Fix

kill says what's wrong. fix says what to change, what each change takes, and how far that gets the score. It works from evidence, and when no honest set of changes reaches 70 it says so and shows the best score within reach. It never edits the verdict: the fixed score is a projection.

${CLAUDE_PLUGIN_ROOT} is the plugin folder. Safety rules §0, source rules §3, the rubric and the fixed rules §5 of ${CLAUDE_PLUGIN_ROOT}/skills/kill/SKILL.md apply.

Steps

1. Load the idea. ideas/*-<slug>/ with idea.json, verdict.json and report-data.json (competitors, customer voice and personas from a deepdive live in its narrative), and ideas/BRAIN.md when it exists (skills, budget, time). No verdict → run /kill-my-idea:kill first.
2. Find where the points go. Each sub-score loses weight × (5 − score) / 15 points (the criterion's weight, 3 sub-scores per criterion); start from the biggest losses. Also list the rules that cut the total (R2 caps at 69, R6 at 49, R7 takes 10 and caps at 49): lifting a rule is often worth more than any sub-score.
3. Research the levers (at most 10 searches and 5 page fetches, 2 subagents in parallel when available, in the foreground). Reuse the verdict's sources first.
   • Price: the current price page of at least 3 competitors (amount, period, date). Pick a position, undercut, match or premium, and justify it with margin and LTV:CAC.
   • Features: the must-haves (what customers complain is missing, with a link to the complaint or review), what to cut, and the one difference competitors don't have.
   • Customer, channel, model, country, cost: only where the evidence supports a change.
4. Choose the levers (1–6), the smallest set that gets the furthest. For each:
   • change in one line, evidence and its source;
   • moves: each sub-score it changes, from (the verdict's score, exactly) to, with why. Use the rubric anchors in kill skill §5. One lever per sub-score;
   • clears: the kill shots it removes (they must have hit in the verdict);
   • prerequisites: what it takes, each with kind (skill, money, time, partner, legal, data, audience), cost, and have (yes, no, unknown from what the user said or BRAIN.md);
   • difficulty (doable, hard, very-hard as in kill skill §7), effort, cost.
5. Re-score the fixed idea. Copy all 21 sub-scores from the verdict, apply the moves, write the notes for the ones that moved, set kill_shots (a cleared one becomes false), apply the fixed rules again (rules_applied, with R5 coherence), update metrics.ltv_cac when a lever changes it, then compute score_raw, score and verdict with the §5 formula.
   • reachable is true only when the projected score is 70 or more.
   • Not reachable → blockers: what still stands in the way (founder fit, a market with a dominant free option, a channel only a team could run…). Never stretch a sub-score to reach 70. A 62 with honest blockers is the right answer when that's what the evidence says.
6. Plan and kill criteria. 3–7 steps in order (when, step, verify with a number), the cheapest proof first. Kill criteria with a number and a date.
7. Write fix.json (format below), then run python3 "${CLAUDE_PLUGIN_ROOT}/scripts/fix-kit.py" ideas/<folder>. It refuses a sub-score that changed without a lever, a from that isn't the verdict's, a price position that doesn't match the competitor prices, and a score that doesn't recompute. Fix every error and rerun. When it passes it adds the "Path to GO" section to report.html (or "How close it can get").
8. Write fix.md next to the report: the fixed idea in one line, score before → after (projected), the levers in order with their prerequisites, the price table, the features, the plan, the kill criteria, the blockers, and the sources with access dates.
9. Answer in chat, short: before → after (projected) and whether GO is reachable; the 2 levers worth the most points, each with what you need first; the recommended price against the competitors; the first step of the plan. End with: "Build it: /kill-my-idea:build --fix."

Never publish, buy, sign up or contact anyone. Never change verdict.json or idea.json.

fix.json

{
  "schema_version": "1.0",
  "slug": "<verdict.slug>",
  "generated": "YYYY-MM-DD",
  "based_on": {"verdict": "<copied>", "score": 0},
  "summary": "<the fixed idea in one line>",
  "pricing": {
    "current": {"amount": 0, "currency": "USD", "period": "month"},
    "competitors": [{"name": "", "amount": 0, "period": "month", "source": "<pricing page>", "as_of": "YYYY-MM-DD"}],
    "recommended": {"amount": 0, "period": "month", "position": "undercut | match | premium", "why": ""}
  },
  "features": {
    "must": [{"feature": "", "why": "", "source": ""}],
    "cut": [{"feature": "", "why": ""}],
    "differentiator": {"feature": "", "why": "", "source": ""}
  },
  "levers": [{
    "id": "L1", "kind": "price | features | customer | channel | model | country | cost | other",
    "change": "", "evidence": "", "source": "", "difficulty": "doable", "effort": "", "cost": "",
    "moves": [{"sub": "<one of the 21 sub-score keys>", "from": 2, "to": 3.5, "why": ""}],
    "clears": ["<kill shot key>"],
    "prerequisites": [{"what": "", "kind": "skill", "cost": "", "have": "yes | no | unknown"}]
  }],
  "fixed": {
    "criteria_detail": {"<the 21 sub-scores, verdict.json format>": {}},
    "kill_shots": {"<the 8 kill shots>": false},
    "rules_applied": [],
    "metrics": {"ltv_cac": {"value": 0, "basis": "estimate", "reasoning": ""}},
    "score": 0, "score_raw": 0.0, "verdict": "GO | PIVOT | NO-GO"
  },
  "reachable": false,
  "blockers": [""],
  "plan": [{"when": "Week 1", "step": "", "verify": "<with a number>"}],
  "kill_criteria": ["<with a number and a date>"]
}

Plain text in every string. Price competitors compare per the same period. Every URL is one you found or fetched.