PRD drill: resume from disk, review-page renderer, one-write authors, length caps - #3
Draft
d6veteran wants to merge 5 commits into
Draft
PRD drill: resume from disk, review-page renderer, one-write authors, length caps#3d6veteran wants to merge 5 commits into
d6veteran wants to merge 5 commits into
Conversation
run.json can be stale: a workflow cannot write it, and a session can end before the notification arrives (drill 5a was killed by a ten-minute wait ceiling during the rewrite). The resume check now reads the run directory and the PRD's Appendix A to choose startAt, and continues without asking when the session cannot ask. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014aqcjmXas9EFDx3pstfpyE
…s the PRD itself Drill 5b: resumed at the rewrite, draft.js declared the panel had not run instead of reading the earlier run's memo and lens files, and the skill then improvised in the main session. Fixed both; the PRD and architecture contracts cap the document at 3,000 words; drill log updated. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014aqcjmXas9EFDx3pstfpyE
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014aqcjmXas9EFDx3pstfpyE
…ffort Drill 5c finished /ck:prd end to end at $13.93. The journal showed River's rewrite taking 117 turns and 182k output tokens by editing the document piece by piece, the memo synthesis 234 turns, and the main session writing the review page's HTML by hand. scripts/render-review.py (standard library only) now builds the page in one command; every author prompt says to write the file once and not re-read it; the draft, rewrite, and memo stages run at medium effort. Test 12 covers the renderer. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014aqcjmXas9EFDx3pstfpyE
…s the appendices Drill 5d finalized the PRD from the review state and set status final. With no review comments it invented the author's premortem answer; the finalize prompt now leaves the question open in that case. The finished PRD ran 407 words over the cap, all in Appendix A's verbatim panel record, so the cap applies before the appendices. Drill log: passes 5c and 5d. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014aqcjmXas9EFDx3pstfpyE
d6veteran
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What the PRD drill showed
Four passes of
/ck:prdon the game fixture (tests/drill/2026-09-09.md, drill 5). The pipeline works end to end: the skill mints the run, launchesck:draftby name, the draft is validated, the nested panel argues it on three models, River rewrites it with a challenged-claims appendix and a premortem, the skill publishes a review page and waits, and a second run finalizes from the review state. One complete PRD from scratch: 2,995 words at review, 3,407 finished, statusfinal.It cost $16.33 for that one PRD against $2.65 estimated, and the run journal says why: River's rewrite took 117 turns and 182k output tokens editing the document piece by piece, the memo synthesis 234 turns, and the main session wrote the review page's HTML by hand.
The fixes, in the order the drill found them
skills/prd/SKILL.md: the resume check reads the stage from what is on disk instead of trustingrun.json, which a workflow can never update; it relaunches withstartAt: panelwhen a result says the panel did not run; it never edits the PRD in the main session; finalize leaves the premortem question open when the author left no answer instead of inventing one.workflows/draft.js: a run resumed at the rewrite reads the earlier panel's memo and lens files; the draft and rewrite stages run at medium effort and write the file once.scripts/render-review.py(standard library only): builds the review page in one command;skills/review-page/SKILL.mduses it. Test 12 covers it.workflows/panel.js,brief.js,team.js: every author writes the file once and does not re-read it; the memo synthesis runs at medium effort.skills/prd-artifact/SKILL.md,skills/architecture-artifact/SKILL.md: 3,000 words before the appendices.Verified
tests/run.sh: 64 checks pass. The PRD incode-katz/.githubcarries the same scripts, skills, and contracts, and the measured cost. The next/ck:prdrun re-measures after these fixes.🤖 Generated with Claude Code
https://claude.ai/code/session_014aqcjmXas9EFDx3pstfpyE