Ciunkos/skills

Business idea validator

Usage: /business-idea-validator <idea text or path to a file> [--lang en|pl] [--quick] [--out <dir>]

Score one business idea on ten rubrics, apply a fixed verdict rule, write a report. The judgment is in the scores and their reasoning; the verdict is arithmetic over the scores so that two ideas get the same treatment and a reader can check it by hand.

Run from a working directory of the user's choice (--out <dir>, default the current directory). Never write into this skill's folder.

Files you will use

• references/rubrics/01-problem.md … 10-risks.md: one file per rubric. Read the one you are scoring.
• references/report-style.md: writing rules. Read once before any scoring.
• references/intake.md: when to ask the founder questions first.
• references/verdict.md: the rule, for your understanding; the code is scripts/verdict.mjs.
• assets/rubric-output.schema.json: the exact JSON one rubric pass returns.
• assets/report.schema.json: the whole report.
• assets/examples/fieldnote/: a complete example (invented idea) of idea.md, report.json, report.md.
• scripts/check.mjs: validates, computes the verdict, renders Markdown. ${CLAUDE_SKILL_DIR} points at this folder in Claude Code; elsewhere use the path the skill was installed to.

Steps

1. Intake

Take the idea from the argument, the prompt, or a file path. Decide the report language: the language of the idea text, unless --lang says otherwise.

Apply references/intake.md. If the text is thin (under about 40 words, or missing who it is for as a role or person, what it does, or how it makes money), ask the founder once, up to five questions in one message, and wait. If you cannot ask (non-interactive, or --quick), continue with assumptions and say so in synthesis.notes.

Store what you have as idea.text and idea.answers (for, money, today, advantage, stage; only the ones given).

2. Score the ten rubrics

Read references/report-style.md first. Then, for each rubric in the order problem, customer, market, alternatives, monetization, distribution, feasibility, moat, founder, risks:

• Give the pass the idea text, the answers, the rubric file, the style rules and the output schema.
• The pass returns one JSON object matching assets/rubric-output.schema.json: id, score 1 to 5, confidence low/medium/high, evidence (exact quotes), assumptions, reasoning (at most 160 words), questions, improve.

Run the passes in parallel through the business-idea-rubric subagent when the Agent tool is available (one agent per rubric, brief it with the file paths and the idea; ask it to return only the JSON). Otherwise score them yourself, one rubric at a time, reading the rubric file fresh each time so the anchors stay in front of you.

Create <out>/rubrics/ and save the ten results there as <id>.json as you go. Validate each one: node ${CLAUDE_SKILL_DIR}/scripts/check.mjs --rubric <out>/rubrics/<id>.json. Fix any problem it reports (usually a list that is too long or reasoning over 120 words) before moving on.

3. Critic pass

Skip with --quick. Otherwise give the business-idea-critic subagent (or do it yourself) all ten results, the idea, and the rubric files. Its job: try to refute every score of 1, 2 or 5 using the anchors and the text, and flag any rubric where the confidence does not match how settled the score is. It returns { "changes": [{ "id", "from", "to", "confidence", "why" }], "notes": "" } where confidence is same or the new level; why and notes are in the report language. Apply each change to the rubric result: set the score and confidence, and rewrite that rubric's reasoning so it argues for the new score (a result must not contradict itself). Keep the critic object for the report.

4. Synthesis

With the final scores, write the synthesis object (assets/report.schema.json, synthesis): summary (three sentences), riskiestAssumption (one), experiments (one to three, 14 days max, each with a numeric metric), pivots (only real ones; empty is fine), openQuestions (the rubric questions, deduplicated, most important first, at most eight), notes (how the report was made when it matters, else empty). Follow the synthesis rules in references/report-style.md.

5. Assemble and render

Write <out>/report.json with version: "1", lang, idea, rubrics (all ten), critic (if run), synthesis, verdict: {} (the script fills it; never write a verdict by hand), meta (model: the model you are running on as you know it, createdAt: ISO date, generator: "skill").

Run node ${CLAUDE_SKILL_DIR}/scripts/check.mjs <out>/report.json. It validates everything, writes the verdict into the file and writes <out>/report.md. Fix problems until it prints OK.

6. Hand over

Tell the user: the verdict and the rule that produced it, the riskiest assumption, the first experiment, and where the two files are. Keep it under ten lines; the report has the rest. Do not restate the scorecard in chat.

Rules that override everything above

• Quote, do not paraphrase, in evidence. Anything not in the text is an assumption and is labeled as such.
• No invented numbers. Market sizes and prices are "assuming …" with a range.
• The verdict comes from scripts/check.mjs. Never write it by hand and never argue with it in the report; if it looks wrong, a score is wrong, fix the score with a reason.
• Report language follows the idea text (or --lang). JSON keys stay English.
• Style: short sentences, no em dashes, no hype, no "not just X but Y".