ready-check

name

ready-check

description

The PM-side readiness gate — "airport security for PRDs." A PM (or anyone) gives a PRD as any source — JIRA / Linear / Notion / Confluence / Google Doc / Slack link / GitHub issue / Figma URL, pasted text, an image, or a plain description — and the agent scores it against the Definition of Ready, surfaces the blockers the PM can fix themselves in plain language, and (when ready) mints a tamper-evident gate code + a clean hand-off for dev. Runs identically in Claude Code and Cowork. Use when the user says "ready check", "is this PRD ready", "check my PRD before dev", "gate this requirement", or invokes `/ready-check <source>` / `/ready <source>`. This is Level 1; the dev-side deep pass is `/contract pickup`.

Ready Check (PM-side gate)

The world this is built for: a PM has a PRD and is about to hand it to engineering for grooming and estimation. Today, half-baked PRDs get through, dev chases the PM for basics, and dev eats the delay. Ready Check is the cheap, 2-minute check the PM clears first. Green → a gate code + hand-off; anything missing → the exact gaps, in plain language.

It is deliberately a floor, not the deep judge. The precise, code-grounded review (regression, security, feasibility, estimate) is Level 2 — skills/contract-pickup, dev-side. Read reference/READY-CHECK-RUBRIC.md for the rubric and the PM-vs-dev dividing line.

Works for EVERY PRD — frontend, backend, API, data, infra

This gate is stack-neutral. It is not a UI-only tool, and there is no "this is backend, the tool doesn't apply" escape. Every rubric item has a sensible reading for a pure-backend or API PRD — treat the item, not the label:

item	frontend reading	backend / API / data reading
design	Figma / mock link	interface / contract spec — OpenAPI, proto, GraphQL schema, event schema
states	empty / loading / done / error copy	response & failure contract — 2xx/4xx/5xx codes, error bodies, ret/idempotency
flows	screens touched	endpoints, jobs, queues, consumers touched
scope	platforms (iOS/Android/web)	services in scope, and what's explicitly out
l10n	copy final & localised	response/payload format finalised (schema version, error codes)
writer	writer/creator surface	downstream consumers (services, webhooks, subscribers)
events	analytics events	metrics / logs / traces (latency, error-rate, throughput)

A backend PRD is not cleared by N/A-ing the UI items — it's cleared by answering the backend equivalents above. If a PM says "it won't work, this is backend," that's the signal to run it with these readings, not to skip it.

Runs in both surfaces

This skill is plain markdown + one deterministic script, so it behaves identically in Claude Code and Cowork. The offline web page (gate/prd-readiness-gate.html) is the same rubric for PMs with no Claude open; it mints the same gate-code format, so a code from any surface verifies in all of them.

Opening & auto-syncing the live panel in Cowork (the artifact)

Prefer the chat flow. Running /ready-check <link> in chat is the primary, most reliable experience — it renders a live, panel-like scorecard (Phase 4) and fixes are a one-line reply. It has none of the sandbox limitations of the artifact (no per-user connector wiring, no snapshot/version drift, no flaky in-page bridge), so lead with it. The sidebar panel below is an optional convenience for PMs who want an always-open surface; when a PM hits panel trouble (a link won't fetch, a stale build), the answer is "run `/ready-check

` in chat" — it just works.

In Cowork, you can also give the user a persistent sidebar panel — the artifact built from gate/ready-check-cowork.html, where the deterministic gate runs in-page and the edge-case review is done by Claude via window.cowork.askClaude (no backend). The plugin can't ship the artifact itself (plugins bundle skills/commands, not artifacts), so create it on demand.

The pinned panel is a saved snapshot, not a live link to the plugin file. Cowork stores a copy of the HTML when the artifact is created and never re-reads the plugin file on its own — so the panel does not magically update when the plugin updates. It only refreshes when THIS skill reconciles it with mcp__cowork__update_artifact. To make that reliable, run the reconciliation below at the start of every Cowork invocation of ready-check (not only /ready-check panel / "open the panel") — that way the pinned page self-heals to the current build whenever the PM uses Ready Check. It's cheap: one list_artifacts plus at most one Read.

Reconcile (create-or-update, id ready-check):

1. List. Call mcp__cowork__list_artifacts; find the artifact whose id is ready-check and note its path.

2. Compare build numbers — read them from the file contents, not the path. Both the plugin's bundled gate/ready-check-cowork.html and the stored artifact carry <meta name="ready-check-build" content="N">. Read the bundled file and, if the artifact exists, Read the file at its path, and parse the integer N from that meta tag in each. (Reading N "from the path string" — the old instruction — never worked; that's why the panel went stale. You must open the file and read the tag.)

3. Act on the comparison:

   • No ready-check artifact yet → create it (fields below).
   • Bundled N > stored N → the plugin shipped a newer panel: this is the step that was being skipped. Call mcp__cowork__update_artifact with id ready-check, html_path = the baked scratch file (the bundled panel with __RC_ATLASSIAN_MCP__ replaced by THIS user's discovered prefix — see the bake step under the create fields), a one-line update_summary (e.g. "Ready Check panel → build N"), and re-pass mcp_tools with THIS user's Atlassian tools. Both matter: the id is unique per user, so a panel that shipped with someone else's id (or none) can't fetch links until it's re-baked + re-granted here. Approve and the pinned page updates in place.
   • Stored build unparseable / missing tag (a pre-build-tag artifact) → treat as older and update it.
   • Equal (or stored ≥ bundled) → do nothing; only mention "already current" if the user explicitly asked to open/refresh the panel. Never create a second artifact with the same id.

   Discover this user's Atlassian prefix first — ToolSearch for getJiraIssue and take its mcp__<id>__ prefix. The connector id is unique per user, so this must be discovered in THIS user's session, never hardcoded.

   BAKE that prefix into the panel HTML before creating/updating — this is what makes link-fetch work reliably for every user (not just whoever's id was shipped). Don't pass the bundled file directly: Read gate/ready-check-cowork.html, replace the literal token __RC_ATLASSIAN_MCP__ with the discovered prefix (e.g. mcp__abc123__), Write the result to a scratch file, and use THAT scratch file as html_path for both create_artifact and update_artifact. (If Atlassian isn't registered, leave the token as-is — the panel then cleanly prompts "paste the text / use /ready-check in chat" instead of calling a wrong id.) A manual setAtlassianConnector('mcp__<id>__') in the panel console still overrides.

   Then call mcp__cowork__create_artifact with:

   • id: ready-check
   • html_path: the baked scratch file (not the bundled file)
   • description: "Ready Check (live) — PM-side PRD readiness gate; smart edge-case review via Claude, no backend."
   • mcp_tools: the concrete per-user tool names — <ATLASSIAN>getJiraIssue, <ATLASSIAN>getAccessibleAtlassianResources, <ATLASSIAN>getConfluencePage, <ATLASSIAN>editJiraIssue, <ATLASSIAN>createJiraIssue, mcp__figma-rest__get_file_version, mcp__figma-rest__audit_design. Drop any whose connector the user hasn't registered. The baked prefix and the mcp_tools you pass must be for the SAME connector.

4. Tell the user the panel is in their Cowork sidebar (created / refreshed / already current) and persists across sessions. Worth surfacing what it does:

   • Deterministic score & gate — checking a PRD drafts answers from the PRD (each marked ✨ to review) and scores them; the result is cached per PRD (by content hash), so re-checking the same PRD always returns the same score — no drift. Edit the PRD to re-analyze. The gap findings are a helper (advice); a slow review bridge can't lock the PM out, and a cached PRD re-checks instantly with no bridge call.
   • Playbook (header) — good-vs-weak examples for all 13 fields, plus an inline "see example" on each field.
   • Verify (header) — dev pastes a hand-off → authentic / tampered / not-ready / invalid. In the PRD field paste a JIRA/Confluence link and it fetches the live doc to check source-version drift + content hash. Same freeze check as pickup, one click.
   • Draft-from-idea — "✨ Only have an idea?" drafts a starter PRD skeleton.
   • Readiness score ring + strong/weak sample buttons; green/red status dot.
   • Design audit — the panel is a sandbox and cannot reach the local figma-rest connector, so when a Figma link is present it tells the PM to run the design audit in chat (/ready-check <figma-link>) or dev-side pickup. The audit itself (platforms, states, placeholder copy) runs there — see Phase 3 item 6. The panel is deliberately just the gate + hand-off (checklist, score, smart review, JIRA/Confluence link fetch, gate code) — it no longer generates or publishes a PRD; dev picks up the hand-off with /contract pickup.

5. Prompt for the connectors the panel uses, and note which are missing so the user can authorize them in Cowork connector settings:

   • Atlassian — powers JIRA-issue & Confluence-page fetch and publish (update a linked ticket / create a new one). Without it the panel still works on pasted text; links just can't be fetched or published.
   • Figma REST — (for the v0.22 design-freeze: validate a design link, capture its version, export a snapshot). Note: the Figma Dev Mode MCP is NOT the one to use here — it needs the desktop app with the file open and exposes no file version, so it can't validate arbitrary links or pin versions. Design-freeze needs a Figma REST connector / token. The panel ships with ATLASSIAN_MCP and FIGMA_MCP (mcp__figma-rest__) set. If a user hasn't registered one of those connectors, blank its CONFIG line at the top of the panel HTML and drop its tools from mcp_tools so the panel degrades gracefully instead of erroring.

Bump the ready-check-build number in gate/ready-check-cowork.html whenever you change that file, so the reconciliation above can tell "newer" from "same." Two things gate whether a PM actually sees a new panel: (1) the plugin must be updated in their Cowork — the pinned page is only ever as new as the bundled gate/ready-check-cowork.html; if the plugin is still on an old version there's nothing newer to push. (2) the skill must run so it can call update_artifact — Cowork does not watch the file, so "auto" here means "on your next /ready-check", not "instantly on plugin update." If a PM is stuck on an old panel, have them update the Manifest plugin, then run /ready-check once (or /ready-check panel) to pull the latest build.

Only works in Cowork (the artifact bridge is Cowork-only). In Claude Code there is no sidebar panel — the /ready-check chat flow is the way. If the bundled HTML can't be found, fall back to telling the user to open gate/ready-check-cowork.html from the plugin folder directly.

Optional — JIRA/Confluence-link fetching in the panel. The panel can fetch a pasted Atlassian link through the user's connector, but the connector-id is an opaque per-install hash, so it must be discovered, not hard-coded. When you create/update the artifact:

1. Discover the prefix. ToolSearch for getJiraIssue (and getAccessibleAtlassianResources). Take the returned tool name and extract its mcp__<connector-id>__ prefix.
2. Pass the tools in mcp_tools (…getJiraIssue, …getAccessibleAtlassianResources, …getConfluencePage, and, if publishing, …editJiraIssue / …createJiraIssue).
3. Point the panel at that prefix. The panel reads the prefix from localStorage['rc-atlassian-mcp'] (falling back to a documented default), and exposes setAtlassianConnector('mcp__<connector-id>__') for a dev to set it by hand from the console. Tell the user the one-liner if their links aren't being picked up. Broaden-URL handling now also accepts sharable/Rovo links and links with a little surrounding text, so a plain browse/KEY isn't required.

If no Atlassian connector is registered, Atlassian links simply prompt the PM to paste the text or use /ready-check <link> in chat (which fetches via the connector regardless of the panel).

The engine vs the judgement

Two layers, mirroring how contracts work (validate.mjs + critics):

• Deterministic engine — scripts/ready-check.mjs. Scores presence + specificity of every required item, mints/verifies the gate code, renders the hand-off. Never an LLM guess. You MUST call it for the verdict and the code — do not eyeball readiness or invent a code.
• Judgement — you. On top of the engine, run the text-judgeable critics so the PM sees real blockers, not just empty fields: is the goal measurable, are obvious edge cases missing, are the UI states named, is the metric a number. This is what makes the check smart instead of a checklist.

Process (5 phases)

1 Fetch + expand   Read the PRD from whatever source was given
2 Map to answers   Fill the 11 items from the text; ask the PM only what's missing
3 Judge            Run text-judgeable critics; flag blockers in plain language
4 Score + mint     Call ready-check.mjs → verdict; if ready, gate code + hand-off
5 Hand off         PM pastes the hand-off into the ticket / dev thread

Phase 1 — Fetch + expand

Accept any input: a URL (JIRA, Linear, Notion, Confluence, Google Doc, Slack message, GitHub issue, Figma), pasted text, an image (screenshot of slides / a Figma frame), or a plain description. Fetch it with the matching MCP or reader. If it's a Figma link, treat that as the design artifact for item 3. If multiple sources are pasted, use them all.

Fresh fetch, every run — never answer from a previous PRD. Each invocation is about the source given right now. Follow this every time, with no shortcuts:

1. Fetch the current source before you judge anything. Do not emit a verdict, score, or gate code until you have actually fetched THIS source in THIS run and can quote specifics from it. A confident-looking answer that isn't grounded in a fresh fetch is the bug PMs reported ("it showed results for the previous PRD").
2. Echo what you fetched back to the PM first — the issue key / page id and the title (e.g. "Fetched ENG-1421 'FTC enhancement — pre-auth flow'; analysing this one."). This lets a wrong or stale doc get caught immediately.
3. A new link/source discards the previous one. If the PM gives a different PRD than a moment ago, start clean — do not carry over the earlier PRD's answers, findings, or verdict. When in doubt, treat it as a brand-new check.
4. If the fetch fails, say so and stop — ask the PM to paste the text or connect the source. Never fall back to answering from memory or an earlier run.

Discover the connector before deciding it's missing (Cowork especially). In Cowork, connector tools are deferred — they do not appear in your tool list until you load them with ToolSearch. So a Confluence/JIRA/Notion link does not mean "not connected" just because you don't see the tool yet. When the source is such a link:

1. Search for the fetch tool first. Call ToolSearch with the obvious keywords before concluding anything — e.g. confluence page, getConfluencePage, jira issue, getJiraIssue, notion page. If a matching tool comes back, it's connected; load it and fetch. (Atlassian typically exposes getConfluencePage, searchConfluenceUsingCql, getJiraIssue, getAccessibleAtlassianResources under a mcp__<connector-id>__… prefix; the <connector-id> is an opaque hash, so match by the tool name, not a guessed prefix.)
2. If ToolSearch returns nothing, then the connector genuinely isn't registered — now ask the PM to paste the text or upload the doc, and mention they can connect Atlassian in Cowork connector settings to skip this next time.
   • Sharable JIRA links work too, not just browse/<KEY>. Extract the issue key by scanning the URL for the first [A-Z][A-Z0-9]+-\d+ (covers selectedIssue=, /issues/<KEY>, browse/<KEY>), then getJiraIssue. See reference/JIRA-SYNC.md.
3. Never say "X isn't connected" without having run ToolSearch for it. That was the failure mode — reporting a connector absent when it was only deferred.

If nothing usable is given, ask the PM the 11 questions directly — the rubric doubles as an interview.

Phase 2 — Map the text to the 11 items

For each rubric item, pull the answer from the source if it's there. Build an answers object:

{ "title": "<feature name>",
  "items": {
    "goal":   { "detail": "..." },
    "design": { "detail": "https://figma.com/..." },   // or { "na": true } if no UI
    "oldbeh": { "na": true },                            // brand-new feature
    ... } }

Only skippable items (design, oldbeh, states, writer) may take N/A — and N/A is never a free pass. A bare { "na": true } no longer clears the item; it only clears as { "na": true, "reason": "<why it genuinely doesn't apply>" } with a reason ≥ 3 chars, which rides into the hand-off for dev to accept or push back. Prefer the stack-appropriate answer over a skip: for a backend/API change, design is the interface/contract spec (OpenAPI, proto, schema) and states is the response & failure contract (status codes, error bodies) — answer those rather than marking them N/A. Reserve N/A for items that truly have no analogue (e.g. a pure internal cron with no consumers → writer N/A "no downstream consumers"). For everything you can't find, ask the PM in plain words, one short batch. Do not invent answers.

Waivers — when the PM genuinely can't answer a required item. Any required item may be waived with a written reason (e.g. the metric is owned by another team, a decision is pending). Record it as { "waived": true, "reason": "..." }:

"metric": { "waived": true, "reason": "owned by growth; target decided this week" }

A waiver clears the item so the PRD can move, but the engine hashes the reason into the code and surfaces it in the hand-off under a "Waivers (dev to accept or push back)" block. An empty/absent reason does not clear. Offer a waiver only when the PM truly can't answer — not as a shortcut past a question they could answer.

Phase 3 — Judge (make it smart — actually find the edge cases)

This is the heart of the check, and it must go beyond "is the field filled." Actively run the text-judgeable critics on the PRD and surface real findings the PM can fix themselves — do not just confirm presence.

1. Edge cases — always run the critic-edge-cases critic on the PRD. Enumerate the failure/boundary cases this specific feature implies and check each against the PRD: empty, offline, timeout, expired, concurrent edit, permission denied, partial failure, mid-flow cancel, and any domain-specific ones. List every case the PRD doesn't cover, one plain line each. A PRD whose edge answer misses cases the critic finds is not ready just because the field is non-empty — reopen it.
2. Clarity (critic-minimality) — is the goal a real problem and the metric an actual number with a window? "Improve engagement" fails; "+6% D1 retention in 4 weeks" passes.
3. UI states (critic-comms-completeness) — for a user-facing change, is there copy for nothing-yet / loading / done / error?
4. Instrumentation (critic-instrumentation) — will the stated events actually let us measure the success metric? If not, name the gap.
5. Scope / parity — explicit out-of-scope present? Behavior consistent across the stated platforms?
6. Design audit (when a Figma link is present). If the design item is a Figma URL and the figma-rest connector is available, call its audit_design tool (tool name depends on install — e.g. mcp__figma-rest__audit_design or mcp__Figma_REST_Ready_Check__audit_design; discover it via the available tools). Read the file and fold the gaps in:
   • Platforms — if the PRD scopes mobile but the file has no mobile-width frames (only desktop/tablet), that's a blocker on scope. Web-only when the PRD implies mobile → flag it.
   • UI states — if no error / empty / loading screen is found by name, surface each as a warning on states (frame-naming is heuristic — say so, and ask the PM to confirm those states exist rather than hard-failing).
   • Copy — any placeholder/lorem/TODO text layers are a blocker on design ("N text layers still contain placeholder copy"). This is the design-side of Ready Check. It runs here, in chat (and at dev-side /contract pickup) — the Cowork panel is a sandbox and can't reach a local connector, so the panel defers the design audit to this flow. Say which frames/counts you saw so the PM can trust it.

Present these as a short, plain-language blocker list the PM can act on immediately. The point of Ready Check is that the PM sees — and fixes — the edge cases before dev, not that dev finds them later.

Guardrail: don't push code-level decisions onto the PM (schema choices, security design, perf strategy). Surface the product-level gap; if the answer needs the repo, note it for Level 2. Finding a missing edge case is the PM's job here; deciding its implementation is dev's.

Access modes — the depth scales with what's available

Run at the deepest mode you have access to; each is a strict superset:

• Text only (always works). Judge purely from the PRD. Catches most product-level edge cases and every completeness gap. No setup.
• + code-context cache. If .manifest/.cache/code-context.json exists, call ready-check.mjs --cache-check <answers.json> .manifest/.cache/code-context.json and fold the notes in — event-naming drift, a flow that touches more surfaces than the PRD lists, known past failures on this surface. No repo credentials needed.
• + repo read access (opt-in). If the PM/session has read-only access to the product repo(s) in .manifest/repos.yml, ground the edge-case pass in real code: check that the flows named actually match the surfaces in the code, and pull historical edge cases from prior handling of the same screen. This is the same evidence Level 2 uses, run earlier — so PMs can catch more themselves. Grant this deliberately: read-only, per the team's policy; it's optional, and the check degrades cleanly to the modes above without it.

Whatever mode you're in, say which one you used, so the PM knows how deep the edge-case check went.

Phase 4 — Score + mint (deterministic), rendered as a live scorecard

Chat is the primary experience — render a panel-like scorecard, not a wall of prose. The chat flow is the reliable one (full connector access, the real model, no sandbox), so make it feel like the panel: one glance to read, one reply to fix. Write the answers to a temp JSON and run the engine:

node scripts/ready-check.mjs --scorecard answers.json   # readable panel-like scorecard
node scripts/ready-check.mjs --check     answers.json   # verdict JSON, exit 1 if not ready
node scripts/ready-check.mjs --handoff   answers.json   # the hand-off block (includes the code)

Show the --scorecard output as-is — it's already the panel: a progress bar, the verdict, a numbered "To fix" list with a concrete prompt per gap, the grouped "Answered / N/A / Waived" items, and (when ready) the gate code. Pass meta.source (e.g. JIRA FTC-1421 / Confluence <id>) so the card names which PRD it is. Then overlay the smart findings from Phase 3 (specific missing edge cases / vague answers) as a line or two under the relevant numbered item, so the PM sees what to fix, not just that something's missing.

The card already invites the fix — just run the loop. The scorecard ends with a numbered gap list and an "↳ e.g. …" reply hint, so the PM can answer by number or in free text. When they reply, fold their answers into the answers object, re-run --scorecard, and show the updated card — the progress bar fills and the score climbs. Accept free-text ("metric is +6% D1 in 4 weeks"), a numbered list ("2: …; 4: …"), or N/A: <item> — <reason> / waive <item> — <reason>. Keep looping until it clears. Read like a panel, fix like a chat.

• Not ready — show the scorecard + the fix prompt above. Never mint a code.
• Ready — show the scorecard (now ✅), then the hand-off block from --handoff. The gate code is in it.

Phase 5 — Hand off to dev

Tell the PM exactly where the block goes: the JIRA ticket, or the dev Slack thread. If a Slack/JIRA MCP is connected and the PM asks, post it for them in their voice. The block carries the design link and the gate code so the feature and its design travel together.

Generate the PRD + publish to JIRA (the "right format")

Ready Check holds the structured answers, so it can also author a clean, publish-ready PRD — the same 11-section format as gate/examples/PRD-good-*.md:

node scripts/ready-check.mjs --prd answers.json    # prints the 11-section PRD (markdown)

The PRD carries the gate code and lists any waivers under "Waivers (dev to accept)". Offer it once the PM has the basics in — even Not-Ready renders (it marks missing sections), but it's most useful at Ready.

Publish (ask the PM which):

• Update the linked ticket — if the PRD came from a JIRA link, write the PRD into that ticket's description via the Atlassian MCP editJiraIssue (fields.description, contentFormat: "markdown"). Safest — no new tickets.
• Create a new ticket — createJiraIssue with projectKey, issueTypeName (ask; default Story/Task), summary = the feature title, description = the PRD. Confirm the project key with the PM first; never guess it.
• Copy only — hand the PM the PRD markdown to paste in themselves.

Always show the PM the PRD and confirm before writing to JIRA. (The Cowork panel exposes the same three via a Generate PRD / Copy / Publish row.)

Enforcement — "no code, no grooming" (automatic)

The gate code is content-bound (a djb2 hash of the answers), so it's checkable, not decorative. Dev runs nothing extra. When the dev runs /contract pickup <ticket>, pickup reads the Ready-Check: line from the ticket/hand-off and verifies it automatically, then acts:

• VALID → proceed; record readyCheck: RC-… in the contract frontmatter.
• STALE (PRD changed after it cleared) → stop and ask the PM for a fresh Ready Check.
• INVALID / missing → refuse pickup: the PRD never passed the gate.

So the only person who ever thinks about the code is the PM (they paste the hand-off once). Dev just picks up as usual and the gate enforces itself. The manual ready-check.mjs --verify <handoff> exists as a fallback for anyone who wants to check by hand, but the normal flow needs no command.

Freeze check (v0.22) — did the PRD or design change after sign-off?

When the PRD was fetched from JIRA/Confluence (and, if Figma REST is wired, has a design), the hand-off carries a freeze stamp: Source: <type> <id> v<version>, optionally Design-Version: <n>, PRD-Hash: H…. At /contract pickup, after verifying the gate code:

1. Re-fetch the current source version (JIRA updated / Confluence version.number) and, if applicable, the Figma file version (mcp__figma-rest__get_file_version).
2. Write them to a temp JSON { "version": …, "designVersion": …, "prdHash": … } and run:

   node scripts/ready-check.mjs --verify-freeze <handoff.txt> current.json
   

3. It prints VALID / STALE-SOURCE / STALE-DESIGN / STALE-CONTENT.
   • VALID → proceed.
   • any STALE-* → stop: the PRD doc or the design changed since it cleared. Ask the PM to re-run Ready Check and re-sign.
   • No freeze lines in the hand-off → skip (older hand-off; nothing to check).

This is what makes "the PRD/design can't quietly change after the gate" real — you detect the drift at the moment dev picks it up.

Scenarios to handle

• Backend / API-only feature — do not just N/A the UI items. Answer their backend equivalents: design = the interface/contract spec (OpenAPI, proto, schema), states = the response & failure contract (status codes, error bodies, retry/idempotency). Only mark an item N/A when it has no analogue at all, and then only with a reason (e.g. writer → N/A "internal job, no downstream consumers"). Everything else is required, read the backend way.
• Brand-new feature — mark oldbeh N/A (nothing exists yet).
• No writer/creator impact — mark writer N/A with that reason.
• Figma-only hand-off — the Figma URL satisfies design; still need the other 10.
• Vague "make it better" PRD — expect Not Ready; the value is showing the PM precisely what's missing.
• Re-check after edits — re-run; a changed answer re-mints the code, so an old hand-off correctly goes STALE.
• No code-context cache — the check runs on text alone; skip the cache notes without erroring.
• Can't answer an item — waive it with a reason ({ "waived": true, "reason": "..." }). It clears but shows in the hand-off's Waivers block for dev to accept or push back. Empty reason doesn't clear.
• N/A a skippable item — same rule: { "na": true, "reason": "..." }. A bare N/A no longer clears; the reason rides into the hand-off. Prefer the backend-equivalent answer over N/A wherever one exists.
• Not a PM — a dev pre-checking their own pickup can run it too; same gate.

Anti-patterns

• Don't eyeball the verdict or hand-write a gate code — always run the engine.
• Don't let a PM N/A a non-skippable item to slip through — and don't accept a bare N/A on a skippable one either; it needs a reason, or better, the backend-equivalent answer.
• Don't treat a backend/API PRD as out of scope — run it with the backend readings in "Works for EVERY PRD" above.
• Don't emit a verdict from a previous run — fetch the current source first and echo back what you fetched (issue key + title) before judging.
• Don't demand implementation detail from the PM (that's Level 2's job) — keep blockers to product decisions judgeable from text.
• Don't fetch source through non-approved means if an MCP fails — say so and ask the PM to paste the text.