Humanity’s Last PRD
Latest drop: Haiku 5.5 · 8 Oct
Tasks / Metrics & Experimentation

Draft OKRs

Can the model turn a team's wish list into a few outcome-based OKRs that add up to the company's goals?

Task type v1.1 · 2 tasksLast changed 6 Oct 2026 · ChangelogDifficulty

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

From 16 graded outputs by 8 models. 56% were usable with at most a quick edit.

Reliably right

  1. Builds the key results on the data100% pass
    Key results 2 and 3 are built directly on the first-week joining data and time-to-shared-note data, with baselines from the brief and targets.
    GPT-6.1 Sol · API · Eleven key results and a bonus
  2. Addresses the actual decision97% pass
    The output commits to a clear rewritten OKR set and a note explaining the changes, framed for the squad, and states that targets are proposed pending feasibility review.
    GPT-6.1 Sol · API · Eleven key results and a bonus
  3. Focuses on the big rock94% pass
    The output reduces the draft to one objective with three KRs, explicitly drops dark mode, AI summaries, and NPS, and explains why.
    GPT-6.1 Sol · API · Eleven key results and a bonus

Where it slips

  1. Flags the bonus link59% pass
    Mentions bonus multipliers only to justify outcome KRs; does not name the risk of safe targets or suggest decoupling OKR scores from bonuses.
    Haiku 5.5 · API · Eleven key results and a bonus
  2. Avoids unsupported claims69% pass
    Presents 'the gap is correlation' and 'the safe play is to pick easy KRs' as facts without labeling them as hypotheses.
    Sonnet 5.5 · API · Eleven key results and a bonus
  3. Identifies material uncertainty69% pass
    The output does not name any specific unknowns that could change the OKRs or how they would be resolved.
    GPT-6 Luna · API · Eleven key results and a bonus

How it’s graded

Each check is built on how world-class product and growth leaders do this work, as they’ve described it on Lenny’s Podcast, and names the leader it comes from.

  1. Key results are outcomes, not output

    Is every key result a measurable change in customer or business behaviour, rather than something the team will ship or do?

    Passes when Every key result names a metric with a baseline and a target; shipping work appears only as initiatives that serve them.

    Matt LeMayLenny’s Podcast

  2. Focuses on the big rock

    Does the output cut the goals down to the few that matter most, rather than covering everything the team could do?

    Passes when Few objectives, each with a small number of key results, and says what was dropped and why.

    Christina WodtkeLenny’s Podcast

  3. Shows how the goals add up

    Does the output show how each team-level key result contributes to the company-level goals, and flag any that don't?

    Passes when Every key result is linked to a company goal, with the reasoning or data that connects them, and any that serve no company goal are flagged or cut.

    Matt LeMayLenny’s PodcastClaire Hughes JohnsonLenny’s Podcast

Plus our standard checks

Uses the supplied evidence correctly · Addresses the actual decision · Respects explicit constraints · Identifies material uncertainty · Avoids unsupported claims · Produces the required deliverable · and 2 written for each task, which you’ll see in the tasks below.

The tasks

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

The brief

You're the PM for Fieldnote's onboarding squad. The squad drafted its OKRs for next quarter and asked you to tidy them up before they go to leadership on Friday. Rewrite them, and add a short note to the squad on what you changed and why. No more than 600 words in all. What we know is below.

What the model was given6 items: About Fieldnote, The company's goal for next quarter (set by leadership), The squad's draft, Data, A note from HR, The squad
About FieldnoteShared notes and docs for teams, sold self-serve. Teams sign up free and pay once they pass five members.
The company's goal for next quarter (set by leadership)Objective: make Fieldnote the place new teams get started. Key result: raise the share of new teams still active in week 4 from 34% to 40%.
The squad's draftObjective 1: Ship the new onboarding. Key results: launch an onboarding checklist; ship 5 templates; redesign the invite flow; launch onboarding emails. Objective 2: Improve quality. Key results: fix 50 bugs; get P1 bugs to zero. Objective 3: Delight users. Key results: raise NPS from 31 to 40; launch dark mode; ship AI summaries; run 10 user interviews; publish 4 help articles.
DataNew teams where three or more people join in the first week: 58% still active in week 4. Everyone else: 21%. Today 29% of new teams get three or more people in the first week. Median time from sign-up to the first note someone else opens: 3 days.
A note from HRFrom this year, each squad member's bonus multiplier is set by their squad's OKR score.
The squadFive engineers, one designer and you.
What a strong answer doesThe answer key the graders mark against

Rewrites the draft into one objective tied to the company's (new teams get started) with two or three outcome key results built from the data, such as raising the share of new teams with three or more people in week one from 29% to a stated target, and cutting the time to the first shared note from 3 days to a stated target, with the company's week-4 retention as the result they serve. The checklist, templates, invite redesign and emails become initiatives, not key results. Drops what doesn't serve the goal (dark mode, AI summaries) or moves it out of the OKRs; keeps bug work as a health measure rather than an objective. Flags that tying bonuses to OKR scores will push the squad to set safe targets, and suggests keeping them apart. Explains the changes to the squad briefly and warmly.

Critical failures (cap the score at 40)
  • Keeps the draft's key results essentially unchanged
Task

Mid-level · v1.0

The outputs

8 setups answered this brief.

GPT-6.1 SolAPI100Usable with a quick editGPT-6 AstraChatGPT98Usable with a quick editHaiku 5.5API96Usable with a quick editSonnet 5.5API90Usable with a quick editOpus 5.5Claude79Needs more workGPT-6 LunaAPI75Needs more workGemini 3.5 Flash-LiteGemini71Needs more workGemini 3.8 FlashAPI65Needs more work
Haiku 5.5API

Usable with a quick edit10 of 11 checks passed

The output is accurate and well-structured; the only missing piece is an explicit note on the safe-target risk, which could be added with a quick edit without re-checking evidence or reasoning.

Output

Onboarding squad OKRs, next quarter (draft for Friday)

Objective: Make Fieldnote the place new teams get started.

  • KR1: Raise the share of new teams still active in week 4 from 34% to 40%. (Company KR, shared with leadership.)
  • KR2: Raise the share of new teams with three or more members joining in week one from 29% to 50%.
  • KR3: Cut the median time from sign-up to the first note opened by a second member from 3 days to 1 day.
  • Guardrail: Zero open P1 bugs at quarter end, with no new P1s in onboarding flows.

Delivery plan (tracked, not scored): - Invite flow redesign and onboarding emails first, since they target KR2 and KR3. - Onboarding checklist next, then the five templates. - Ten user interviews with teams that stalled after sign-up, to learn why people didn't invite others. - Dark mode, AI summaries, and the four help articles are paused. If leadership has committed to any of them, let's discuss before Friday.

---

Note to the squad

Hi all. I reworked the draft before it goes to leadership. Here's what changed and why.

1. One objective, tied to the company goal. The draft had three objectives, and several items (dark mode, AI summaries) didn't connect to the week-4 number. Everything now points at one goal: new teams getting started.

2. We're targeting the lever the data points to. Teams with three or more people joining in week one stay active at 58% in week 4. Everyone else stays at 21%. Only 29% of new teams get there today. Getting more teams to that threshold is the biggest lever we can see. It's correlation, though, not proof. Teams that invite friends early may already be more committed. The invite flow work is our test of whether the link is causal, so we'll watch it closely.

3. Outcomes over outputs. Since bonus multipliers now depend on the OKR score, I want the score to measure whether we moved the number, not whether we shipped things. A "ship 5 templates" KR scores 100% even if nobody uses the templates. The delivery plan still shows what we're building, and it has no bearing on the score.

4. Fewer KRs. Ten KRs across three objectives meant spreading five engineers and a designer very thin. Three KRs plus a guardrail give us a clear focus.

5. Bugs stay in as a guardrail. I cut "fix 50 bugs" because a count rewards small fixes over important ones. Zero P1s is the real bar: we shouldn't trade quality for onboarding speed.

6. Targets I'd like your pushback on. The 50% figure for KR2 is my estimate of what would move week-4 activity toward 40%, based on the 58% and 21% split. If the link is weaker than it looks, that target may be wrong, and I'd rather find out in week three than week twelve. Likewise, the one-day target in KR3 is a judgment call. Tell me if the invite flow can't realistically get there.

Thanks for the work so far. The draft's raw material was solid. It just needed one thread running through it. Let's review progress at week six, and I'll bring any changes to the team first.

Grades and run details

Decision model 100 · LLM judge 92
Decision model checks
  • passUses the supplied evidence correctly79%
  • passAddresses the actual decision89%
  • passRespects explicit constraints35%
  • passIdentifies material uncertainty86%
  • passAvoids unsupported claims70%
  • passProduces the required deliverable94%
  • passBuilds the key results on the data100%
  • passFlags the bonus link91%
  • passKey results are outcomes, not output79%
  • passFocuses on the big rock100%
  • passShows how the goals add up76%
Run
Run
#1
API response time
15 s
Submitted
8 Oct 2026
Sonnet 5.5API

Usable with a quick edit9 of 11 checks passed

The output is clear and actionable; the unsupported correlation claim can be fixed with a quick wording edit, and the rest is solid.

Output

Onboarding squad OKRs, next quarter

Objective 1: New teams get their teammates in during week 1, and stay (supports the company goal: make Fieldnote the place new teams get started)

  • KR1: Raise the share of new teams with 3+ people in week 1 from 29% to 45%.
  • KR2: Cut the median time from sign-up to the first note someone else opens from 3 days to 1 day.
  • KR3: Raise the share of new teams active in week 4 from 34% to 40% (shared with the company KR).

Planned bets, not KRs: onboarding checklist, invite flow redesign, 5 templates, onboarding emails. We'll drop or swap any of them if they aren't moving KR1 and KR2.

Objective 2: Keep the product dependable while we change onboarding

  • KR1: Zero open P1 bugs at quarter end, and no P1 open longer than 5 days at any point.

Backlog, not this quarter: dark mode, AI summaries, NPS target, help articles.

---

Note to the squad

Thanks for the draft. Here's what I changed and why.

1. Outcomes, not deliverables. Most of the draft listed things we'd ship. We could ship all of it and leave week-4 activity flat. The checklist, invite redesign, templates and emails are now the bets we make to hit the KRs.

2. One clear line to the company goal. Teams where 3+ people join in week 1 are still active in week 4 58% of the time, versus 21% for everyone else, and only 29% of teams reach 3+ today. So KR1 is about getting more teams to that point. Reaching 40% overall probably needs about 45% of teams there, assuming the relationship holds. KR2 measures how quickly a new team sees value from a teammate.

3. Cut and moved. Dark mode, AI summaries and the NPS target have no clear link to the company goal, and NPS isn't something onboarding alone can move. They go to the backlog, not the bin. User interviews and help articles are useful, but as activities inside Objective 1, not KRs. "Fix 50 bugs" rewarded volume over impact, so it became a P1 guardrail.

4. Two caveats to settle before Friday. - The data implies a week-4 rate of about 32% (0.29×58% + 0.71×21%), not the 34% in the company KR. Likely a definition or timing difference, but I'll confirm which baseline we're using. - The 58% vs 21% gap is correlation. Teams that invite people quickly may already be more committed, so the real lift from nudging invites will probably be smaller. That's why KR3 sits next to KR1, so we'll see whether the outcome really moves.

5. The bonus link. Since scores now affect everyone's multiplier, the safe play is to pick easy, output-style KRs. I've deliberately not done that. KR1 and KR2 are stretch targets, and I'd rather be upfront about it than quietly sandbag. Before Friday I'll ask leadership and HR how scores convert to multipliers, and I'll push for 0.7 on stretch KRs to count as a good result. I'll share the answer as soon as I have it. If you think a target is unrealistic, tell me before Friday so we can adjust it openly.

Please reply with objections or number challenges by Thursday noon.

Claims the judge couldn’t find in the brief

  • The 58% vs 21% gap is correlation.
  • The safe play is to pick easy, output-style KRs.

Grades and run details

Decision model 95 · LLM judge 83
Decision model checks
  • passUses the supplied evidence correctly59%
  • passAddresses the actual decision81%
  • partialRespects explicit constraints23%
  • passIdentifies material uncertainty96%
  • passAvoids unsupported claims49%
  • passProduces the required deliverable80%
  • passBuilds the key results on the data100%
  • passFlags the bonus link100%
  • passKey results are outcomes, not output73%
  • passFocuses on the big rock100%
  • passShows how the goals add up59%
Run
Run
#1
API response time
25 s
Submitted
5 Oct 2026
GPT-6.1 SolAPI

Usable with a quick edit11 of 11 checks passed

The output is clear, evidence-based, and directly usable by a PM with at most minor wording tweaks; it flags a data discrepancy and bonus risk proactively.

Output

Proposed onboarding OKRs — next quarter

Objective: Help new teams experience shared value quickly and keep collaborating.

Key resultBaseline → target
1. Increase the share of new teams still active in week 4. This is our contribution to the company’s retention goal, shared with other squads that influence it.34% → 40%
2. Increase the share of new teams with three or more people joining in their first week.29% → 40%
3. Reduce median time from sign-up to the first note opened by another team member. The share of teams reaching this milestone within seven days must not decline.3 days → 1 day

The targets for KRs 2 and 3 are proposed, pending a feasibility review before leadership approval. They are not forecasts of retention impact.

Measurement and quality guardrails

  • Freeze cohort eligibility, “active,” membership, and shared-note definitions before the quarter starts. Use consistent definitions for baselines and results.
  • Score week-4 retention only on cohorts that have reached week 4; report immature cohorts separately.
  • Reconcile the data before approval: the supplied segment figures imply approximately 32% overall retention, rather than 34%. Confirm whether they use different cohorts or definitions.
  • Track onboarding reliability alongside these KRs. Critical onboarding defects are release blockers, not a bug-count target.
  • Improve collaboration through genuine value—not forced invitations or changes that artificially inflate membership.

Candidate initiatives — not scored key results

Prioritize the checklist, invite-flow improvements, a small set of useful templates, and lifecycle emails according to their expected contribution to these outcomes. Use interviews and funnel analysis to choose and refine experiments.

Defer dark mode and AI summaries unless evidence shows they address a major onboarding barrier. Publish help content where research identifies a specific need.

Note to the squad

I replaced three broad objectives and a shipping checklist with one focused objective and measurable user outcomes. Launches, interviews, articles, and bug fixes are ways to make progress; completing them does not prove that onboarding improved.

Early team participation is our strongest available signal: teams with three or more people joining in week 1 have much higher week-4 retention. That is correlation, not proof of causation, so we’ll test whether our changes improve retention rather than assume invitations alone will do it.

I also narrowed the scope for our five-engineer, one-designer squad. NPS is too broad to diagnose this onboarding problem, while dark mode and AI summaries would compete with the core work.

Because OKR scores now affect bonuses, I want targets, scoring rules, guardrails, and shared ownership agreed before the quarter—not adjusted afterward. We should review the bonus linkage with leadership and HR so it does not reward metric manipulation or discourage honest experiments.

Grades and run details

Decision model 100 · LLM judge 100
Decision model checks
  • passUses the supplied evidence correctly58%
  • passAddresses the actual decision47%
  • passRespects explicit constraints24%
  • passIdentifies material uncertainty59%
  • passAvoids unsupported claims70%
  • passProduces the required deliverable73%
  • passBuilds the key results on the data100%
  • passFlags the bonus link98%
  • passKey results are outcomes, not output98%
  • passFocuses on the big rock100%
  • passShows how the goals add up56%
Run
Run
#1
API response time
27 s
Submitted
5 Oct 2026

Check by check

Where the outputs differ comes first. Open a check to read the judge’s reasons. “Mixed” means the two graders disagreed.

Where they differ 3

Uses the supplied evidence correctlyRightMixedRight
Haiku 5.5 · API

All claims about the current situation are directly supported by the brief or supplied data.

Sonnet 5.5 · API

Claims that the gap is correlation and that the safe play is to pick easy KRs are not supported by the brief or arithmetic.

GPT-6.1 Sol · API

All factual claims about the current situation are taken directly from the brief or derived by correct arithmetic, with no invented facts.

Avoids unsupported claimsRightMixedRight
Haiku 5.5 · API

Hypotheses (e.g., committed teams) are labelled as 'may', and no unsupported certain claims are made.

Sonnet 5.5 · API

Presents 'the gap is correlation' and 'the safe play is to pick easy KRs' as facts without labeling them as hypotheses.

GPT-6.1 Sol · API

Hypotheses and causes are clearly labelled (e.g., 'correlation, not proof of causation'), and no interpretations are presented as established fact.

Flags the bonus linkMixedRightRight
Haiku 5.5 · API

Mentions bonus multipliers only to justify outcome KRs; does not name the risk of safe targets or suggest decoupling OKR scores from bonuses.

Sonnet 5.5 · API

Flags risk of safe targets due to bonus link and suggests pushing for 0.7 on stretch KRs to count as good.

GPT-6.1 Sol · API

The output flags the risk of bonus-linked OKRs leading to safe targets or manipulation, and suggests reviewing the linkage with leadership and HR.

All got right 8

Addresses the actual decisionRightRightRight
Haiku 5.5 · API

Commits to a single revised set of OKRs and explains changes to the squad, meeting the brief's request.

Sonnet 5.5 · API

Commits to rewritten OKRs and note, and says squad feedback by Thursday could change targets.

GPT-6.1 Sol · API

The output commits to a clear rewritten OKR set and a note explaining the changes, framed for the squad, and states that targets are proposed pending feasibility review.

Respects explicit constraintsRightRightRight
Haiku 5.5 · API

Output is under 600 words (~502) and includes both rewritten OKRs and a note to the squad.

Sonnet 5.5 · API

Output is a rewritten OKRs and note under 600 words, addressed to the squad.

GPT-6.1 Sol · API

The output is a note to the squad with rewritten OKRs, well under 600 words, and respects the requested form and reader.

Identifies material uncertaintyRightRightRight
Haiku 5.5 · API

Explicitly notes the correlation is not proven, estimates may be wrong, and outlines how they'll test causality and adjust.

Sonnet 5.5 · API

Names baseline discrepancy and correlation vs causation, and says how to resolve them; also addresses bonus conversion uncertainty.

GPT-6.1 Sol · API

It identifies the data discrepancy, the correlation vs. causation risk, and the need for feasibility review, and says how to resolve them.

Produces the required deliverableRightRightRight
Haiku 5.5 · API

Provides complete, clear OKRs and a well-structured note to the squad that the reader could act on.

Sonnet 5.5 · API

Provides complete OKRs and note within word limit, usable by the squad.

GPT-6.1 Sol · API

The deliverable is a complete, usable OKR rewrite and squad note, in the right form and within the word limit.

Builds the key results on the dataRightRightRight
Haiku 5.5 · API

KR2 uses the first-week member data (29% baseline) and KR3 uses time to first shared note (3 days), both with targets.

Sonnet 5.5 · API

KR1 uses the 29% baseline and KR2 uses the 3-day median, both from the data.

GPT-6.1 Sol · API

Key results 2 and 3 are built directly on the first-week joining data and time-to-shared-note data, with baselines from the brief and targets.

Key results are outcomes, not outputRightRightRight
Haiku 5.5 · API

All key results are outcome metrics with baselines and targets; shipping work is listed as initiatives, not KRs.

Sonnet 5.5 · API

All KRs are outcome metrics with baselines and targets; shipping work is listed as initiatives.

GPT-6.1 Sol · API

All key results are measurable outcomes with baselines and targets; shipping work appears only as candidate initiatives.

Focuses on the big rockRightRightRight
Haiku 5.5 · API

One objective, three KRs plus guardrail, and explicitly drops dark mode, AI summaries, help articles, and bug-count KR.

Sonnet 5.5 · API

Cuts to one main objective with three KRs and a quality guardrail, drops unrelated items and explains why.

GPT-6.1 Sol · API

The output reduces the draft to one objective with three KRs, explicitly drops dark mode, AI summaries, and NPS, and explains why.

Shows how the goals add upRightRightRight
Haiku 5.5 · API

Each KR connects to the company goal via the shared objective and the data on first-week invites and time to shared note.

Sonnet 5.5 · API

Links KR1 and KR2 to company goal via the data, and flags items that don't serve it.

GPT-6.1 Sol · API

Every key result is linked to the company retention goal, with KR1 directly shared and KRs 2-3 as leading indicators, and the reasoning is explained.

Results

Every setup we’ve tested on this task type, across all its tasks and repeats, graded on the current checklist. Calibrated. Where our PM checked the graders by hand, they agreed on 100% of 12 checks.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6 AstraChatGPT97.7100.02None
2GPT-6.1 SolAPI97.7100.02None
3Haiku 5.5API95.595.82None
4Sonnet 5.5API90.991.72None
5GPT-6 LunaAPI88.679.22None
6Opus 5.5Claude93.279.22None
7Gemini 3.8 FlashAPI65.962.52None
8Gemini 3.5 Flash-LiteGemini61.445.82None

What you get for the cost

Haiku 5.5 · API scores within 3 points of GPT-6 Astra · ChatGPT at about a thirtieth of the cost.

Task score against relative cost for each setup. The table below has the same figures.4050607080901001×3×10×30×100×Cost per output, relative to the cheapest setup (log scale) →↑ Task scoreGPT-6 Astra · ChatGPT: task score 99, 83× the cheapestGPT-6.1 Sol · API: task score 99, 16× the cheapestHaiku 5.5 · API: task score 96, 2.1× the cheapestSonnet 5.5 · API: task score 92, 44× the cheapestGPT-6 Luna · API: task score 82, the cheapestOpus 5.5 · Claude: task score 82, 78× the cheapestGemini 3.8 Flash · API: task score 61, 13× the cheapestGemini 3.5 Flash-Lite · Gemini: task score 51, 2.1× the cheapestGPT-6 Astra · ChatGPTGPT-6.1 Sol · APIHaiku 5.5 · APISonnet 5.5 · APIGPT-6 Luna · APIOpus 5.5 · ClaudeGemini 3.8 Flash · APIGemini 3.5 Flash-Lite · Gemini
Best valueNobody beats these on both score and cost

Cost is each output’s tokens at the model’s list price, shown relative to the cheapest setup. Runs done in an app have no usage meter, so theirs are estimated from the same setup’s measured runs. How we compare cost

The PM job

Setting a team's goals for the quarter.

Why it matters

Most OKRs are task lists in disguise. Teams ship everything on them and nothing moves.

What good looks like

  • Key results are outcomes, not things to ship
  • Few enough to focus on
  • Each team's goals visibly add up to the company's
  • Kept apart from performance ratings

Deliberately not measured

  • OKR software or formatting
Capability tested

Goal setting

The failure we’re looking for

A task list dressed up as key results

Grading

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