validate-idea

The first skill in the pack. It does one job: pin the idea down tightly enough that the rest of the pack can test it, and set the pass/fail bar before anyone has seen a single complaint. People who set the bar after seeing the data always find a reason to pass.

What the user typed after /validate-idea: $ARGUMENTS

The tool

python3 "${CLAUDE_SKILL_DIR}/idea.py" check validator/idea.json
python3 "${CLAUDE_SKILL_DIR}/idea.py" lock  validator/idea.json

Below, IDEA means python3 "${CLAUDE_SKILL_DIR}/idea.py". If the path is not expanded, use the base directory Claude Code printed for this skill. The shape of the file is in ${CLAUDE_SKILL_DIR}/idea.example.json.

Step 1: the idea, the person, the problem

If $ARGUMENTS is empty, ask for the idea in one line. Then ask, one question at a time, only for what is still missing:

• Who has the problem. Push for a role and a context, not a demographic. "Small businesses" is too wide. "Solo recruiters who work leads over email, LinkedIn and WhatsApp" is something you can search for.
• What goes wrong for them today, in their words. A missed reply, a lost hour, a bill they hate. Not the absence of your product.
• What they do about it now. The workaround is the best clue there is.

If the user already knows this from their own customers, use it and say which parts came from them and which are your guesses.

Step 2: the hypothesis

One sentence the data can prove wrong:

[who] struggle with [problem] and would pay [rough price] for [outcome].

If the user has no price in mind, leave a range and say /validate-wtp will test it.

Step 3: what to search for

Write search.terms the way a frustrated person would type a post, not the way a landing page talks. Three to eight terms. Quote exact phrases: "unified inbox", "too many messaging apps". A term without quotes matches posts that contain every word, in any order.

Good terms name the pain or the workaround: "missed a message", "spreadsheet to track", "switched from <competitor>". Bad terms are your product category in marketing language, because nobody complains in that voice.

search.subreddits (optional): communities where this person hangs out. search.competitor_apps (optional): apps people already use for this. Look up each App Store id with the miner's search:

python3 "${CLAUDE_SKILL_DIR}/../validate-mine/sources.py" find-app "<app name>"

Pick the right one from the list with the user. A wrong id mines the wrong app.

Step 4: the bar

Explain this once: the bar is what the evidence must show for the idea to pass, and it is set now because nobody judges fairly after seeing the data.

bar.evidence is what the mined data has to show:

• min_complaints: distinct complaints about this problem. Default 100.
• min_money_signals: complaints that mention paying, a price, or switching tools. Default 10. This is the number that matters most.
• max_age_months: only count complaints this recent. Default 24.

bar.real_world is what a real test has to show later (/validate-smoke plans it). Ask the user, and suggest numbers if they have none, for example "200 waitlist signups from strangers" and "10 paid, refundable preorders".

Step 5: write, check, lock

1. Write validator/idea.json in the shape of the example.
2. Run IDEA check validator/idea.json and fix everything it lists.
3. Show the user the idea, the hypothesis, the terms and the bar in a few lines. Let them correct it.
4. Once they agree, run IDEA lock validator/idea.json.

After the lock, terms, subreddits and apps can still change freely. The bar cannot. If the user later wants to move it, run IDEA lock ... --reason "<why>". The move is kept in validator/idea.lock, and /validate-verdict reports it. Never edit idea.lock by hand, and never move the bar without the user asking.

Output

validator/idea.json, validator/idea.lock.

Next: /validate-mine reads the search terms and competitor apps.