broad channel with a narrow one. Torres treats sales and support input as raw material for interview topics, not as opportunities5; the HEART authors want metrics "triangulated with findings from other sources" and note that metrics are "primarily useful for evaluation of launched products"7. Together the pattern is a division of labor: analytics, tickets and sales conversations are wide and cheap and say where something is happening; a specific story from a person says what it is. Neither source states it that way, and both prescribe without outcome data. The wide channels are the ones an engineer can reach without asking anyone, and the temptation is to treat what they show as the problem rather than the place to look. Shape Up draws the budget consequence: if the team cannot find a specific pain point, "we'll walk away from it and work on something else"1. Inputs that never resolve into a story do not become projects.

5. Personas: the design archetype, read at full length

5.1 What the 2005 paper prescribes

SVPG's PRD white paper puts the user model in Step 3, "Define the User Profiles, Goals and Tasks," and the order inside the step is the content2 (pp. 3-6). First profiles: classify the user types, "determine who the primary users are," and characterize the important constituencies "in considerable detail." The profiles "are sometimes called 'personas' and while they should be fictitious, they should be as representative, realistic and plausible as you can make them. The idea is to come up with an archetype which captures the essence of this type of customer." Then goals, then tasks, and only then features: "features should be in support of required tasks that map to customer objectives"2.

The example is Leon the Power Seller, an eBay seller, and the declared use is decision support: "As we consider new features, we can ask ourselves what Leon's response to the feature would likely be." The attached discipline is selection: "Trying to please everyone is futile, and typically ends up pleasing no one"2.

5.2 What is actually in Leon

The jobs critique (Section 5.3) treats demographic profiles as merely correlational. Leon's paragraph lets you test how demographic a persona actually is. The paper's persona runs to nineteen statements by my count (another reader might split sentences differently); Table 3 sorts them2.

Table 3. The statements in the "Leon the Power Seller" persona, sorted by type (writer's count and classification of the paragraph in Step 3, p. 4). The last column is the writer's judgment.
Type of statementCount (of 19)What the persona saysPlausible link to a feature decision
Biographical646 years old; male; lives in Fresno; married with two teenage sons; owns two big Harleys; drives a 1993 Toyota PickupNone is stated; reads as color
Business context4Runs a small motorcycle parts business with a small shop; almost all sales come from eBay; wide range of items, saddle bags for Harley Davidsons the most popularDependence on the platform
Behavior and routine5About 400 items per month; seldom uses anything except eBay and e-mail; lists items Monday; most auctions end by Friday; ships within a few hours of paymentWorkflow constraints a selling feature must fit
Experience and attitude4Bought his computer to use eBay; three years of selling; a feedback rating over 5000 he takes pride in; finds changes to the selling process aggravatingSensitivity to changes in the selling flow

Six of nineteen statements are purely biographical; the other thirteen describe a business, a routine and a disposition toward change. The persona is closer to a thick description of someone in a working context than the "46-year-old male in Fresno" caricature suggests, and the paper's own warning, that the team will otherwise "assume that your customer is more like you than he really is," shows its purpose: a defense against projection, not a prediction of purchase from demographics2. It also lacks things: no trigger (what prompted Leon to list an item or to leave another channel), no named alternative he compares eBay against, and no account of the progress he seeks beyond "selling." The separate Goals step is where the paper expects the last of these2.

5.3 What the jobs critique hits, and what it does not

The critique is Christensen's, as reported by HBS. Segmenting a market by "product categories, such as function or price" or by "target demographics, such as age, gender, education, or income level" does not work well, he argues, because "the fact that you're 18 to 35 years old with a college degree does not cause you to buy a product. It may be correlated with the decision, but it doesn't cause it"6. That applies directly to the six biographical statements: age and income are covariates that move with purchase without being the variable that moves it. It does not apply to the routine, the platform dependence or the aggravation at change, which are circumstance and disposition.

What survives is a difference in unit of analysis. A persona indexes a type of person; a job indexes a situation that arises in a person's life. The Christensen Institute says so directly: "people 'hire' products or services when 'jobs' arise in their lives"8. In modeling terms the first clusters users and the second clusters episodes, and one user can contribute episodes to several clusters; the Institute says milkshake customers "were hiring milkshakes for different Jobs throughout the day," hunger and a more interesting drive in the morning and a treat for the kids in the afternoon8. A persona asks "what would Leon want?"; the jobs frame asks "in what circumstance does someone reach for this, and what else could they reach for?" Section 6 gives the case behind the second question and Section 7 shows the literature does not agree on how to answer it.

6. Jobs-to-be-done: the milkshake case and what it claims

6.1 The milkshake case, step by step

Christensen's teaching note "Integrating Around the Job to be Done" details the case; the HBS Working Knowledge article of 14 February 2011, by Carmen Nobel, is the fetched account6. The stepper below follows that account and adds nothing to it.

Observe The job Alternatives, response Marketing segmented by product and demographics, then asked people to list an ideal shake: thick, thin, chunky, fruity, chocolaty Result: sales did not improve A researcher spent a full day logging who bought, when, and whether they stayed to drink. 40% bought first thing in the morning, by commuters, to go (orange) Next morning he interviewed buyers leaving with a shake: what job did you hire it for? The job, as reported Long, boring commute Not hungry yet; will be by 10 Something to last until noon Constraints In a hurry Wearing work clothes At most one free hand Hired instead of a bagel or a doughnut: relatively tidy and appetite-quenching, and a thick liquid through a thin straw gives a hand a task Response Morning: thicker, with fruit chunks A separate job, a treat for young children: a thinner shake
Figure 3. The milkshake case as reported by HBS: segment-and-ask fails, observation finds a morning to-go cluster, interviews at the point of purchase yield the job, constraints and alternatives, and the product response6. The 40% share is the account's only figure and is single-sourced.
  1. Segment and ask. The chain divides its market by product and target demographic, asks people who fit the demographic to list an ideal milkshake's characteristics, and responds. Sales do not improve6.
  2. Watch first. A researcher spends a full day in one restaurant documenting who buys, when, and whether they drink on the premises. He finds that 40 percent of the milkshakes were purchased first thing in the morning, by commuters who ordered them to go6.
  3. Interview at the moment. The next morning he interviews customers leaving with a milkshake in hand, asking what job they hired it to do6.
  4. State the job, constraints and rivals. A long, boring commute; an extra hand to keep busy; not yet hungry but certain to be by 10 a.m., wanting to hold off hunger until noon; "in a hurry," "wearing work clothes," "(at most) one free hand." The milkshake was hired "in lieu of a bagel or doughnut because it was relatively tidy and appetite-quenching," and the thick liquid through a thin straw "gave customers something to do"6.
  5. Respond to each job separately. For the morning job, a milkshake "even thicker (to last through a long commute) and more interesting (with chunks of fruit)"; for a separate job, a treat for young children, a thinner milkshake so parents are not kept waiting6.

6.2 From correlation to circumstance

Christensen's stated mechanism, as reported: "We developed this idea because we wanted to understand what causes us to buy a product, not what's correlated with it. We realized that the causal mechanism behind a purchase is, 'Oh, I've got a job to be done'"6. The function-of-a-product view "really originates from your competitors or your own employees deciding what you need," whereas jobs-thinking "causes you to crawl into the skin of your customer and go with her as she goes about her day, always asking the question as she does something: Why did she do it that way?"6

Three details carry more engineering content than the slogan. First, Steps 2 and 3 run behavior first, question second: a day of observation located where and when to ask, and the interviews happened at the point of purchase, the broad-then-narrow pattern of Section 4.3. Second, the job statement contains constraints that filter the alternative set; the rival is whatever else passes the filter, a bagel or doughnut, not another milkshake. The Institute's summary agrees: "the competition wasn't other fast food milkshakes; rather, bananas and bagels vs. other kids' treats"8. Third, such findings are rare, in Christensen's account, because of how data is shaped: "All the data is organized by product category or customer category because that's easy to get. To go out and get data about a job is really hard"6.

The third point reads directly for anyone who instruments software, and I label it a reading: a telemetry schema keys events by account, plan, version, device and model, and none of those is a column for what the person was trying to accomplish. A job is therefore never in the event stream and has to be obtained by watching and asking. HEART's remark that web logs make it "difficult to know which task the user was trying to accomplish" is the nearest sourced statement of the same limit7.

6.3 A persona and a job, field by field

Side-by-side cards: fields of the Leon persona by type, and fields of the morning-milkshake job by type Persona: Leon the Power Seller Job: the morning milkshake 46, male, Fresno; married, two sons Two Harleys; a 1993 Toyota Pickup Parts shop; almost all sales via eBay About 400 items per month Lists Monday; auctions end by Friday Ships within hours of payment Feedback rating over 5000; proud of it Aggravated when the selling flow changes Long, boring commute Not hungry yet; hungry by 10 a.m. In a hurry Wearing work clothes At most one free hand Rivals: a bagel, a doughnut Progress: no hunger until noon Progress: something for the idle hand biographical (6) context, routine (9) attitude (4) circumstance constraint rival progress Missing: trigger, rivals Missing: who the buyer is, beyond "commuter"
Figure 4. What each form records. Left: the Leon persona's statements, grouped by the writer into the Table 3 types (counts in parentheses are of 19)2. Right: the fields of the HBS milkshake account6. The persona records a person and a routine; the job records a situation, the limits on what can be done in it, and what else could be hired.

The figure declares no winner. The persona's thirteen non-biographical statements are the facts a team needs to design a selling flow; the job card holds what the persona lacks, namely the triggering circumstance, the constraints on what is feasible in it, and the rivals. Designing an interaction for an existing population favors the persona's fields; deciding what to build and against what you compete favors the job's.

6.4 A second case, and what the cases have in common

The Christensen Institute's page carries a second case from Bob Moesta: a Detroit-area condominium builder targeting retirees and divorced single parents with high-end finishes. Foot traffic was high and conversions low; focus groups suggested bay windows, which were added, and "sales didn't improve." Moesta asked buyers what job they had hired the condominium to do. Prospects said they did not need a formal dining room, yet buyers kept bringing up the dining room table: "As soon as I figured out what to do with my dining room table, then I was free to move." He hypothesized that what held buyers back was "the anxiety that came with giving up something that had profound meaning." The builder enlarged dining rooms and added moving services, storage and a sorting room. The Institute reports that prices rose by $3,500 and that in 2007, "when industry sales were down 49%, the developers had actually grown business by 25%"8.

The structure matches the milkshake case: segment by demographic and ask for preferences, see no improvement, then interview people at or after the decision about the circumstance and find a cause the preference question could not reach. The jobs differ in composition: the milkshake job is built from circumstance and constraint, the condominium job from anxiety, a force that holds a decision back. The Institute's definition covers both: Jobs to Be Done "reveals the circumstances—or forces—that drive people and organizations toward and away from decisions," and a job is "the progress they're trying to make as they strive toward a goal or aspiration within particular circumstances"8.

An inconsistency inside the proponent literature is informative. The Institute glosses the milkshake job as "to satiate hunger, make their drive more interesting, and easily hold food in the car" and elsewhere as "Help me relieve stress with an easy-to-hold snack"; the HBS account words it as a boring commute, an idle hand and hunger to be staved off until noon68. Three phrasings from two proponent accounts suggest the job statement is an interpretation the analyst writes, not a field the interviewees supply.

6.5 What the evidence is

  • The milkshake facts (40% morning share, the commute constraints) are one secondhand report of a teaching note; the Institute's retelling gives no share, so the 40% has no second fetched source68.
  • The condominium figures ($3,500; industry sales down 49%; developer up 25%) are the Institute's telling of a consulting engagement, with no independent source in the map8.
  • Procter & Gamble. HBS says P&G's product success rate "rose dramatically" after it segmented by a product's job, "Christensen says," with no figure, period or definition of success6.
  • The book. The publisher's description states the thesis: "Customers don't buy products or services; they 'hire' them to do a job," and "Understanding customers does not drive innovation success, he argues. Understanding customer jobs does." That shows what the book claims. Its text was not fetched, so its mechanism, evidence and counter-cases are unread, and a blurb is not evidence that the framework works9.

An explanation built after the fact from interviews is easy to make coherent, and each case above is one. The stronger test is a prediction made before the change and checked after; none of the fetched accounts reports one. That is a gap in what this page can pass on, not a finding that the framing is wrong. The cases are also consistent with the simpler lesson of Sections 1 and 2: stated preferences about product attributes are a weak input and a specific episode is a strong one.

7. What counts as a "job": two schools that share a slogan

"Jobs to be done" names at least two practices, and each school's institution publishes its own definition: the Christensen line carried by Bob Moesta and the Christensen Institute8, and Tony Ulwick's Outcome-Driven Innovation, published by his firm Strategyn10. The second source is a consultancy page, so every figure on it is a vendor claim; I use it for what the school says it does, not for whether it works.

7.1 Definitions and methods

Strategyn defines a job as "the task, goal, or objective a customer is trying to accomplish in a given situation" and tells the analyst to "define the core functional job in a single, solution-free statement," for example "cut a piece of wood in a straight line"10. Its core claims are that "jobs are stable; products are not" (listening to music on the go persisted through records, cassettes, CDs, MP3 players and streaming) and that needs are measurable: "desired outcomes" are "the metrics customers use to measure success when getting a job done," and "a single market typically contains 50 to 150" of them. The method is staged: qualitative research and a "Universal Job Map" of eight steps (define, locate, prepare, confirm, execute, monitor, modify, conclude) to enumerate outcomes; a survey of "a statistically valid population" rating each outcome's importance and current satisfaction, which "reveals which needs are underserved (unmet) and which are overserved"; then outcome-based segmentation10.

The Institute's version centers on circumstance and force. Jobs to Be Done is "a lens that reveals the circumstances—or forces—that drive people and organizations toward and away from decisions" across "functional, social, and emotional" dimensions. The method visible on the page is to ask people who made a decision what they were trying to accomplish, as in both cases above. Its college example scales that: Michael Horn and Bob Moesta "drew on more than 200 detailed stories of students making the decision to attend some form of postsecondary education" and arrived at five jobs with their own "push and pull" forces8. The page describes neither an interview protocol nor a quantification step.

Table 4. Two jobs-to-be-done practices as each publishes itself. "Not stated" means not on the fetched page. Strategyn is a vendor page; none of its figures is independently verified here810.
DimensionChristensen / Moesta (Institute page, HBS account)Ulwick / Strategyn (vendor page)
Definition of a job"the progress they're trying to make as they strive toward a goal or aspiration within particular circumstances"8"the task, goal, or objective a customer is trying to accomplish in a given situation"10
Unit of outputStories of decisions; functional, social and emotional forces50 to 150 measurable desired outcomes per market
QuantificationNot stated; counts of stories only (200+ in the college example)Survey of importance and satisfaction per outcome
SegmentationBy job and circumstance, not by demographics or product categoryBy outcome; segments "cannot be found through demographics, psychographics, or behavioral data alone"
Outcome evidence on the pageTwo cases told by proponentsCompany cases and a stated 86% success rate against a "roughly 17%" industry average; no method given

7.2 Is the commute the job or the context?

A reported dispute is whether the milkshake's long boring commute is the job or its context. The fetched Strategyn page never mentions the milkshake or contrasts itself with Christensen; its comparisons are with design thinking and lean startup, and it credits Ulwick with introducing the concept to Christensen10. A secondhand blog account says Ulwick reads the milkshake's commute as context and the job as getting breakfast on the go. What is verifiable from the two definitions is narrower. Both place the situation inside the definition ("in a given situation"; "within particular circumstances"). They differ in what the analyst writes down afterward: Strategyn strips out the solution, keeps a functional statement stable across technologies, and measures outcomes against it; the Institute keeps the story and its forces. For a team deciding what to build, that is a difference in output format: a stable statement plus a rated outcome list can be compared across segments and over time, while a story carries the context that explains why a rating would be low.

7.3 Reading the schools as measurement designs

Both schools hide a measurement design. A survey of importance and satisfaction can only rate outcomes someone listed first, so its precision is conditional on the completeness of the qualitative step, and the "Opportunity Algorithm" the page names is not defined there10. A story-based approach may surface the unexpected cause (the dining table), but its stories are few, selected by whoever agreed to be interviewed, and the output is an interpretation. Neither source quantifies its own error, so nothing here ranks the two instruments, whose failure modes are complementary.

8. Opportunity solution trees: organizing what the stories say

8.1 Structure and the type check at each level

Teresa Torres's Product Talk page defines an opportunity solution tree as "a simple way of visually representing the paths you might take to reach a desired outcome." The root is the outcome, "the business need that reflects how your team can create business value." Below it sit the opportunity space ("customer needs, pain points, and desires that, if addressed, will drive your desired outcome"), the solution space, and assumption tests, "how we'll evaluate which solutions will help us best create customer value in a way that drives business value." It is built for a product trio of product manager, designer and software engineer, and the page dates the tree to 20165.

For an engineer the useful reading is an explicit search tree with a type check per level: the outcome decomposes into opportunities, each admits several solutions, each solution carries falsifiable assumptions. The page supplies a check for each level (Table 5). A flat backlog hides which opportunity an item serves and which alternative was never considered; the tree makes unexplored branches visible.

Table 5. The four levels of an opportunity solution tree, what belongs at each, and the page's own check for a misplaced item. Source: Product Talk5; the 22-to-25% outcome is Torres's illustrative example, not data.
LevelWhat belongs hereCheck for a misplaced itemExample on the page
Outcome (root)A change the team can influence: a business outcome (financial metric), a product outcome (customer behavior or sentiment in the product) or a traction metric (adoption of one feature). A product outcome is preferred.Business outcome: teams "don't always know how to impact this metric." Traction metric: scope "often too narrow" for discovery."Increase the percentage of first-time users who get to our aha moment during their first session from 22 to 25%"
Opportunity spaceCustomer needs, pain points and desires, drawn from interview stories"Is there more than one way to address this opportunity?" If not, it is a solution"I don't have time to cook dinner"
Solution spaceA product, feature, service, workflow, process or documentation offered to address one opportunityConsider more than one solution for the target opportunity"Order takeout"; "Subscribe to a meal prep service"
Assumption testsThe assumptions each solution depends on, riskiest firstTest the riskiest assumption of each candidate, then compareNot worked on the page
A hypothetical opportunity solution tree for a developer-facing on-device runtime: one outcome, three opportunities, three solutions under the target opportunity, and an assumption test under each solution OUTCOME (product outcome, direction only) More first-time integrators get a feature running on-device in their first session I cannot tell why it is slow on this device My model's format is not loadable I do not know which devices it will fit Per-stage timing report Guided "why slow?" page Auto-tuned default config Feasibility: do timing hooks stay cheap? Value: do developers follow the page? Viability: who keeps the defaults current? Solution space: several candidates for ONE target Assumption tests: riskiest assumption of each candidate HYPOTHETICAL: structure from Product Talk; contents invented for illustration. No figures; no source covers a runtime. Orange = target opportunity.
Figure 5. Skeleton of an opportunity solution tree5 filled with a hypothetical instance for a developer-facing on-device runtime. Only the target opportunity (orange) is expanded into solutions, and every solution carries a test of its riskiest assumption. The labels on the tests use the value, feasibility and viability vocabulary of Page 1; that mapping is the writer's.

8.2 The root: which kind of outcome, and a tension on the page

Torres's book distinguishes three outcome types, which the page restates: business outcomes "measure the health of the business and are typically financial metrics like grow revenue, increase market share, or reduce churn"; product outcomes "typically measure a customer behavior in the product or a customer sentiment about the product"; traction metrics "measure adoption of a single feature." The preference is about control: "When we assign a team a business outcome, they don't always know how to impact this metric, whereas product outcomes are directly within their control," and traction metrics "aren't good discovery outcomes because their scope is often too narrow"5.

An outcome has two parts. In Torres's example, "increase the percentage of first-time users who get to our aha moment during their first session from 22 to 25%," the direction is the outcome and 25% the target; a new tree is needed only when the direction changes. Outcomes are one per team, split where several teams share one "by customer segment" or "by customer journey step"5. The 22% and 25% are the author's hypothetical, not a measured rate.

The page also contains a tension, which I report without resolving. It recommends a product outcome at the root. Yet asked "Why does the outcome focus on business value and not customer value?" it says "An opportunity solution tree starts with business value (represented by the outcome) because when we create business value, we earn the right to serve our customer over time," citing Google Reader, which Google shut down "because it didn't create enough value for Google," and Dark Sky, bought and shuttered by Apple5. One reading: the product outcome stands in for a business outcome the team cannot move directly. Another: business value is the reason the tree exists and the root is the nearest controllable quantity. The page does not say how the link from product outcome to business outcome is checked, and gives no example of a product outcome that failed to move the business metric. That link is a hypothesis embedded in the root, exposed to the same base-rate problem as any idea (Page 1).

8.3 The opportunity space, built from stories

The page insists that opportunities come from interviews: "When we generate opportunities off the top of our heads, we bring our own biases and half-truths into the picture," whereas limiting the space to what is heard in interviews "guarantee[s] that we are working on real unmet customers needs, pain points, and desires." "Guarantee" overstates what three or four conversations establish: a story is a real episode, but whether the need recurs across customers is a sampling question the page does not address. The prerequisites include "three to four story-based customer interviews before starting your first opportunity solution tree"; without good inputs, trees are "nearly impossible to create"5.

Structuring is "a messy process of synthesis" in four steps: build an experience map of all the stories, map its moments to top-level opportunities, group each customer's individual opportunities under the moment they occurred, and structure each branch with parent-child and sibling relationships. The top-level moments "shouldn't change too often." The revision rule is "for every three to four interviews that you conduct, you want to revise and refine the opportunity space": at one interview a week, a revision every three to four weeks, about three to four in a 13-week quarter (writer's arithmetic from the page's cadence). Segments sit at the top of a tree only when the team's remit spans audiences, such as a marketplace's two sides or an electronic medical record used by physicians, nurses and office administrators5.

They are called opportunities rather than problems because products also serve desires: "Disneyland entertains me; ice cream is delicious; mountain biking is fun... These products were designed to address my desires, not solve my problems"5. Compared with the problem story of Section 2, an opportunity is deliberately one level up: a story is one episode, an opportunity a need that several episodes share.

8.4 Choosing a target opportunity, and why effort is left out

Torres wants one small opportunity at a time, citing kanban's limit on work in progress, because that is what makes multiple solutions explorable. Choose the target "as soon as you have a first draft of the opportunity space," with a "crummy first draft" mindset, because "this is where most of the learning in discovery happens." She names four assessment factors from her book's Chapter 6: opportunity sizing, market factors, company factors, and customer factors ("how important is each opportunity to your customers and how satisfied are they with existing solutions")5. The last pairs importance with satisfaction, the pair the Strategyn survey rates per outcome; the Product Talk page does not cite Ulwick, so the resemblance is only mine to note.

Effort is excluded on purpose: "Solutions take effort. But we aren't exploring solutions yet. Any given opportunity can have both easy and difficult solutions." The failure she wants to avoid is that "many teams assume that an opportunity will be too hard to address, and then once they choose it, they uncover a dead-simple solution"5. That is the logic of Section 3.3 applied to the tree: do not let a guess about build cost veto a need before a cheap solution has been sought. Page 3 covers scoring methods that divide by effort; no fetched source tests which treatment of effort chooses better.

8.5 Solutions, assumption tests and "ready to build"

The page defines a solution broadly ("a product, a feature, a service, a workflow, a process, documentation, or anything else that we offer to customers") and prefers small solutions matched to one opportunity each. It recommends "more than one solution for their target opportunity," citing "decision-making research" that says comparing options yields better decisions (no study is cited), and concedes "we don't always have time to do this": compare "when there's risk in the solution, when the opportunity is a differentiator, or you need to uncover an innovative solution"5.

After choosing a target: brainstorm solutions for that opportunity only, choose three, "break your solutions down into their underlying assumptions and prioritize which assumptions to test first," test the riskiest of each, then discard, refine one, or abandon the opportunity5. Readiness to build depends on three questions: "How much risk does the solution have? How much risk can our organization bear? How much time do we have to make a decision?" And: "Whenever we put a solution in our delivery backlog, we are making a bet. Discovery helps us reduce the risk of that bet." Because "when we are new to discovery, we aren't very good at assessing risk," do a little more discovery than you think you need5.

The page offers no taxonomy of assumptions. Page 1's four risks (value, usability, feasibility, business viability) are the natural one, and SVPG describes prototypes as "testing these prototypes against these four risks"11; the mapping is the writer's, not Torres's. For the engineer on the trio, feasibility assumptions are the ones he can name sharply and often test cheapest; Figure 5 shows one on each solution because feasibility is the engineer's native risk, not because the page requires it. The remaining assumptions, such as whether developers follow a guided page or who maintains tuned defaults, are the ones he is still expected to notice.

8.6 Limits the page states, and the ones it does not

  • Garbage in. A tree built from imagined opportunities contains no information; the page repeats the prerequisites ("I know I'm starting to sound like a broken record")5.
  • One tree per trio. On a company-wide tree: "I don't recommend this. I've seen many companies try and it never turns out well." For a company-wide view she suggests a KPI tree and a company-wide experience map5.
  • Sales and support are inspiration. They feed interview topics and do not add leaves unless confirmed (Section 4.1)5.
  • Commercial position. The author discloses "I also license AI services to Vistaly," recommends it if a tree "continues to be a mess after six to eight interviews," and Product Talk sells courses on the method. That does not make the advice wrong, but the page is prescription by the vendor of the training, and its listed benefits ("faster learning cycles," "simpler stakeholder management") are asserted, not measured5.
  • No outcome data. The page compares no teams with and without trees; "there are no black and white answers in discovery"5.

9. What a product-minded engineer does with a spec he is handed

A spec arrives as a bundle of mechanism, scope, deadline, acceptance criteria and edge cases, and each presupposes something the spec does not state. Reading it as an engineer who thinks about product means reconstructing the presupposed parts. Gergely Orosz's nine observed traits are the fetched account of this behavior; they rest on his observation, report no data on any trait or career outcome, and do not address infrastructure or platform teams4.

Table 6. What the parts of a spec presuppose, the question that recovers it, and the source for the move. The pairing of parts to questions is the writer's composite; no source lists these rows together.
Part of the specWhat it presupposesQuestion to ask backSource for the move
A named mechanism ("add a rule," "support format X")A problem story that this mechanism resolves"When did this come up, and what was the person doing?" Is there more than one way to address it?Shape Up's "when" question1; Torres's test5; Orosz trait 14
A scope listAn unstated appetite and set of no-gosHow much is this worth to us, and what is excluded?Appetite and no-gos1
A deadlineAn outcome that depends on the date, with discovery done to back itWhat happens if it ships later? Is this a commitment someone has paid discovery work for?High-integrity commitments "come at a cost"3 (mechanics on Page 3)
Acceptance criteriaA success measure and a person who will look at it afterwardHow will we know it worked, and who checks the result?"how you will measure success"3; Orosz trait 84
A list of edge casesThat each must be handled in codeWhat happens if we do nothing for this case? Can support or a retry cover it?Orosz trait 64

Two traits describe mechanisms worth stating. Trait 5: where many engineers cut effort and then work out what that does to the feature, product-minded ones "attack this problem from both angles," often returning to the product manager "suggesting a completely different feature to be built, given the product impact would be similar, but the engineering effort vastly smaller"4. That is the Section 3.2 argmax in practice: the engineer sees the cost surface and can propose a different point on it. Trait 6 gives a worked example: "if one in a thousand users might be hit by an error, they will consider the effort to fix it and think about what happens if they don't do anything. Can customer support help the person in this case, during validation? Can the user just retry and succeed the next time? Can the product be slightly modified, so this edge case won't occur?"4

The social cost is acknowledged. Such engineers "ask many questions, doing this frequently," and "manage not to annoy people, as they've built up strong relationships with them," finding "answers they can, by themselves" first4. The questions in Table 6 spend relationship capital built beforehand, by understanding how the company makes money, working with the product manager, and bringing "well-backed product suggestions" with both effort sides stated4. SVPG's stakeholder paper frames the exchange from the other side: the requester supplies the problem, the person and the success measure, and the product team is "empowered and accountable to discover a solution"3. Without that empowerment, asking is a favor, not a role. SVPG's 2026 FAQ adds the opposite caution for product managers, which transfers to any engineer who answers every spec with questions: many or even most requested ideas "won't actually solve the problem," yet "your job is not to be the person to say no, but rather be an integral part of the product team that can get to a solution that actually works"11.

10. Framing a developer-facing runtime: where the evidence ends

None of the sources on this page frames a runtime that other developers integrate. Orosz's list is silent on infrastructure and platform teams4; Torres's examples are dinner, streaming, email, a first-session "aha moment," a marketplace and a medical record5; the milkshake and condominium cases are consumer purchases68. The one fetched statement naming developers as customers is a sentence on the Strategyn page about business-to-business roles10. This section is therefore a labeled hypothetical: the mechanics of Sections 2 to 8 applied to a plausible structure, with unsupported steps marked.

10.1 Who is "the user" when the customer is a developer?

Strategyn separates three customer roles that collapse into one person in consumer markets: the job executor, "the primary user of the product, the person who employs it to accomplish the core functional job"; the product lifecycle support team, "those who install, transport, repair, maintain, upgrade, or dispose of the product"; and the buyer, "the individual or group responsible for making the financial purchase decision." In business-to-business markets they are distributed: "developers use it, IT staff maintain it, and C-level executives approve its purchase"10. Torres puts customer segments at the top of the tree when a team's remit spans audiences5. Table 7 applies the three roles to an on-device runtime; the occupants and jobs are my hypotheses, not findings.

Table 7. The three Strategyn customer roles10 mapped to a hypothetical developer-facing on-device runtime. The role names are from the source; every entry in the other columns is an illustrative hypothesis.
RoleCandidate occupant (hypothetical)Job they might hire the runtime for (hypothetical)How the team could see them (hypothetical)
Job executorThe developer integrating the runtime into an appGet a model's output into a feature at a latency the app's users accept, on the devices the app supportsIntegration attempts, issues, questions: the Table 2 channels
Lifecycle supportWhoever ships, updates and debugs the app in the fieldKeep the feature working across OS, device and model updatesCrash reports, regressions after an update
BuyerWhoever approves the dependency or its licenseAdopt a component without adding unbounded risk to the productProcurement conversations; rarely visible in usage
Outside the frameworkThe end user of the downstream appNone with the runtime; never sees itOnly through the developer

The last row is the structurally new element. In the milkshake case the constraints (hurry, work clothes, one free hand) define the job; for a runtime the end user's circumstances (offline, battery, a hurried glance at a phone) plausibly become constraints in the integrating developer's job, inherited rather than observed. If so, the team needs a story two hops deep, and the Section 4 distortions apply at each hop: the developer is a reporter of the end user's circumstances, not the user.

10.2 What transfers, and what has been tested only on other products

Table 8. Mechanisms from Sections 2 to 8, the evidence the sources give, and the status for a developer-facing runtime. "Applied" means the writer is extending a result beyond the domain of its source.
MechanismEvidence in the sourcesStatus for a developer-facing runtime
Problem as one story; ask "when"Basecamp anecdotes1; Torres's story-based interviews5; prescriptions without outcome dataApplied; untested
Break-even for framing workThe writer's derivation from Shape Up's principle1The algebra transfers; \( B, s, c, p \) for a runtime are unknown
Channel distortions, personas, jobs, treesConsumer and SaaS examples: HEART, PRD, HBS, Institute, Product Talk25678; Strategyn's business-to-business roles are a vendor claim10Applied; no developer persona, developer input channel or developer-product tree appears in the sources

The table is the honest boundary of this page. Figure 5 borrows her first-session activation outcome; whether that outcome is meaningful when integration spans days or weeks is untested. For an engineer on a runtime team the transferable core needs the least evidence to defend: a request is a solution until a story says otherwise; the budget governs how hard to look; each input channel distorts in a known direction; a tree makes unexplored branches visible. The parts that need evidence, namely that jobs beat personas for developer products, that trees raise the hit rate, or that framing time pays back at a given \( p \), are open in the sources.

Where experts disagree

Two questions are bundled under "understanding users" and the sources answer them differently. The first is what to write down about a user: a person-type (persona), a situation and its forces (Christensen and Moesta's jobs), or a set of measured outcomes (Ulwick's jobs). The second is how much effort the problem deserves relative to the solution at all. The first is the logged disagreement for this page; the second surfaced while reading SVPG's 2026 FAQ against Shape Up and Product Talk. None of the four positions below is backed by outcome data in the fetched sources, so the comparison is of mechanisms and costs.

Personas as design archetypes

Marty Cagan, SVPG PRD white paper, 20052

Characterize the few most important user types in detail as fictitious but realistic archetypes, then derive goals, tasks and features from them, so the team stops assuming the customer resembles itself.

  • Cheap, and a shared reference for design decisions ("what would Leon do?").
  • Thirteen of the nineteen statements in the example are routine, context or disposition, not demographics (Table 3).
  • Indexes a person-type, not a situation; no trigger and no named rivals in the example.
  • Biographical attributes correlate with, but do not cause, a purchase6.
  • A 2005 prescription without outcome data.

Jobs as circumstance, learned from decision stories

Clayton Christensen, Bob Moesta and the Christensen Institute; HBS report of the milkshake case68; the 2016 book's thesis, publisher's wording9

A purchase is caused by a job arising in a circumstance, so segment and interview by the job. Customers "hire" a product for the progress they want; the rivals are whatever else fits the same constraints.

  • Surfaces causes that preference questions miss, in two cases.
  • Names constraints and rivals, which a persona does not.
  • Two proponent-told cases; the 40% share is single-sourced.
  • The job statement is an analyst's interpretation (three phrasings of one job).
  • No quantification step; the book's text was not read.

Jobs as solution-free statements with measured outcomes

Tony Ulwick, Strategyn (vendor page, © 2025)10

State the core functional job without a solution, enumerate 50 to 150 desired outcomes, survey importance and satisfaction, and segment by unmet outcome.

  • Comparable numbers across segments and over time.
  • Separates job executor, lifecycle support and buyer, which matters in business-to-business products.
  • The headline figures (86% against roughly 17%) come with no method.
  • A survey can only rate outcomes someone listed; the "Opportunity Algorithm" is not defined on the page.

Understanding the problem is the cheap part

Marty Cagan, "Build To Learn FAQ," 2026-04-2711

Strategy picks the problem. Understanding it "is not difficult and doesn't take us much time," misunderstandings "we will see this very quickly once we start testing prototype solutions," and "the hardest part, by far, is almost always solving the problem." Treating discovery as confirming that the problem is real is the wrong job.

  • Matches the economics of cheap prototypes, which also underlie Andrew Ng's argument12.
  • Guards against framing work that never ends.
  • Conflicts with Basecamp's anecdotes of a six-week build becoming a one-day change1.
  • Presumes a problem chosen well by strategy and prototypes cheap enough to test.
  • A failed prototype shows a frame was wrong, not what the right frame is. No evidence is offered.
Assessment

Persona or job. The two answer different questions and can be used together. Persona-style fields (routine, business context, disposition toward change) are what a team needs to design an interaction for a population it already serves; job-style fields (circumstance, constraints, rivals) are what it needs to decide whether to build and against what it competes (Figure 4). Between the job schools, stories come first because they are cheap and expose the unexpected cause; an outcome survey earns its cost only if the decision is large and the outcome list is believed complete. A head-to-head trial would settle it: teams given the same decision, using personas, job interviews or an outcome survey, scored on whether the change moves a pre-registered metric, with Page 1's base rates as the yardstick. No fetched source reports one.

How much problem work. The break-even model of Section 3.3 reconciles Cagan with Basecamp. Framing pays when a build is expensive relative to the investigation (large \( B \), small \( c \)); Cagan's claim holds where prototypes are cheap, and Basecamp's examples apply where a six-week build was the alternative. My judgment, not a sourced finding: for on-device inference work, where a new format or kernel path is a multi-week build, \( B \) is large, so the weight shifts toward Shape Up and Product Talk. Time-to-correct-frame data from comparable teams working prototype-first against story-first would settle it; the map has none.

What is changing

Two statements point the same way, and neither is data. Andrew Ng wrote on 2025-01-15 that "writing software, especially prototypes, is becoming cheaper. This will lead to increased demand for people who can decide what to build," arguing from complements: when one good gets cheaper, demand for the other rises, as with cars and gasoline12. SVPG's 2026-04-27 FAQ restates that "as the cost of product delivery continues to drop, the bottleneck and our competitive advantage moves to product discovery"11. Ng adds a line that matters to an engineer: "Perhaps engineers will step in to do some of this work"12. Treat both forecasts as likely to move within a year.

Section 3.3's model adds a caution. Shrinking \( B \) raises \( p^{*} \): at \( B = 5 \), \( s = 1 \), \( c = 1 \), \( p^{*} = 1/(5-1) = 25\% \), against 3.4% at \( B = 30 \). On cost grounds alone, cheaper builds make up-front investigation less valuable. Both sources claim the scarce resource shifts to choosing what to build, which a cost-only model does not price; the claim is unmeasured. SVPG adds a restraint that cheap building makes more relevant: "it's very easy for even the most loyal customers to start feeling abused because of constant, erratic change," so rapid testing belongs with opt-in users and prototypes11.

References

  1. officialSinger, R. (37signals / Basecamp). Shape Up: Stop Running in Circles and Ship Work that Matters. Web edition (page states only "Copyright ©1999-2026 37signals LLC"). https://basecamp.com/shapeup. Accessed 2026-10-26. Used for: Ch. 3 "Set Boundaries" (/shapeup/1.2-chapter-03: permissions example, "What's really going wrong?", "good is relative", appetite, grab-bags, boundaries); Ch. 6 "Write the Pitch" (/shapeup/1.5-chapter-06: Ingredient 1, the single specific story, test of fitness); Ch. 2 case study "The Dot Grid Calendar" (/shapeup/1.1-chapter-02). Practitioner assertion and anecdote; no measured data.
  2. blogCagan, M. How To Write a Good PRD. SVPG white paper, 26 pp., © 2005 (PDF re-hosted on svpg.com 2024-07). https://www.svpg.com/wp-content/uploads/2024/07/How-To-Write-a-Good-PRD.pdf. Accessed 2026-10-26. Used for: Step 2 (elevator test, prioritized and measurable objectives, pp. 2-3); Step 3 (profiles, the Leon persona p. 4, goals, tasks, pp. 3-6); Step 1 (competitors as homework).
  3. blogJones, C. and Cagan, M. Stakeholders and the Product Model. Silicon Valley Product Group, 2025-12-01. https://www.svpg.com/stakeholders-and-the-product-model/. Accessed 2026-10-26. Used for: "Framing Work As Problems To Solve" (reframing requests; problem, who, success measure); "Sharing Business Context"; high-integrity commitments.
  4. blogOrosz, G. The Product-Minded Software Engineer. The Pragmatic Engineer, 2019-09-05 (last updated 2022-02-01). https://blog.pragmaticengineer.com/the-product-minded-engineer/. Accessed 2026-10-26. Used for: traits 1, 5, 6, 8 and the "tips to become a more product-minded engineer"; the page's silence on infrastructure and platform teams.
  5. blogTorres, T. Opportunity Solution Trees: Visualize Your Discovery to Stay Aligned and Drive Outcomes. Product Talk (page undated; author discloses a commercial tie to Vistaly). https://www.producttalk.org/opportunity-solution-tree/. Accessed 2026-10-26. Used for: tree structure and levels, outcome types and the 22-to-25% example, opportunity vs solution test, story-based interviews, revision cadence, target-opportunity factors, assumption tests, limits.
  6. analysisNobel, C. Clay Christensen's Milkshake Marketing. Harvard Business School Working Knowledge, 2011-02-14. https://www.library.hbs.edu/working-knowledge/clay-christensens-milkshake-marketing. Accessed 2026-10-26. Used for: the milkshake case (40% morning share, job, constraints, rivals, response), the correlation-versus-cause argument, the data-organization remark. Secondhand report of Christensen's teaching note "Integrating Around the Job to be Done."
  7. paperRodden, K., Hutchinson, H. and Fu, X. Measuring the User Experience on a Large Scale: User-Centered Metrics for Web Applications. CHI 2010 (4-page note; author's version hosted by Google Research). https://research.google.com/pubs/archive/36299.pdf. Accessed 2026-10-26. Used for: page-view ambiguity (PULSE Metrics), "not a substitute for early or formative user research" and triangulation (Related Work), the Google Finance 2008 surge (Adoption and Retention), log data and task intent (Task Success).
  8. officialChristensen Institute. Jobs to Be Done Theory. Undated organization page (footer: © 2016-2026 Clayton Christensen Institute). https://www.christenseninstitute.org/theory/jobs-to-be-done/. Accessed 2026-10-26. Used for: the definition (circumstances or forces; progress in particular circumstances), the morning-milkshake summary, the condominium case (Moesta) and its self-reported figures, the college example (200+ stories, five jobs). Added by the writer beyond the source map.
  9. textbookChristensen, C. M., Hall, T., Dillon, K. and Duncan, D. S. Competing Against Luck: The Story of Innovation and Customer Choice. HarperCollins, 2016-10-04, 352 pp. (Google Books catalog page). https://books.google.com/books/about/Competing_Against_Luck.html?id=zGd_CwAAQBAJ. Accessed 2026-10-26. Used for: the publisher's statement of the book's thesis only. The book's text was not read; the publisher description is not evidence that the framework works.
  10. blogUlwick, A. Jobs-to-be-Done (A Comprehensive Guide). Strategyn (consultancy vendor page; undated; © 2025). https://strategyn.com/jobs-to-be-done/. Accessed 2026-10-26. Used for: Ulwick's definition, solution-free job statement, desired outcomes, eight-step job map, survey stage, three customer roles; vendor-stated success figures cited as claims only. Added by the writer beyond the source map.
  11. blogCagan, M. Build To Learn FAQ. Silicon Valley Product Group, 2026-04-27. https://www.svpg.com/build-to-learn-faq/. Accessed 2026-10-26. Used for: "what is the hard part: picking, understanding or solving the problem"; the four-risks testing statement; PM role passages; the cost-of-delivery bottleneck argument; the "constant, erratic change" caution. Likely to change within a year.
  12. blogNg, A. The Batch, Issue 284 (letter on AI Product Management). DeepLearning.AI, 2025-01-15. https://www.deeplearning.ai/the-batch/issue-284. Accessed 2026-10-26. Used for: cheaper prototypes and the complements argument; "Perhaps engineers will step in to do some of this work." A forecast, not data.