Humanity’s Last PRD
Latest drop: Sonnet 5.5 · 30 Sept
Tasks / Define

Develop product strategy

Can the model make a coherent choice grounded in the evidence, rather than list aspirations?

Measures the modelTask v1.1 · 2 casesDifficulty

What AI gets right here, and what you’ll still have to catch

From 12 graded outputs by 6 models. 75% were usable with at most a quick edit.

Reliably right

  1. Produces the required deliverable100% pass
    It is a complete one-page strategy for the CEO/board, with staged funding, metrics, and risk controls, usable as-is with light edits.
    GPT-6 Astra · ChatGPT · Supply or demand for a stalled marketplace
  2. Makes a clear choice100% pass
    It chooses matching quality as the one bet and names what it forgoes: broad paid acquisition, broad tutor recruitment, and unrelated marketplace expansion.
    GPT-6 Astra · ChatGPT · Supply or demand for a stalled marketplace
  3. Diagnosis before prescription100% pass
    It diagnoses the mismatch between abundant supply and failed matching, and the proposed actions directly target tutor choice and repeat bookings.
    GPT-6 Astra · ChatGPT · Supply or demand for a stalled marketplace

Where it slips

  1. Proposes tests that could fail35% pass
    The months 7–12 gate says 'gains persist and contribution economics support expansion' without a numeric threshold or defined contribution target, so not every gate names a threshold.
    GPT-6 Astra · ChatGPT · Supply or demand for a stalled marketplace
  2. Uses the supplied evidence correctly40% pass
    The output makes unsupported claims about the current system (open-ended tutor browsing as default journey, existence of data on subject/exam-board segments) that are not in the supplied context.
    GPT-6.1 Sol · API · Supply or demand for a stalled marketplace
  3. Avoids unsupported claims48% pass
    Presents as fact that the current default journey is open-ended tutor browsing and that data on highest-volume subject/exam-board segments exists, neither of which is in the supplied evidence.
    GPT-6.1 Sol · API · Supply or demand for a stalled marketplace

Case viewer

Read the brief, then put up to three outputs side by side, each with the LLM judge’s verdict on every check. Highlights mark what a PM had to fix.

The brief

Our tutoring marketplace has stalled at about 4,000 bookings a month. Write a one-page strategy for the next 12 months (about 500 words) for our CEO to take to the board. We can fund one major bet.

ScenarioTutorly connects secondary-school students with independent tutors for online lessons at £30–45 an hour; we take a 20% commission. There are 6,500 active tutors and 11,000 active students. We have 18 months of runway and about £1.2m to put behind the one bet.
MetricsTutor supply up 60% year on year; bookings up 8%. 70% of tutors get zero bookings in a typical month. Repeat booking rate is 64% for students who book twice.
ResearchStudents cite 'finding the right tutor' as their top problem. Tutors cite 'not enough students'.
Board viewThe board believes paid acquisition of students is the answer.
What a strong answer does

Recognise the contradiction: supply is abundant but matching fails, so buying demand into a broken match pours water into a leaky bucket. Choose matching quality and say what is given up.

Critical failures (cap the score)
  • Recommends paid demand acquisition without addressing matching
Case

v1.9 · synthetic · marketplace, board pressure

GPT-6 AstraChatGPTUsable with a quick editGPT-6 LunaAPIUsable with a quick editGPT-6.1 SolAPIUsable with a quick editSonnet 5.5APIUsable with a quick editOpus 5.5ClaudeNeeds more workGemini 3.5 Flash-LiteGeminiNeeds more work
GPT-6 LunaAPI

Usable with a quick edit8 of 10 checks passed

The output is well-structured and evidence-based; a quick edit to add a numeric threshold for the pilot kill criterion would make it fully usable without re-checking evidence or redoing reasoning.

Output

Tutorly: 12-month strategy — make every good match count

Recommendation: Put our £1.2m behind improving the first tutor match, not broad paid acquisition. Tutor supply is growing far faster than bookings, 70% of tutors get no booking in a typical month, and students say finding the right tutor is their main problem. Buying more student traffic before fixing that experience risks paying to send more families into the same frustration. The encouraging signal is that students who book again show a 64% repeat rate: when a match works, there is reason to believe students stay.

The bet: a guided matching and first-booking service

Build a faster, more trusted path from a student’s need to a first lesson. Ask about subject, exam level, goals, budget, availability and preferred teaching style; then show a short, ranked list of tutors who are available and suitable—not a directory students must search alone. Make tutor profiles clearer on relevant experience and approach, and prompt tutors to respond quickly. For students who still struggle, offer human-assisted matching.

Start with the subjects and exam levels that already have the strongest demand and enough suitable tutors. Use actual enquiries, bookings, repeat behaviour and student feedback to improve recommendations. This is one integrated liquidity bet: make existing demand easier to serve with existing supply, rather than adding supply or buying broad traffic.

How we will execute

In the first 90 days, analyse the booking funnel and pilot assisted matching in a few high-demand subject/exam segments. Compare it with the current journey on first booking, time to match, cancellations and student satisfaction. In months 4–6, automate the most successful parts and improve availability and profile information. In months 7–12, expand only where the pilot improves bookings without harming quality. Help tutors without bookings improve their profiles or availability; stop actively promoting tutors who are unavailable or consistently unresponsive.

Allocate the £1.2m across the year: £450k product and data, £300k matching and student support, £250k tutor onboarding and quality, £100k testing and measurement, and £100k contingency. Release funding in stages, with a formal review after the pilot.1

What success looks like

By month 12, target 6,000 monthly bookings (up from 4,000) and at least 45% of active tutors receiving a booking in a typical month (up from 30%). Also improve first-booking conversion and time to first lesson, while maintaining or improving repeat booking, cancellation and satisfaction rates. Report results by subject and exam level so growth cannot hide weak or overloaded segments.

Hold broad paid acquisition for now. Run only small, measurable acquisition tests in segments where matching is working; scale spend only if the resulting students book and repeat at acceptable acquisition cost. If the pilot fails to improve first bookings, stop and reassess before committing the remaining funds. This gives the board a measurable growth bet while protecting our 18-month runway.

What a PM had to fix

  1. 1Invented evidenceVerify or remove the claimTargeted repair

    What we’d changeTreat £1.2m as a funding ceiling, not a spending commitment. Assuming one-hour lessons, reaching 6,000 monthly bookings adds only £12k–18k in monthly commission revenue before costs. That target alone does not justify the investment. Cost the 90-day matching pilot first; release further funding only if incremental completed bookings, repeat behaviour and service costs support a credible path to covering the investment within our financing horizon. Assess that path against actual cash burn and remaining runway before scaling.

Check by check

Got wrong · 1
  • Proposes tests that could failThe pilot kill criterion lacks a numeric threshold; it only says 'fails to improve first bookings' without a specific measurable target.
Mixed · 1
  • Uses the supplied evidence correctlyAll statements about the current situation are taken directly from the supplied context with no invented facts.The two graders disagreed on this one.
Got right · 8
  • Addresses the actual decisionCommits early to one choice (improving matching, not paid acquisition), framed for the CEO, and says the pilot failing to improve first bookings would trigger reassessment.
  • Respects explicit constraintsDelivers a one-page strategy of about 500 words for the CEO to take to the board, respecting the form, reader and length.
  • Identifies material uncertaintyIdentifies that the effect of matching improvement on bookings is uncertain, and specifies that a pilot will resolve it, with a clear stop condition if it fails.
  • Avoids unsupported claimsInterpretations and forecasts are presented as risks or hypotheses, not as established facts; factual claims are all supported.
  • Produces the required deliverableThe output is a complete, actionable strategy memo that the CEO could take to the board with minimal edits.
  • Surfaces the central contradictionClearly surfaces the contradiction: supply grew 60% but bookings only 8%, 70% of tutors idle, and students can't find the right tutor.
  • Makes a clear choiceChooses one bet (matching quality) and explicitly forgoes broad paid acquisition, naming the sacrifice.
  • Diagnosis before prescriptionDiagnoses the crux (broken matching despite abundant supply) and all major actions directly address that crux.

Grades and run details

Decision model 80 · LLM judge 10 of 11 checks
Decision model checks
  • failUses the supplied evidence correctly42%
  • passAddresses the actual decision100%
  • passRespects explicit constraints66%
  • passIdentifies material uncertainty95%
  • partialAvoids unsupported claims11%
  • passProduces the required deliverable96%
  • passSurfaces the central contradiction100%
  • passMakes a clear choice100%
  • passDiagnosis before prescription100%
  • partialProposes tests that could fail80%
Run
Run
#1
API response time
21 s
Submitted
29 Sept 2026
Sonnet 5.5API

Usable with a quick edit8 of 10 checks passed

The strategy is clear and actionable; minor unsupported claims (e.g., £50 acquisition cost) can be removed or labelled as assumptions with a quick edit, without redoing the reasoning.

Output

Tutorly 12-Month Strategy: Fix Matching, Then Scale

Recommendation

Put the £1.2m behind one bet: guided matching with a first-lesson guarantee. Paid acquisition comes second, released in stages once the funnel converts.

Why not paid acquisition first

Tutor supply is up 60% and bookings only 8%, so the constraint is not volume. Seventy percent of tutors get no bookings, and students say their biggest problem is finding the right tutor. Tutors say "not enough students," but that is a symptom of poor discovery, not a shortage of demand alone.

The economics are also tight. At an average £37.50 lesson and 20% commission, we earn about £7.50 per booking. A £50 acquisition cost needs roughly seven bookings from a student just to break even. Our best evidence of retention is that 64% of students who book twice book again, but we don't yet know how many get to a second booking. Paying to send more students into a search experience they find frustrating risks buying churn.

The bet

1. Guided matching (about £600k). Replace browse-and-search with a short intake (subject, exam board, grade target, learning needs, availability, budget) that returns three ranked tutors. Add structured tutor profiles and quality signals such as response time, outcomes and reviews. Route new demand to under-booked tutors who fit well, rather than only to the top few.

2. First-lesson guarantee (about £300k). If the first lesson isn't right, the student is rematched and the lesson is free, funded by us. This targets the drop-off before the second booking, which is where the strong repeat behaviour begins.

3. Supply curation (about £100k). Stop passive tutor onboarding. Admit tutors by subject-level demand, and give inactive tutors coaching or profile support, or deprioritise them.

4. Gated acquisition reserve (£200k). Held back and released only when the gates below are met.

Gates for releasing acquisition spend

Two consecutive months of: - first-to-second booking conversion up at least 10 points from baseline; - median time to first booking under 48 hours; - modelled payback on acquisition inside 6 months.

Until then, test small paid channels only to measure cost per matched student.

Targets (month 12)

  • Bookings: 4,000 → 6,000+ per month
  • Tutors with at least one booking a month: 30% → 50%
  • Students reaching a second booking: +10 points
  • Student satisfaction after first lesson: tracked from launch

Timeline

  • Months 1–2: Instrument the funnel (search → enquiry → first lesson → second lesson). Establish the baseline. Interview 30 students and 30 tutors.
  • Months 3–6: Launch matching to one or two high-demand subjects, such as GCSE maths and sciences. Run against a control group.
  • Months 7–9: Roll out across all subjects, and launch the guarantee if the pilot shows it pays back.
  • Months 10–12: Release acquisition spend if the gates are met. Report to the board.

Risks

  • Matching doesn't lift conversion. Pilots with control groups will show this by month 6, leaving 12 months of runway to change course.
  • Guarantee costs exceed benefit. We cap exposure per student and monitor abuse.
  • Tutor churn from curation. We prioritise communication and coaching before any deprioritisation.

Ask of the board

Approve the matching bet and the staged acquisition gate. If the board is not willing to defer scaled paid acquisition, the fallback is a capped £150k test with cost per second booking as the success measure.

Check by check

Got wrong · 2
  • Uses the supplied evidence correctlyOutput states that students find the search experience frustrating, that we don't know how many get to a second booking, and implies a £50 acquisition cost, none of which are in the supplied context.
  • Avoids unsupported claimsPresents 'search experience they find frustrating' and a £50 acquisition cost as facts without support, and states 'we don't yet know how many get to a second booking' as if known.
Got right · 8
  • Addresses the actual decisionCommits to guided matching with first-lesson guarantee as the one bet, defers paid acquisition, and provides a fallback if the board disagrees.
  • Respects explicit constraintsRespects the one-major-bet constraint, delivers a one-page strategy of about 500 words for the CEO to take to the board.
  • Identifies material uncertaintyNames risks (matching not lifting conversion, guarantee costs, tutor churn) and sets gates with numeric thresholds and a fallback plan.
  • Produces the required deliverableDelivers a complete one-page strategy with recommendation, rationale, plan, timeline, risks, and ask, usable by the CEO.
  • Surfaces the central contradictionHighlights that supply is up 60% but bookings only 8% and 70% of tutors get no bookings, surfacing the matching failure.
  • Makes a clear choiceChooses matching as the one bet and explicitly forgoes immediate paid acquisition, naming what is sacrificed.
  • Diagnosis before prescriptionDiagnoses poor discovery as the crux, and each major action (guided matching, guarantee, curation) directly addresses it.
  • Proposes tests that could failGates have numeric thresholds (10-point conversion lift, under 48 hours, 6-month payback), measurement windows (two consecutive months, by month 6), and actions (release spend or change course).

Claims the judge couldn’t find in the brief

  • we don't yet know how many get to a second booking
  • search experience they find frustrating
  • A £50 acquisition cost

Grades and run details

Decision model 75 · LLM judge 9 of 11 checks
Decision model checks
  • failUses the supplied evidence correctly12%
  • passAddresses the actual decision100%
  • passRespects explicit constraints27%
  • passIdentifies material uncertainty100%
  • failAvoids unsupported claims14%
  • passProduces the required deliverable78%
  • passSurfaces the central contradiction100%
  • passMakes a clear choice99%
  • passDiagnosis before prescription100%
  • partialProposes tests that could fail62%
Run
Run
#1
API response time
20 s
Submitted
29 Sept 2026

Results

Every setup we’ve tested on this task, across all cases and repeats, graded on the current checklist. Calibrated: the graders match our PM on 84% of checks.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6 LunaAPI87.290.52None
2GPT-6.1 SolAPI87.285.92None
3Sonnet 5.5API79.285.92None
4GPT-6 AstraChatGPT89.775.52None
5Opus 5.5Claude78.971.42None
6Gemini 3.5 Flash-LiteGemini58.357.72None

The PM job

Writing a strategy memo that leadership can act on.

Why it matters

Strategy that doesn't choose isn't strategy. Models write fluent strategic prose easily; making a choice the evidence supports and naming what it gives up is harder.

What good looks like

  • Diagnoses the one obstacle that matters before prescribing
  • Makes one clear choice and names what is sacrificed
  • Grounds the choice in the supplied evidence
  • Surfaces the tension or contradiction in the data
  • Defines how we would know it is working

Deliberately not measured

  • Market sizing accuracy beyond the supplied data
  • Financial modelling
Capability tested

Choosing where to play and what not to do, from supplied evidence

The failure we’re looking for

Strategic-sounding aspirations without a diagnosis or a choice

Grading

Decision model and LLM judge, calibrated against a blind PM review