Nuke

One command, the whole pipeline, one folder. Each step is the existing skill, run as written: read its SKILL.md and follow it. Nothing is copied or shortened here, and nuke adds only the order, the stop rule and the summary.

${CLAUDE_PLUGIN_ROOT} is the plugin folder.

Before starting

Say what the run does and what it costs, then ask once (in a headless run, "Just go" counts as yes):

Full run: kill, a deep dive if it scores 50 to 69, fix, a 7-day test kit, a starter kit and post drafts. About 15 to 25 minutes and $5 to $10 of plan usage on Opus. Nothing is published, bought or sent. Go?

Ask the kill skill's questions (country, budget, goal…) in the same message, so the run doesn't stop again. Note the start time.

The steps, in order

1. kill: ${CLAUDE_PLUGIN_ROOT}/skills/kill/SKILL.md. Write down kill_score, the score it gives.
2. deepdive, only when kill_score is 50 to 69 (a PIVOT, where more evidence can change the verdict): ${CLAUDE_PLUGIN_ROOT}/skills/deepdive/SKILL.md. Otherwise skipped, with why.
3. fix: ${CLAUDE_PLUGIN_ROOT}/skills/fix/SKILL.md, unless the verdict is already a GO (skipped).
4. Stop rule. When the fixed idea still projects under 50, stop here: mark validate, build and share stopped with why, and ask the user whether to go on anyway. If they say yes, set continued_past_stop: true and continued_why. In a headless run, stop.
5. validate (kit mode): ${CLAUDE_PLUGIN_ROOT}/skills/validate/SKILL.md. Test the fixed idea when there is one.
6. build: ${CLAUDE_PLUGIN_ROOT}/skills/build/SKILL.md, with --fix when fix ran.
7. share: ${CLAUDE_PLUGIN_ROOT}/skills/share/SKILL.md. Posts for the places the user named, or one fitting subreddit and X.

Each step's own script must pass before the next step starts. A step that fails twice is stopped with the error as why, and so is everything after it.

Finish

1. Write nuke.json (format below), with minutes and cost_usd when you know them (else null).
2. Run python3 "${CLAUDE_PLUGIN_ROOT}/scripts/nuke-kit.py" ideas/<folder>. It checks the run against the rules above and every step's file against its own checker, then writes NUKE.md and the "Full run" section at the top of the report. Fix every error and rerun.
3. Answer in chat, short: the verdict card from kill, then one line per step (the fixed score and whether GO is reachable, the test's hypothesis and first target, the kit's name and key metric, where the posts are), then the one thing to do today. Point to NUKE.md and report.html.

nuke.json

{
  "schema_version": "1.0", "slug": "<verdict.slug>", "generated": "YYYY-MM-DD",
  "kill_score": 0,
  "steps": [
    {"step": "kill", "status": "done", "why": ""},
    {"step": "deepdive", "status": "done | skipped", "why": ""},
    {"step": "fix", "status": "done | skipped", "why": ""},
    {"step": "validate", "status": "done | skipped | stopped", "why": ""},
    {"step": "build", "status": "done | skipped | stopped", "why": ""},
    {"step": "share", "status": "done | skipped | stopped", "why": ""}
  ],
  "continued_past_stop": null, "continued_why": "",
  "minutes": null, "cost_usd": null
}

why is required for every step that isn't done.