playtest-prd-prep
Reconcile one validated Continuity Loom author-playtest report against live repo and tracker state, preserve proven strengths, and write one same-stem change portfolio for playtest-to-issues custody before any later /to-prd pass. Invoke explicitly with a reports/playtest-*.md path; this skill diagnoses and prepares but does not implement or publish.
- License 许可
- GPL-3.0
- 最近更新
- 2026-11-10
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- playtest-prd-prep
- description
- Reconcile one validated Continuity Loom author-playtest report against live repo and tracker state, preserve proven strengths, and write one same-stem change portfolio for playtest-to-issues custody before any later /to-prd pass. Invoke explicitly with a reports/playtest-*.md path; this skill diagnoses and prepares but does not implement or publish.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档agents/openai.yaml285 B
- 参考文档references/prep-format.md16.8 KB
- 脚本scripts/validate-prd-prep.mjs48.8 KB
- 脚本scripts/validate-prd-prep.test.mjs32.4 KB
03 SKILL.md 正文
Playtest PRD Prep
Reconcile a .claude/skills/playtest report into one PRD-ready portfolio. The stopping point is
reports/<playtest-report-stem>-prd-prep.md. Do not edit the source report, product code, tests,
docs, skills, specs, tickets, or tracker state; do not invoke /to-prd or satisfy its seam
checkpoint. The closeout must route the completed artifact through $playtest-to-issues before
recommending /to-prd.
Input
Require one canonical reports/playtest-<story-slug>-<timestamp>.md path or an unambiguous glob.
Exclude evidence assets and derivative *-prd-prep.md files. If the path does not resolve, search
case-insensitively under reports/ and proceed only when one report resolves.
Validate the report and inspect its downstream inventory before trusting it:
node .claude/skills/playtest/scripts/validate-report.mjs --report <report>
node .claude/skills/playtest-prd-prep/scripts/validate-prd-prep.mjs --inspect-source <report>
If schema, privacy, or ledger errors make exhaustive disposition unsafe, stop without writing or
updating the prep artifact. The source inspector may downgrade a bounded prompt-evaluation metadata
or disclosure mismatch to a non-blocking validator defect when the finding inventory remains
complete; carry that defect into source validation and skill-maintenance follow-up. Identify every
report or playtest-methodology repair and route skill repair through $skill-audit; never repair
the source during prep. Missing optional evidence, stale commits, unavailable tracker reads, or
failed reproduction do not invalidate the source: continue with explicit limitations and classify
affected items as verification work.
Complete intake only when the canonical source path, source validation, report inventory, run mode, prior-report pointer, prior-report prep state, current same-stem prep state, branch, HEAD, and baseline worktree dirt are known.
Reconcile
1. Establish live provenance
- Retain exact
git status --short --untracked-files=all,git branch --show-current, andgit rev-parse HEADoutput. Treat pre-existing dirt as unowned unless this invocation explicitly adopts it. - Read
AGENTS.md,docs/ACTIVE-DOCS.md,docs/FOUNDATIONS.md, and the active domain authorities for every affected surface. Readdocs/agents/issue-tracker.md,docs/agents/triage-labels.md, andtickets/README.mdbefore assigning destinations or label posture. Treatarchive/**only as explicitly routed historical evidence. - Read the complete named report. Use its Cumulative Finding Ledger as the inventory frontier.
Follow the explicit
prior_reportchain only foropen,repeated, ornot-retestedrows whose decision-relevant detail is absent from the frontier report. Never scan unrelated reports. - Compare the report's launch HEAD with the current tree over affected paths. A changed HEAD is not itself drift; distinguish relevant committed drift, relevant worktree drift, and unrelated report/skill drift.
- Classify the source report's durability using tracked, clean, publication-ref-visible, and publication-ref-content-identical evidence. Summarize machine-local evidence; never cite its path as durable.
- For a continuation with a named
prior_report, derive that report's same-stem*-prd-prep.mdpath. If it exists, read it as historical recommendation input, never as product authority. Reconcile its former first action, every PRD candidate, and every Non-PRD Follow-Up item asconsumed,still live,rejected, orsupersededagainst the frontier report, current authorities, implementation, and tracker state. If it is absent, recordmissing at intake. - Check the current report's default same-stem prep path. If it exists, classify it as current,
partially consumed, stale, superseded, or not relevant. Reconcile every prior recommendation as
consumed,still live,rejected, orsupersededbefore reusing it. - Search current open work and relevant closed tracker work from exact report terms and IDs. Start with projected metadata; fetch full bodies or comments only for likely owners. If tracker access is unavailable, record the limitation and do not claim AFK-ready label certainty.
This step is complete only when every authority and freshness source needed to judge the report is checked or explicitly unavailable, and every recommendation from an existing current same-stem or prior-report prep has a current disposition.
2. Disposition every report row
Give every Cumulative Finding Ledger row exactly one disposition from the output contract. Current Prioritized Findings must already occur in that cumulative ledger; never invent a parallel ID.
Treat report rows as credible observations, not automatic mandates. Promote an observation into a stable change only when current source, tests, repeated evidence, deterministic reproduction, or a governing authority supports the desired rule. Preserve one-run ambiguity and model variability as verification, coverage, or research follow-up unless a safe deterministic mitigation is justified.
Use this evidence ladder:
- report detail and privacy-safe retained evidence;
- current authorities, implementation, and focused tests;
- exact open/closed tracker overlap and closeout evidence;
- focused read-only probes; then
- a targeted browser reproduction only when drift or contradiction still blocks classification.
For a targeted browser reproduction, read
../playtest/references/browser-driver.md, reuse the
blank-key guarded localhost boundary, and keep diagnostic state under /tmp. Do not start another
full author journey, accept prose, click provider controls, or make OpenRouter requests.
Every strength / preserve-strength row is a preservation constraint. Map it to global or to
the candidates it constrains and name regression evidence. Do not turn a strength into new scope.
This step is complete only when every source row has one disposition, current evidence, a change destination, and PRD impact; every strength has one preservation row; and no uncategorized blocker remains.
3. Determine the change portfolio
Follow each report-derived rule across every mechanically affected code, test, active-doc, and skill surface. Exclude adjacent issues that do not trace to a report row, preserved strength, or required consistency consequence.
Route work to its smallest governed destination:
- broad or risky product behavior becomes a PRD candidate and names any required spec, ADR, active doc, test, or skill changes;
- narrow, already-scoped product work becomes a ticket candidate;
- pure playtest-methodology or report-schema work becomes skill maintenance;
- an active authority that misstates current behavior becomes a doc correction;
- contradictions with closed work become verification/reopen work;
- insufficient evidence becomes coverage or research follow-up; and
- covered, resolved, preference-only, or disproportional work becomes covered or rejected.
Bundle PRD rows only when they share the same product rule, decision point, implementation seam, and acceptance proof. Default to one recommended first PRD plus deferred follow-ons. Use a multi-PRD program only when independently implementable outcomes have a real dependency or require one shared ratified direction. It is valid to recommend no new PRD.
Rank the first PRD by constitutional/invariant risk, blocker or author-trust impact, core-loop
severity, dependency order, evidence strength, proof readiness, and strength-regression risk.
Keep First operational action separate from the first new PRD when verification, a reopen, or
skill maintenance must happen first.
First operational action is the first substantive portfolio action that custody must resolve,
not the mandatory custody or publication handoff. Never set it to $playtest-to-issues,
playtest-to-issues, or /to-prd; use none - <reason> when no substantive action precedes the
remaining portfolio.
Describe behavior, authority, the existing seam, affected surfaces, acceptance, and testing at PRD-ready depth. Do not prescribe patch order, volatile symbols, components, or new abstractions unless current architecture makes the constraint unavoidable.
If one material product-scope, sequence, or stewardship choice remains after the evidence ladder, ask one focused question with a recommendation. Do not conduct a routine requirements interview or offload classifications to the user.
This step is complete only when the portfolio has one publication verdict, one first operational action, explicit non-PRD destinations, deferred work, rejected alternatives, and preservation constraints.
4. Write the same-stem prep artifact
Read Prep artifact contract in full immediately before writing. Write or update the default same-stem artifact; never fork numbered or timestamped copies. On rerun, carry the prior-recommendation consumption ledger into the updated artifact. For a continuation, include the prior report's prep recommendations in that same ledger when its derived prep exists.
Always write the current Prep contract version from the artifact contract. An unversioned or
older same-stem prep is historical recommendation input, never a template to preserve byte-for-byte.
The producer alone migrates it: inventory its prior recommendations, classify each one against live
evidence, and rewrite the same-stem artifact at the current version. Do not delegate migration to
$playtest-to-issues.
Consult /to-prd for house style only: publication-package vocabulary, source durability, label
posture, browser-visible checklist mapping, and recommended testing seams. Do not draft or stage a
PRD body, create an issue, apply labels, or ask /to-prd's seam-confirmation question. Record that
the seam checkpoint remains owed.
Keep the artifact privacy-safe. Do not copy full prompts, record payloads, raw assistance output,
candidate or accepted prose, API keys, localhost URLs, or machine-local paths. Summarize temporary
evidence and classify it as summarized, not cited.
This step is complete only when /to-prd could select and draft the intended package without
re-reading the full playtest report to recover provenance, scope, preserved strengths, authority
impact, sequencing, or deferred work.
5. Validate and close
Keep the ## Completion Self-Check fields at their draft pending values while validating shape.
Run:
node .claude/skills/playtest-prd-prep/scripts/validate-prd-prep.mjs --draft <report> <prep-artifact>
Fix every structural or privacy error. Compare the emitted source and output counts with the report
tables; the helper proves shape and coverage, never semantic correctness. Manually review each
disposition, candidate grouping, authority impact, preservation constraint, and evidence claim,
then complete the privacy scan. Immediately before completing the freshness scan, re-run
git status --short --untracked-files=all and git branch --show-current; replace the artifact's
final branch and worktree ledger with that exact snapshot and classify every remaining path as the
intentional prep artifact, pre-existing, or concurrent/unowned. Only after those checks actually
complete, replace the three draft self-check values with the final values from the artifact contract
and run:
node .claude/skills/playtest-prd-prep/scripts/validate-prd-prep.mjs <report> <prep-artifact>
Fix every final-validation error and rerun until it passes. Never leave final completion claims in
an artifact that has not passed the final command. Immediately after a pass, re-run
git status --short --untracked-files=all and git branch --show-current and compare both outputs
with the validated snapshot. If either differs, update the snapshot and classifications, repeat the
affected freshness and privacy review, rerun final validation, and compare again. Stop only when
the post-validation comparison matches the validated snapshot exactly. Do not run root gates merely
for a report-only prep; report focused tests, probes, browser checks, and broader gates skipped.
The run is complete only when the source inspector reports passed or nonblocking-defects with
the required skill-maintenance follow-up, every source row and strength is covered, the final prep
validator passes, semantic review is recorded, privacy and freshness scans are clear, and the final
worktree boundary is reconciled. A closeout that points directly to /to-prd without the exact
$playtest-to-issues custody handoff is incomplete.
Before sending the final response, render this exact keyed block as the final content from the
validated artifact. Keep every key, use none where applicable, and list every remaining path
individually under Remaining dirt rather than collapsing an untracked directory:
Source report: <canonical source report>
Prep artifact: <same-stem prep artifact>
First operational action: <substantive action or none - reason>
Publication verdict: <package and intended first PRD, or no-new-PRD result>
Deferred work: <every deferred PRD candidate, or none>
Non-PRD work: <every item and destination, or none>
Issue-custody handoff: $playtest-to-issues "<prep-artifact>"; required before /to-prd
Tracker and durability: <tracker limitations; source durability; artifact durability>
Intentional change: <authored prep artifact>
Remaining dirt:
- <path>: <pre-existing, concurrent/unowned, or intentional classification>
Non-actions: no implementation, source-report edit, tracker mutation, PRD publication, or /to-prd seam checkpoint occurred
04 同仓库更多 Skills
triage
Move issues and external PRs through a state machine of triage roles — categorise, verify, grill if needed, and write agent-ready briefs.
triage
Move issues and external PRs through a state machine of triage roles — categorise, verify, grill if needed, and write agent-ready briefs.
triage
Move issues and external PRs through a state machine of triage roles — categorise, verify, grill if needed, and write agent-ready briefs.
to-prd
Turn the current conversation into a PRD and publish it to the project issue tracker — no requirements interview, just synthesis of what you've already discussed; the seam confirmation is the sole checkpoint.
to-prd
Turn the current conversation into a PRD and publish it to the project issue tracker — no requirements interview, just synthesis of what you've already discussed; the seam confirmation is the sole checkpoint.
to-prd
Turn the current conversation into a PRD and publish it to the project issue tracker — no requirements interview, just synthesis of what you've already discussed; the seam confirmation is the sole checkpoint.