Tasks / Product Craft

Build a roadmap

Can the model sequence bets against capacity and dependencies, and explain the order?

Task type v1.0 · 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. 69% were usable with at most a quick edit.

Reliably right

Produces the required deliverable95% pass
The roadmap is complete, in the right form for a Group PM, within the word limit, and actionable without major gaps.
GPT-6.1 Sol · API · Fourteen requests from one customer
Outcomes, with certainty that falls with distance94% pass
Every item names its outcome (e.g., reduce paper transcription, reduce manual reconciliation), and near-term items are specific while later ones are deliberately looser (e.g., '7 weeks for rollout, workflow improvement and contingency').
GPT-6.1 Sol · API · Fourteen requests from one customer
Handles the big customer with evidence94% pass
It uses Calder's rising usage and the second-hand 'might leave' to propose a conversation first, not committing the report builder, and offers a bounded export if needed.
GPT-6.1 Sol · API · Fourteen requests from one customer

Where it slips

Sees through the request count72% pass
The output never counts requests by requester, does not mention that 14 requests come from Calder or that several are duplicates, and does not compare themes by customer count.
GPT-6 Luna · API · Fourteen requests from one customer
Uses the supplied evidence correctly72% pass
States credit notes as 4 customers, but the request log contains 5 credit note tracking requests.
Sonnet 5.5 · API · Fourteen requests from one customer
Fits the stated capacity73% pass
Q3 Expenses row lacks squad-week allocation, so the committed plan cannot be fully checked against the 11-week capacity.
Opus 5.5 · Claude · A year of spend management, with a hard deadline

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.

Outcomes, with certainty that falls with distance
Does each roadmap item say the problem or outcome it serves, and does precision (dates, scope) decrease for later items?
Passes when Every item names its outcome or problem. Near-term items are specific and later ones are deliberately looser.
Melissa PerriLenny’s Podcast Christian IdiodiLenny’s Podcast
Fits the stated capacity
Does the committed plan fit the team capacity the brief gives, with sums a reader could check, and does it say what was cut or deferred to make it fit?
Passes when The committed work adds up to no more than the stated capacity, with some slack, the sums can be checked, and it names what was cut or pushed later and why.
Gokul RajaramLenny’s Podcast
Tells the story of the bets
Does the roadmap group the work into a few themes and explain why those are the biggest things to invest in now, rather than presenting a scored or ranked list of projects?
Passes when A handful of themes, each with the reason it matters most now, and the items sit under them.
Jiaona ZhangLenny’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 · Sequences around dependencies · 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 Group PM at Fernhill. Build the roadmap for the next half (Q1 and Q2) for our two squads, in no more than 900 words plus a table: what each squad builds, in what order, the outcome each item serves, what we're not doing and why, and what we tell Calder Group. The request log, usage, capacity and goal are attached.

What the model was given5 items: About Fernhill, company_goal_h1.md, feature_requests_last_6_months.csv, feature_usage.md, capacity_and_estimates.md

About FernhillPurchasing and inventory software for restaurants and hotels: orders, deliveries, invoices and stock.

company_goal_h1.md5 lines · Download

# Company goal for the half

Cut churn among independent restaurants (82% of accounts, 61% of revenue) from 2.4% to 1.6% a month. Top churn reason in exit surveys: "too much admin; didn't save us time on deliveries and invoices".

Calder Group renews in May. Their account manager says they "might" leave without the report builder; they haven't raised it with Customer Success, and their usage is up 20% this year.

feature_requests_last_6_months.csv53 rows · Download

request,requester,requester_arr_usd,source,title,theme
FR-300,Customer 553,4353,Support ticket,Credit note tracking,Invoicing
FR-301,Customer 391,9773,Sales call,Match invoices to deliveries automatically,Invoicing
FR-302,Customer 341,4856,Support ticket,Match invoices to deliveries automatically,Invoicing
FR-303,Calder Group,186000,Account manager,ERP export (CSV variant),Integrations
FR-304,Customer 895,9595,Sales call,Mobile receiving (check deliveries on a phone),Receiving
FR-305,Calder Group,186000,Account manager,Report builder with pivots,Reporting
…

feature_usage.md8 lines · Download

# Usage of related features today (last 30 days, 1,140 accounts)

| Feature | Accounts using it | Note |
|---|---|---|
| Standard reports | 212 (19%) | Calder Group uses them daily |
| Receiving checklist (desktop only) | 640 (56%) | 71% of uses start on a desktop, then get re-typed from paper |
| Invoice upload | 870 (76%) | Matching is manual |
…

capacity_and_estimates.md5 lines · Download

# Capacity, Q1 to Q2

Two squads, about 10 squad-weeks of roadmap work each per quarter (40 in total over the half).

Estimates (squad-weeks): Mobile receiving 8. Flag short deliveries 3 (needs mobile receiving). Invoice matching 9 (needs the new invoice parser, 5). Credit note tracking 3 (needs invoice matching). Accounting export with credits 2 (needs credit note tracking). Custom report builder 12. ERP export 4. Approval chains 6. Site-level permissions 4. SSO 3.

What a strong answer doesThe answer key the graders mark against

Builds the half around the churn goal (independent restaurants, whose top exit reason is admin around deliveries and invoices), not the request count. The log has 53 requests, but 14 come from one customer, Calder Group, and several of theirs are the same thing under different names (report builder three times, ERP export twice, bulk SKU twice, scheduled reports twice); across everyone else, receiving and invoicing requests dominate (mobile receiving 9, invoice matching 8, flag short deliveries 7, credit notes 5, accounting export 4). Sequences the dependency chain: mobile receiving (8), then flag short deliveries (3); the invoice parser (5), then invoice matching (9), then credit note tracking (3), then the accounting export (2): about 30 squad-weeks, inside the 40 available with slack. Handles Calder with evidence: their usage is up 20% and they haven't raised leaving with Customer Success, so talk to them before committing the 12-week report builder; offer a smaller item (SSO, 3, or the ERP export, 4) if it fits. Says what's not being done (report builder, approval chains, dark mode, multi-currency) and why, with near-term items specific and later ones looser.

Critical failures (cap the score at 40)

Schedules an item before the item it depends on (for example invoice matching before the parser)

Task

Staff-level · v1.0

The outputs

8 setups answered this brief.

GPT-6.1 SolAPI100Usable with a quick edit Opus 5.5Claude98Usable with a quick edit GPT-6 AstraChatGPT96Usable with a quick edit Sonnet 5.5API94Usable with a quick edit GPT-6 LunaAPI92Usable with a quick edit Haiku 5.5API90Usable with a quick edit Gemini 3.8 FlashAPI71Needs more work Gemini 3.5 Flash-LiteGemini48Needs more work

GPT-6 LunaAPI

Usable with a quick edit11 of 12 checks passed

The roadmap is complete, evidence-based, and directly usable; a PM could adopt it with minor wording tweaks and no re-checking of facts or logic.

Output

H1 roadmap

Priority: reduce admin that contributes to independent-restaurant churn. We’ll automate the invoice-to-credit workflow and make receiving work where deliveries happen. That addresses the stated churn reason and repeated requests across customers. Calder’s renewal matters, but the report-builder risk is unconfirmed, so we’ll validate it before committing 12 squad-weeks to a bespoke capability.

Order	Squad / timing	What we build	Outcome served
1	Squad A · Q1 (8 weeks)	Mobile receiving	Lets staff check deliveries on a phone instead of re-typing paper checklists. Desktop-to-paper re-entry is a clear source of admin.
2	Squad B · Q1 (5 weeks)	Invoice parser	Prerequisite for matching; prepares invoice data for automation.
3	Squad B · Q1–Q2 (9 weeks total: 5 in Q1, 4 in Q2)	Match invoices to deliveries	Reduces manual reconciliation, directly addressing “too much admin”; invoice upload is already used by 76% of accounts. Starts after the parser.
4	Squad A · Q2 (3 weeks)	Flag short deliveries	Surfaces discrepancies at receiving, before they become invoice errors. Starts after mobile receiving.
5	Squad B · Q2 (3 weeks)	Credit note tracking	Extends the automated invoice workflow to credits; starts after matching.
6	Squad B · Q2 (2 weeks)	Accounting export with credits	Makes the completed invoice-and-credit workflow easier to carry into accounting; starts after credit tracking.
—	Squad A · Q1 (2 weeks); Squad A · Q2 (7 weeks)	Field validation, rollout, adoption and measurement for the above	Validate the receiving workflow with restaurants, help teams adopt it, and measure whether it reduces re-entry and invoice admin. Include direct Calder discovery in Q1.
—	Squad B · Q2 (1 week)	Hardening and release buffer	Protects quality across the dependent invoice workflow before broad rollout.

This uses the 40 available squad-weeks: 20 per quarter. The invoice sequence is parser → matching → credit tracking → accounting export; receiving is mobile receiving → short-delivery flags.

Not doing this half

Custom report builder, dashboards and scheduled reports: 12 weeks, with the reported renewal risk coming from the account manager rather than Calder’s CS team; Calder’s usage is up 20% and it uses standard reports daily. We’ll validate the specific reporting job and whether it is truly a renewal blocker before reallocating capacity. A bespoke builder is not currently the best fit for the independent-restaurant churn goal.
ERP export, SSO, site permissions, approval chains, custom fields, catalogue work and other Calder requests: several are substantial or bespoke, and none has stronger evidence against the stated churn driver than the selected workflow improvements. We won’t imply these are committed.
More suppliers and dark mode: lower priority than reducing delivery and invoice admin.

What we tell Calder Group

“We’ve heard that reporting may affect your May renewal. We’re not committing to a custom report builder yet: your teams use standard reports daily, usage is growing, and we haven’t confirmed with you which reporting tasks are blocked. We’d like to meet with your reporting users and Customer Success early in Q1 to understand the must-have use cases and renewal criteria. We’ll come back with a clear proposal; meanwhile, we’ll be investing in reducing delivery and invoice admin across Fernhill.”

If that conversation confirms custom reporting is a hard renewal condition, we’ll bring the trade-off back for an explicit roadmap decision rather than promise a feature we haven’t scoped or prioritized.

Grades and run details

Decision model 92 · LLM judge 92

Decision model checks

passUses the supplied evidence correctly47%
passAddresses the actual decision95%
passRespects explicit constraints25%
passIdentifies material uncertainty98%
partialAvoids unsupported claims38%
passProduces the required deliverable58%
partialSees through the request count64%
passHandles the big customer with evidence97%
passOutcomes, with certainty that falls with distance67%
passFits the stated capacity15%
passSequences around dependencies88%
passTells the story of the bets87%

Run

Run: #1
API response time: 58 s
Submitted: 6 Oct 2026

Haiku 5.5API

Usable with a quick edit11 of 12 checks passed

The output is a complete, evidence-based roadmap with a clear story, actionable table and Calder plan. A PM could use it with light edits and no need to re-verify facts or restructure the reasoning.

Output

Fernhill roadmap, Q1 and Q2

The call: Both squads go to the company goal: deliveries and invoices, the top reasons independents give for leaving. Calder gets a bounded share of capacity, not the roadmap. Calder's report builder can't ship before May, so we tell them that directly and commit to a scoped plan.

Why this split: - The receiving and invoice chains (30 of 40 squad-weeks) serve 32 different non-Calder accounts. Mobile receiving and flagging short deliveries touch 16 accounts. Invoice matching, credit notes and the accounting export touch 16. - Calder has 14 of the 53 requests, but they come from one account. Every Calder-only item totals 29 squad-weeks (report builder, ERP export, site permissions, SSO, approval chains). Taking all of it would consume 73% of the half. - Usage backs the sequence. 56% of accounts use the receiving checklist, but 71% of uses start on a desktop and get re-typed from paper, so mobile receiving removes that step. Invoice matching is manual for the 76% of accounts that upload invoices.

Capacity: 2 squads × 10 squad-weeks per quarter. The plan uses 40 of 40, with a 1-week spike per squad left at the end of Q2 for report builder scoping.

Order	Squad	Item	Weeks	Timing	Outcome it serves	Notes
1	1	Mobile receiving	8	Q1 wk 1–8	Less time per delivery; ends paper re-typing	9 requests, 9 accounts
2	2	Invoice parser	5	Q1 wk 1–5	Prerequisite for invoice matching	Only item with no visible customer output; it gates the invoice chain
3	2	Invoice matching	9	Q1 wk 6–10 to Q2 wk 4	Less manual invoice admin	8 accounts; biggest admin saving
4	1	Flag short deliveries	3	Q1 wk 9–10 to Q2 wk 1	Catch shortages at the door, not on the invoice	Needs item 1; 7 accounts
5	1	ERP export (CSV/ERP format)	4	Q2 wk 2–5	Calder renewal; integration	Calder only; ships before May renewal
6	2	Credit note tracking	3	Q2 wk 5–7	Credits matched to invoices	Needs item 3; 4 accounts
7	2	Accounting export with credits	2	Q2 wk 8–9	Invoices and credits reach accounting without re-keying	Needs item 6; 4 accounts
8	1	Site-level permissions	4	Q2 wk 6–9	Calder multi-site access	Calder only; see below
9	1 + 2	Report builder scoping spike	1 each	Q2 wk 10	Scoped plan and H2 date for Calder	Design only, no build

Dependencies are the risk. Item 2 gates the invoice chain. If the parser slips, items 3, 6 and 7 slip with it, and the invoice outcome misses the half. I'd rather flag that now than hide it inside the plan. If the parser runs over, the first move is to pull the Calder-only site permissions work (item 8) back into the buffer.

Measures: Time per delivery receipt and percentage of invoices auto-matched, tracked per cohort. Independent-restaurant churn (2.4% → 1.6% a month) is the outcome. We won't claim the roadmap alone moves it.

What we're not doing, and why

Full custom report builder (12 weeks). It's 60% of one squad's half, for one account, and the earliest it ships is around the May renewal with no slack. We do a scoping spike and commit to a decision date, not a build date.
SSO (3) and approval chains with 5 levels (6). Calder only, and the 5-level approvals request is a niche configuration. H2 candidates, to be revisited with Calder.
Custom PDF invoices, custom order fields, per-site dashboards, bulk SKU import v2 and bulk SKU edit. All Calder-only, none estimated. Not in H1.
Dark mode (2 accounts), multi-currency (1 account), more suppliers (3 accounts). Low demand, no estimates. Parked, not rejected.
Scheduled reports by email (Calder, 2 requests). Not estimated. It goes into the spike as a candidate for a smaller Calder deliverable.

What we tell Calder Group

Customer Success calls them this week, not the account manager alone. They haven't raised the report builder with CS, and "might leave" is the kind of risk that surfaces too late.
Be direct about the report builder: it's a substantial build, it won't land before May, and we won't promise a date we can't keep.
What they get in H1: the ERP export in the format they asked for, shipping around early May, and site-level permissions by the end of June.
The report builder: we scope it with them in June, and we commit to an H2 decision date by the end of June. We don't commit to a build date until the spike returns.
Ask them two things: which reports they run daily (we know they use standard reports), and what minimum output would cover them until the full builder ships. Their answer decides whether scheduled email reports move up.

Asks for leadership: confirm Calder's renewal status and ARR before the CS call, and approve the spike's two squad-weeks. If Calder won't engage on the scoped plan, we should know before the roadmap is shared externally.

Grades and run details

Decision model 79 · LLM judge 100

Decision model checks

failUses the supplied evidence correctly6%
passAddresses the actual decision95%
partialRespects explicit constraints23%
passIdentifies material uncertainty91%
partialAvoids unsupported claims56%
passProduces the required deliverable69%
passSees through the request count88%
passHandles the big customer with evidence95%
passOutcomes, with certainty that falls with distance80%
partialFits the stated capacity18%
passSequences around dependencies73%
passTells the story of the bets86%

Run

Run: #1
API response time: 56 s
Submitted: 8 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 2

Uses the supplied evidence correctlyRightMixed

GPT-6 Luna · API

All factual claims about the current situation are directly supported by the supplied context.

Haiku 5.5 · API

Every statement about the current situation can be traced to the supplied context or is derived by clear arithmetic from it, with no invented facts.

Sees through the request countWrongRight

GPT-6 Luna · API

The output never counts requests by requester, does not mention that 14 requests come from Calder or that several are duplicates, and does not compare themes by customer count.

Haiku 5.5 · API

It explicitly notes Calder's 14 requests come from a single account, de-duplicates themes, and weights demand by distinct customer count rather than raw request volume.

All got right 10

Addresses the actual decisionRightRight

GPT-6 Luna · API

Commits to a clear roadmap (receiving and invoicing automation) and says the Calder conversation would change the call if reporting is a hard renewal condition.

Haiku 5.5 · API

The roadmap commits unambiguously to investing in deliveries and invoices to address churn, with a clear call for Calder, and states what would change the plan (parser slippage).

Respects explicit constraintsRightRight

GPT-6 Luna · API

The output is under 900 words plus a table, includes the required sections, and addresses the Group PM reader.

Haiku 5.5 · API

The output is well under 900 words plus a table, includes the required elements (ordered items, outcomes, not-do list, Calder plan), and the proposed plan respects the squad and capacity limits.

Identifies material uncertaintyRightRight

GPT-6 Luna · API

Identifies the unconfirmed report-builder risk and states that a conversation with Calder will resolve it and could change the plan.

Haiku 5.5 · API

It names the key dependency risk (invoice parser slip) and how it would be resolved (pull site permissions), and acknowledges the roadmap alone won't guarantee the churn outcome.

Avoids unsupported claimsRightRight

GPT-6 Luna · API

Interpretations like 'desktop-to-paper re-entry is a clear source of admin' are grounded in the usage data; no unsupported confident claims are made.

Haiku 5.5 · API

Interpretations and forecasts (e.g., why mobile receiving reduces admin) are clearly linked to the data, and no claim is presented as fact that isn't supported.

Produces the required deliverableRightRight

GPT-6 Luna · API

Provides a complete roadmap table with order, squad, timing, outcomes, a not-doing list with reasons, and a Calder message; a PM could act on it with light edits.

Haiku 5.5 · API

The roadmap includes a detailed table, explicit not-doing list, and a specific Calder communication plan, all usable by a Group PM as-is.

Handles the big customer with evidenceRightRight

GPT-6 Luna · API

Uses Calder's rising usage and the second-hand 'might leave' to propose a discovery conversation instead of committing the report builder; no small item fits the full capacity, so omission is acceptable.

Haiku 5.5 · API

It uses Calder's rising usage and the second-hand 'might leave' to propose a direct CS conversation and a scoping spike, instead of committing the full report builder.

Outcomes, with certainty that falls with distanceRightRight

GPT-6 Luna · API

Every item names its outcome; near-term items are specific and later items (rollout, hardening) are deliberately looser.

Haiku 5.5 · API

Every roadmap item states the outcome or problem it serves; early items have precise timing, and the later report-builder spike is deliberately scoped rather than committed.

Fits the stated capacityRightRight

GPT-6 Luna · API

The committed work sums to 40 squad-weeks (30 for features + 9 for validation/rollout + 1 buffer), and the output names what was cut and why.

Haiku 5.5 · API

The committed work adds up to exactly 40 squad-weeks, the sums are verifiable, it names the items deferred (full report builder, SSO, approval chains, etc.), and includes a buffer/retreat plan.

Sequences around dependenciesRightRight

GPT-6 Luna · API

All dependencies are respected (parser before matching, matching before credit notes, credit notes before accounting export, mobile receiving before flag short deliveries) and the chains are named.

Haiku 5.5 · API

All dependencies (flag short deliveries after mobile, invoice matching after parser, credit notes after matching, accounting export after credit notes) are respected and explicitly named.

Tells the story of the betsRightRight

GPT-6 Luna · API

Groups work into two themes (invoice-to-credit automation and mobile receiving) tied directly to the churn reason.

Haiku 5.5 · API

The roadmap groups work into the themes of deliveries, invoices, and Calder retention, and explains why each is the most important investment now to cut churn and manage the renewal.

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 79% of 150 checks.

#	Model · Harness	Decision model	LLM judge	Runs	Critical failures
1	GPT-6.1 SolwithAPI	100.0	96.4	2	None
2	GPT-6 AstrawithChatGPT	93.9	100.0	2	None
3	Opus 5.5withClaude	90.2	96.4	2	None
4	Sonnet 5.5withAPI	94.1	92.6	2	None
5	GPT-6 LunawithAPI	92.0	85.4	2	None
6	Haiku 5.5withAPI	74.2	96.4	2	None
7	Gemini 3.8 FlashwithAPI	74.2	66.8	2	None
8	Gemini 3.5 Flash-LitewithGemini	56.9	36.5	2	1 capped

What you get for the cost

No setup scores within 5 points of the leader at half its cost or less here, so the leader is also the best value.

Nobody beats these on both score and costAt least one critical failure

Task score and relative cost by setup
Setup	Task score	Relative cost (cheapest = 1)
GPT-6 Luna · API	87	1.0×
Gemini 3.5 Flash-Lite · Gemini	44	1.1×
Haiku 5.5 · API	86	2.1×
Gemini 3.8 Flash · API	68	5.5×
GPT-6.1 Sol · API	98	6.9×
Sonnet 5.5 · API	94	36×
Opus 5.5 · Claude	94	37×
GPT-6 Astra · ChatGPT	97	39×

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

Turning strategy into a sequenced plan.

Why it matters

A roadmap is where strategy meets capacity. Dated feature lists turn guesses into promises.

What good looks like

Items are problems or outcomes, not just features
Sequencing reflects dependencies
Explicit trade-offs
Commitment falls with distance

Deliberately not measured

Gantt formatting

Capability tested

Sequencing under constraints

The failure we’re looking for

A dated wishlist sorted by excitement

Grading

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

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

Reliably right

Where it slips

How it’s graded

Outcomes, with certainty that falls with distance

Fits the stated capacity

Tells the story of the bets

Plus our standard checks

The tasks

The outputs

Output

H1 roadmap

Not doing this half

What we tell Calder Group

Grades and run details

Decision model checks

Run

Output

Fernhill roadmap, Q1 and Q2

What we're not doing, and why

What we tell Calder Group

Grades and run details

Decision model checks

Run

Check by check

Where they differ 2

All got right 10

Results

What you get for the cost

About the task

The PM job

Why it matters

What good looks like

Deliberately not measured