Skip to content

ndisisnd/msg / .claude/skills/plan-pm

plan-pm

Unverified

Principal PM skill — the autonomous PRD writer. Consumes a graded, fleshed-out row from the INTAKE.md backlog (idea, goal, type, grade) and drafts the full PRD solo — feature/acceptance table, edge cases, error handling — saved to features/planned/prd-[n]-[feature-slug]/. The requirements interview lives in /intake now, not here. Treats the row as settled and picks its own few follow-up questions (refs/principles.md); pauses ONLY for that one batch and for breaking/critical touches. --update edits an existing PRD's content and syncs its intake row; --migrate converts old PRDs to the current template. --review [on|off] runs plan-review at the end (this run, or saved per project). Refuses requests that would skip the PRD stage entirely.

plan-pm

Usage

Invoke: /plan-pm. With no args it lists the INTAKE.md backlog and asks which row to plan.

  • Slash commands: /plan-pm, /plan-pm #<n> (plan a specific intake row), /plan-pm --sub [parent PRD path | number], /plan-pm --update [prd path | number], /plan-pm --migrate [prd path | number | --all]

  • Natural language: "plan this idea", "draft a PRD", "write the PRD for the streaks feature", "turn backlog item 3 into a PRD"

  • Natural language (sub-PRD): "create a sub-PRD", "follow-up fixes for this PRD", "spin off a sub-PRD" — route to the --sub mode in the § Sub-PRD mode section below.

  • Natural language (update): "change PRD 3", "add a feature to this PRD", "drop F2 from the streaks PRD", "the PRD should also cover …" — route to § Update mode below.

  • Natural language (migrate): "migrate this PRD to the new template", "bring PRD 3 up to date with the template", "convert the old PRDs" — route to § Migrate mode below.

  • /plan-pm --review — run /plan-review on the new PRD at the end of this run. --review on / --review off (or "always review my PRDs") — save that for the project.

Auto-review switch. --review on|off or a plain-language toggle writes the switch and reports it, then carries on with any other request:

.claude/scripts/script-policy-set.py --file devkit/policy.json --set policies.plan_pm.review=<true|false>

No devkit/policy.json (a project set up before the file existed) → say /msg --update adds it, and use the value for this run only. The switch is on when the bare --review flag is set or policies.plan_pm.review is true; absent = off. On → after the digest, run Skill("plan-review", "<prd>") (default and --sub only), and its result replaces the plan-review suggestion. Off → the closing message suggests /plan-review.

Modes: default (autonomous top-level PRD from an intake row), --sub (a numbered follow-up nested under an existing parent PRD), --update (change what an existing PRD says), and --migrate (convert existing PRDs to the current template, no content change). When --sub is present — flag or sub-PRD natural-language trigger — read § Sub-PRD mode (--sub) first; all other steps run identically. --update and --migrate each replace the five-step protocol entirely. "Update" means a content change here, as it does in /intake.

Every PRD comes from an INTAKE.md row — there is no other entry path. Prose handed straight to plan-pm is captured into the ledger first (Step 1 calls /intake in capture mode), then planned from the row that comes back. Nothing is drafted against a row that does not exist.

Hard refusal: a request to skip the PRD and jump straight to engineering. State that plan-em requires a PRD and offer to draft one now (from a backlog row) or accept an existing PRD path for plan-em.

Pre-run — devkit reads

Before Step 1, stat-check and read the following in parallel via Bash. Written to devkit/ by /msg --init; CLAUDE.md stays at project root.

FileHow to apply
devkit/AHA.mdSurface relevant entries in the Open questions section; apply self-healing learnings (G5) to the draft so a category-tagged pattern (e.g. [tune:error-cases] …) is avoided this run
CLAUDE.mdExtract tech-stack constraints, conventions, architecture notes; validate feasibility of the drafted features and constrain the autonomous draft where the project setup already determines an answer
devkit/ARCHITECTURE.mdLoad system layers and integration points; validate feasibility and note conflicts in Open questions; source the platform detection in Step 3
devkit/OPEN-QUESTIONS.mdScan for unresolved decisions that constrain the draft; surface relevant entries in Open questions

devkit/GLOSSARY.md and devkit/DESIGN-SYSTEM.md are not read here: the PRD has no Glossary section (new terms are appended to devkit/GLOSSARY.md directly, where the whole pipeline sees them) and no User flow / Key user interactions sections, which were their only stated consumers.

Absent-file rule: If devkit/ does not exist, setup runs by itself first: run /msg --init, then continue this command (../shared/refs/setup-first.md). If an individual file is missing, emit <filename> not found — /msg --update adds it. Proceed without the file; do not create it. Do not ask the user about these files. Do not block. Proceed to Step 1 immediately.

Sub-PRD mode (--sub)

A sub-PRD is a numbered follow-up (prd-<n>.<m>) capturing extra changes/fixes to an existing parent PRD, nested inside the parent's folder, sharing the parent's branch. When --sub is present — flag or natural-language trigger — read refs/protocol-sub.md first and apply its deltas before Step 1 emits; every other step runs unchanged.

Update mode (--update)

--update changes the content of one existing PRD — features, scope, criteria, error cases. Read refs/protocol-update.md and follow it end to end; the five-step protocol below does not run in this mode. Once a PRD exists it is the source of truth and its intake row mirrors it: the run edits the PRD on the template, re-runs the shape gate, stamps reviewed: no so plan-review re-certifies, syncs the changed idea / goal / type / grade cells back to the row through script-intake-stamp.sh --set-cell (plan-pm is the second writer of row content, for PRD-linked rows only), and prints the digest with the changed parts marked.

Migrate mode (--migrate)

--migrate converges a PRD that already exists onto the current template — the one deliberate, user-invoked exception to the rule that nothing on disk is ever rewritten. It never changes what the PRD says. Read refs/protocol-migrate.md and follow it end to end; the five-step protocol below does not run in this mode. It takes one PRD (path or number) or --all.

Two things this mode is not:

  • It writes no INTAKE.md row and reads none. The intake-only rule governs where a PRD's idea comes from, and a migration introduces no idea — this is maintenance on an existing artifact. Explicit carve-out; the hard refusal against skipping the PRD stage is untouched.
  • It changes nothing about how PRDs are read. Every other skill still normalises the old shape in memory, and no writer emits it. The read path is exactly as it was.

The run is two layers: .claude/scripts/script-prd-update.py does the mechanical migration (frontmatter mapping, inline findings moved out to reports/, section renumbering, the §4 heading, the §1 success-signal bullet, AC1 labels), then the model does the style pass (§5 bullets → the four-column table, §1 back to its two bullets). It finishes on script-prd-shape.py --checks 1,2,3,4,5,6 — check 6 is opt-in and requested only here, so fresh-draft gates keep running 1–5.

Step-by-step protocol

Follow refs/protocol-pm.md end-to-end — it owns the steps, their order, and their outputs.

PRD digest (default / --sub): directly above the closing message, print the four-line digest from script-prd-digest.py --terminal — Title (PRD id + name), Change (who + outcome), Builds (feature names), Open (open-question count; omitted when 0). --update prints the same digest with the changed parts marked; --migrate prints no digest, only a one-line Converted <n> PRD(s)… count.

Closing message (default / --sub, every outcome; --update closes per refs/protocol-update.md § Step U6; --migrate per refs/protocol-migrate.md § Step U5): end the run with the closing message per ../shared/refs/closing-message.md — the last chat output, after Step 5's termination output. Two What happened rows are mandatory on every completed run: the PRD path, written verbatim so it can be copy-pasted, and Open questions left: <n> counted from the drafted §5. That count drives the protocol deterministically — 0 ⇒ 🟢, >0 ⇒ 🟡. The one-line "what this run did" slot may run to 2–3 lines here, summarising the feature drafted. Take the next step from the registry's plan-pm row; never compose it.

Harness incidents (all modes): log unexpected script failures, tool errors, retries, and missed writes to devkit/DOCTOR.md per ../shared/refs/doctor-logging.md — logging never changes what the run does next.

PRD status lifecycle

The lifecycle — the frontmatter field table, the status-vs-lane rule, the pre-v5.4 normalisation, and the intake ledger stamp (F4/D14: plan-pm stamps the source INTAKE.md row in-progress + prd cell at Step 5 via script-intake-stamp.sh) — lives in ../shared/refs/prd-lifecycle.md, the one home. plan-pm writes a new PRD status: backlog and stamps nothing else; every stamp goes through the shared writers, never improvised Bash.

References

  • refs/protocol-pm.md — end-to-end five-step autonomous protocol; followed from § Step-by-step protocol
  • refs/protocol-sub.md — the --sub deltas layered over that protocol; read from § Sub-PRD mode when --sub is set
  • refs/principles.md — plan-pm's own rules for which follow-up questions to ask (never re-ask or widen the row; send scope changes to /intake --update); read in Steps 1 and 4
  • refs/protocol-update.md — the standalone --update protocol (resolve + lock gates → edit on the template → shape gate → reviewed: no → re-grade → row sync → marked digest + close); read from § Update mode when --update is set
  • refs/protocol-migrate.md — the standalone --migrate protocol (target resolution → L1 → style pass → checks 1–6 gate → close); read from § Migrate mode when --migrate is set, and it replaces the five-step protocol rather than layering over it
  • .claude/scripts/script-prd-update.py — deterministic v5 → v5.4 PRD migrator (L1); called per target in --migrate Step U2
  • .claude/scripts/script-prd-digest.py --terminal — the four-line PRD digest printed above the closing message (--mark flags the parts an --update changed)
  • .claude/scripts/script-prd-scan.sh — deterministic lane-aware PRD inventory (JSONL); call in Step 2's prior-PRD scan
  • refs/template-prd.md — structured PRD format (seven sections) and its drafting rules, including the §3 F-ID and acceptance-criteria contract and the §4 error-and-edge-case format; used to initialize the file in Step 3
  • refs/error-checklist.md — the standard failure types (Tier 1 always, Tier 2 when the trigger fits) every §3 feature is checked against when §4 is drafted
  • .claude/scripts/script-prd-shape.py — deterministic PRD shape validator; call at the end of Step 3 Part 4
  • .claude/scripts/script-prd-number prd <row-#> — PRD number = intake row number; exit 3 NUMBER_CLASH when that PRD already exists; call in Step 3 Part 1
  • .claude/scripts/script-prd-number sub <parent-n> — deterministic next sub-PRD minor m (counts sub-PRD folders and #<n>.<m> ledger rows); call in --sub D2, before the row is captured (see § Sub-PRD mode)
  • INTAKE.md — the root backlog ledger written by /intake; the only source of a PRD's idea — read in Step 1 to resolve the row (calling /intake to capture one first when the user came in with raw prose), stamped in Step 5
  • devkit/ — project-level agent context directory created by /msg --init; contains AHA.md, GLOSSARY.md, ARCHITECTURE.md, DESIGN-SYSTEM.md, OPEN-QUESTIONS.md
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 467kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 319kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 319kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 319kUpdated Claude CodeCodexVerifiedInstall