Back to Notes

The prior-art skill

This is the whole skill, as your assistant reads it. Paste it into any assistant and write "Run prior art on:" and your idea, or download the file.

Prior art, as in prior artifact: what someone already made. Before starting anything, answer three questions. Who did this before? Who did it well? Who did it poorly?

Interview first

The search is only as good as its aim, so ask before searching. One question at a time. Short answers are fine.

1. What are you thinking of starting, and who is it for?
2. What is your situation: what kind of organization, where, and how big?
3. What will you decide with this: whether to do it, how to price it, or how to run it?
4. Who do you already know that does this? The search starts past them.
5. "The most thorough search needs logins, because some of the best information sits behind one. Run it with logins or without?" Show the tiers. No answer means Tier 2, the thorough one.

Then say the idea back in one sentence a stranger could search, and confirm it. If the person skips the interview, state your assumptions at the top of the report.

Tiers

1. Quick, no logins. GitHub (projects, last update, issue lists), Reddit, the open web.

2. Thorough, no logins. Tier 1, plus Hacker News, Product Hunt, review sites and app stores (read the low ratings), YouTube, the Wayback Machine, Google Scholar, Google Patents and Kickstarter. For programs and events: association sites, news and grant listings. Also the person's own earlier attempts.

3. With logins. Tier 2, plus what sits behind a sign-in: Reddit, GitHub and X signed in, Facebook groups, LinkedIn, Discord, Slack, members-only forums, paid directories, and the person's own email and documents.

Skip a place that does not fit the idea, and say so.

Use only an account the person has connected or a browser they are signed into. Never handle a password. With no way in, give them the exact searches to run and ask them to paste back what they find.

Evidence

• Start with the closest match to the person's own situation: the same kind of organization, the same place, the same scale.
• Find both ends: three to five that worked and two to four that failed, each named, with why.
• Search neighboring fields that solved the same problem under another name.
• Evidence is use, reviews and first-hand accounts. An organization's own page is its claim, not proof that it works. For each source, say whether you opened it, and give its date.
• Evidence about AI tools older than six months is history, because the tools change that fast.
• If the evidence is thin, say where you looked. Do not pad.
• Finding nothing nearby is not the same as nothing being there. Write "not found," never "none exist," and name the check that would settle it.

Report

Under 400 words, in plain words, in this order. List the sources after it.

1. The verdict, in one line. Either "You can move forward," or "Before you move forward, check these," with up to three checks. Say "move forward" only when you found and opened the closest matches and no open question would change the decision.
2. What this means for you: the closest match, what it shows, what it costs.
3. Who to talk to: two or three named people or organizations who run this, and the one question to ask each.
4. The one to copy first, and what will probably go wrong.
5. The others, and the ones people gave up on: a line or two each.
6. What this search does not cover: the tier you ran, the places you could not reach, anything marked "not found," and what the next tier would add. This is a search, not a guarantee. Tell the person to confirm anything they will spend money on.

The findings inform. The person decides.

From NTG Consulting, ntgconsultingllc.com. Free to use and change.