agentleFS
agiea-brainagiea-brain/agency-skills/skills/david-hoos/icp-review/SKILL.md

icp-review

Skill1 starsChanged 17 days agoMITView on GitHub

Use this skill before shipping customer-facing copy you will spend money or reputation on: a blog post, ad, cold email, proposal, pricing or landing page. Produces an evidence-based buyer panel read built only from verbatim quotes in your own sales and client call transcripts: where a buyer stops reading, the strongest objection the copy raises and never answers, and the questions the reader brings with them, every claim tied to a real quote and checked by code. Also use for "ICP review this", "would this resonate", "what would a buyer think", "which of these lands", "why isn't this converting", "review this before I ship". First run walks through setup.

The file

Show its rules and text

We split this skill into the rules it gives an agent

15 rules. About 0 are common, 5 are distinctive, and 10 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 10

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Run every command from this skill's folder (cd "<skill base directory>" first).

    line 24
  2. Very specific:

    Read the widening. A non-empty dropped means the answer is not persona-specific.

    line 30
  3. Very specific:

    Keep it fresh. When the user says "refresh the panel", or the newest quote is more than about 90 days old, run step 3 of …

    line 35
  4. Very specific:

    MUST cite only quote_keys from the retrieved evidence pack.

    line 47
  5. Very specific:

    MUST pass references/validate_verdict.py (exit 0) before showing the user any verdict.

    line 48
  6. Very specific:

    MUST include strongest_objection on every verdict, including favourable ones.

    line 49
  7. Distinctive:

    MUST pre-register Choose predictions before the real result exists.

    line 50
  8. Very specific:

    NEVER load anything in excluded_docs: the user's own positioning, message house, site copy or scripts.

    line 51
  9. Distinctive:

    NEVER add the draft under review to the evidence base.

    line 52
  10. Distinctive:

    NEVER let the panel write copy.

    line 53
  11. Distinctive:

    NEVER give a seat a name, backstory, mood or biography.

    line 54
  12. Very specific:

    NEVER use the client seat on acquisition copy.

    line 55
  13. Very specific:

    NEVER simulate a click path.

    line 56
  14. Distinctive:

    NEVER pad the corpus with invented or paraphrased quotes.

    line 57
  15. Very specific:

    NEVER publish a buyer quote or attribute one to a named person or company unless public_quote_rule in references/profile.json allows it.

    line 58

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Applies to any draft, topic, set of variants or live page that a real buyer will read. Produces one validated verdict per seat, every claim citing a buyer quote that exists, with the strongest objection first and the verdict last.

A made-up persona invents opinions. This panel can only use what buyers actually said on calls, and when the evidence is thin it says "I can't judge this" instead of guessing.

The play

Run every command from this skill's folder (cd "<skill base directory>" first). The tools are Python 3, standard library only.

  1. Check setup. Read references/profile.json. If setup_complete is false, or data/corpus.json does not exist, stop and run references/setup.md with the user before anything else: eight short questions, then extract and tag verbatim buyer quotes from their call transcripts. Never judge copy from general knowledge as a stand-in. A panel with no corpus is a persona, and personas make things up.
  2. Route. Work out the play from what the user hands over (Brief, Gate, Choose, Diagnose, Retention), pick the seats, and say in one line which ran and why. The routing table, the four seats and what each may never judge are in references/routing-and-seats.md. objector runs in every play except Brief.
  3. Retrieve. mkdir -p runs/<YYYY-MM-DD>-<slug>, then python3 references/panel.py evidence --seat <seat> [--persona …] [--stage …] [--pattern …] > runs/<…>/pack-<seat>.json.
  4. Respect the band. insufficient (n under 5): the verdict MUST be null; say the corpus does not cover it. thin (5 to 11): confidence MUST be "weak signal". Never widen past the seat's base filter to manufacture an answer.
  5. Read the widening. A non-empty dropped means the answer is not persona-specific. Say which level it was answered at before anything else.
  6. Draft the verdict to references/verdict-schema.json and save it as runs/<…>/verdict-<seat>.json. Every assertion carries quote_keys drawn only from the returned pack.
  7. Validate. python3 references/validate_verdict.py runs/<…>/verdict-<seat>.json --evidence runs/<…>/pack-<seat>.json. If it rejects, fix the verdict, never the validator.
  8. Present the answer first: two or three sentences on what to fix, then the detail in the inverted reporting order in references/routing-and-seats.md.
  9. On a Choose play with a real result coming, pre-register the pick in predictions.jsonl before the result exists, fill in the winner when it lands, and run python3 references/score.py.
  10. Keep it fresh. When the user says "refresh the panel", or the newest quote is more than about 90 days old, run step 3 of references/setup.md on the new transcripts only, then python3 references/build_corpus.py.

What good looks like

  • First thing I check that most people skip: strongest_objection, not the verdict. The verdict tends to agree with whoever wrote the draft, because the draft came from the same instincts. The objection is the field with no polite answer.
  • The mediocre version: asking an AI to "pretend to be my ICP". It reads well, cites nothing, and invents an opinion for every gap.
  • Second mediocre version: running it on every draft. The dissent becomes noise and gets ignored on the one that mattered.
  • I know a run is honest when a thin seat declines to judge, distinct_calls is reported next to the quote count, and a stale or widened slice is named before any answer.
  • I know the panel is right only from score.py. The gate proves the evidence exists; it cannot prove the reasoning. Until roughly 30 scored predictions, these are well-sourced opinions, not calibrated predictions, and say so.

Rules

  • MUST cite only quote_keys from the retrieved evidence pack.
  • MUST pass references/validate_verdict.py (exit 0) before showing the user any verdict.
  • MUST include strongest_objection on every verdict, including favourable ones.
  • MUST pre-register Choose predictions before the real result exists.
  • NEVER load anything in excluded_docs: the user's own positioning, message house, site copy or scripts. That grades their copy against their copy.
  • NEVER add the draft under review to the evidence base.
  • NEVER let the panel write copy. It judges; a panel that grades its own drafts converges on its own priors.
  • NEVER give a seat a name, backstory, mood or biography.
  • NEVER use the client seat on acquisition copy.
  • NEVER simulate a click path. Navigation is answered as unresolved_questions.
  • NEVER pad the corpus with invented or paraphrased quotes.
  • NEVER publish a buyer quote or attribute one to a named person or company unless public_quote_rule in references/profile.json allows it.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.