extract-build-log

name	extract-build-log
description	Extracts the build conversation for a recorded run in the spec-build-lab and writes it to evals/<run>/session-log.md. Use this skill ANY time the user wants to capture, save, snapshot, extract, export, or render the conversation for a specific build run — including phrasings like "extract the build log", "save the conversation for run-02", "snapshot run-02", "export the build log", "render the session log", "log this build", "produce the session log for run-NN", or "/extract-build-log run-NN". Reads the transcript recorded by Claude Code (located via the .session breadcrumb at evals/<run>/.session) and slices it between an opening bookend (the /prepare-build announcement) and a closing bookend (a fuzzy diagram/structure phrase by default, or a user-supplied --until pattern). Idempotent — re-runnable with different bookends. Requires an explicit run-NN; if none is given, list the runs in builds/ and ASK which one — never default. Do NOT use this skill to evaluate the run — that's /eval-spec-vs-build or /eval-materials-vs-build.
argument-hint	<run-NN> [--until="phrase"]
allowed-tools	["Read","Write","Bash","Grep"]

extract-build-log

Slice the build conversation out of Claude Code's session transcript and render it as a markdown log at evals/<run>/session-log.md.

Steps

1. Verify cwd is a course folder. Confirm spec.md, materials/, builds/, evals/ exist. Refuse otherwise: "Run /extract-build-log from inside a course folder — cd courses/<name> first." Stop.
2. Require a run. If $ARGUMENTS (or its first whitespace-separated token) is empty:

   ls -1 builds/ 2>/dev/null | grep -E '^run-[0-9]+$' | sort
   

   List those runs and ask: "Which run should I extract? Pick one — I won't default to the latest." Stop until they answer.
3. Validate the run. Confirm evals/$ARGUMENTS/.session exists. Refuse otherwise: "evals/$ARGUMENTS/.session is missing — this run wasn't allocated by /prepare-build, so I can't locate its transcript." Stop.
4. Run the extractor. From the project root:

   python3 "$CLAUDE_PROJECT_DIR/.claude/skills/extract-build-log/scripts/extract.py" \
     "evals/$ARGUMENTS" \
     [--until="<user-supplied phrase>"]
   

   Pass --until only if the user gave one. The script:
   • Reads evals/$ARGUMENTS/.session to find the transcript
   • Slices between the opening bookend (the /prepare-build announcement matching this run number) and the closing bookend (--until regex if given, else a fuzzy default, else end of transcript)
   • Writes evals/$ARGUMENTS/session-log.md (overwrites if present — re-run freely with different bookends)
5. Report. Echo the script's stdout — it prints the path written plus a one-line summary of how many turns were extracted and which bookends were used.
6. Hand off. Tell the user: "Ready to evaluate run-$ARGUMENTS — I can run /eval-spec-vs-build $ARGUMENTS and /eval-materials-vs-build $ARGUMENTS next. Say 'go ahead' to proceed, or pick one if you only want to run one of them."

Don't

• Don't generate or judge content — this skill is a transcript slicer, not an analyst.
• Don't read or write to builds/<run>/ — that's the build artifact, out of scope.
• Don't pick a default run. List the runs in builds/ and ask.
• Don't fabricate a .session file if it's missing — that means the run wasn't allocated normally; refuse and ask the user to investigate.