Compare Ideas

Options

• --refresh (also when the user says "refresh" or "update the ranking"): re-research stale verdicts before ranking (step 2). Without it, compare only reads the verdicts. This option was requested by a user, so a ranking built on old research can be brought up to date in one command.
• --refresh all (or "refresh all of them"): refresh every verdict, stale or not.

Steps

1. Find the verdicts. Read every ideas/*/verdict.json (and idea.json) in the current project.
   • Fewer than 2 found → say so, and suggest /kill-my-idea:kill.
   • A verdict whose evaluated date is more than 90 days before today is stale. Without --refresh, mark it and suggest /kill-my-idea:compare --refresh.
   • A verdict whose confidence.level is low is marked low confidence: its rank is provisional.
2. Refresh (only with --refresh). For each stale verdict (or every verdict with --refresh all), one idea at a time or one subagent per idea:
   1. Copy the current mode, evaluated, score, verdict and source count into a new entry at the end of verdict.history.
   2. Rerun the kill skill's research steps 3–5 for that idea (${CLAUDE_PLUGIN_ROOT}/skills/kill/SKILL.md), with its safety rules (§0) and source minimums. Keep idea.json as it is; the user's answers still hold.
   3. Update verdict.json: evaluated = today, the metrics, sub-scores, score, rules, confidence, kill shots, rescue paths and sources. Keep mode.
   4. Rewrite report-data.json (kill skill §6) with the new verdict and narrative, including premortem, why_now and icp, then rebuild report.html with python3 "${CLAUDE_PLUGIN_ROOT}/scripts/build-report.py" ideas/<folder>.
   5. Run python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check-idea.py" ideas/<folder> when Bash is available, and fix every error. Tell the user which ideas were refreshed and how each score moved ("42 → 47").

3. Build the ranking table. Sort by score, then by time_to_first_revenue_weeks (the faster one wins), then by confidence (high before medium before low).

   #

   Idea

   Verdict

   Score

   Confidence

   Best rescue (score, difficulty)

   LTV : CAC

   Break-even

   Time to first revenue

   Primary kill shot

   Read the values from verdict.json: rescue[0], metrics.ltv_cac, metrics.break_even_customers_per_month, metrics.time_to_first_revenue_weeks, primary_kill_shot. A missing metric is unknown, never a guess. Mark a score that rules_applied capped ("69, capped: no channel"). Show the last history entry when a verdict changed ("42 → 47"). Add a column per country when ideas target different markets.

4. Decide like a co-founder:
   • Focus on ONE idea this week, with the reason, its first action and a deadline. Prefer a lower score with high confidence over a higher score with low confidence, and say so.
   • Pause: what to park, and why. A NO-GO with a doable rescue path above 60 is a pause, not a kill.
   • Kill: NO-GOs that are still in ideas/ and should be archived.
   • Opportunity cost: what the focus idea beats the others on, in one line.
   • Read ideas/BRAIN.md when it exists: the user's Me profile (budget, time, goal) and the kill-shot patterns weigh in the choice.
   • Save the focus as ideas/focus.json, a strict JSON object: {"slug": "<verdict slug>", "why": "<one line>", "first_action": "<one action>", "deadline": "YYYY-MM-DD", "decided": "<today>"}. Each value is at most 300 characters.
5. Write ideas/RANKING.md with the table, the decision and today's date.
6. Write the dashboard and the brain. With Bash, run python3 "${CLAUDE_PLUGIN_ROOT}/scripts/build-brain.py" ideas, then python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check-idea.py" --board ideas. The script computes everything from the verdicts and focus.json: ideas/board-data.js, a fresh copy of ideas/board.html, and the generated block of ideas/BRAIN.md (focus, all ideas, what keeps killing your ideas, score changes, lessons). It keeps the Me section and the user's notes. Without Bash, do it by hand:
   • Write ideas/board-data.js in the format below, ideas in ranking order.
   • Copy the board page as-is. Never rewrite it. With Bash: cp "${CLAUDE_PLUGIN_ROOT}/skills/compare/board.html" ideas/board.html. Otherwise Read it and Write it unchanged.
   • Check it with python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check-idea.py" --board ideas when Bash is available, and fix every error.
7. Answer in chat with the table and the decision. Keep it short. Name the kill shot that hit the most ideas, if it hit 2 or more. End with "Open ideas/board.html in your browser. Your notes: ideas/BRAIN.md."

Without --refresh, do not re-research the ideas. Only compare what the verdicts contain.

Board data

build-brain.py writes this file; the format is here for the no-Bash case. ideas/board-data.js is one statement, a strict JSON literal (double quotes, no functions, no comments, no template literals), following ${CLAUDE_PLUGIN_ROOT}/skills/compare/board-data.schema.json:

window.KMI_BOARD = { "generated": "YYYY-MM-DD", "ideas": [ { "slug", "title", "verdict", "score", "confidence", "evaluated", "time_to_first_revenue_weeks" (number|null), "ltv_cac" (number|null), "primary_kill_shot" (string|null), "best_rescue": {"change","score_after","difficulty"}|null, "countries": [..], "report": "<folder>/report.html", "stale": bool } ] };

Never put a value in the board that is not in the verdict or the idea brief.