PROMPTS.md
Audit agent prompts
Canonical evaluation criteria per target. Hand these to the Explore agent along with the input sample. Versioned here so future sessions use the same rubric.
tier1
For each entry in audits/inputs-YYYY-MM-DD/tier1-sample.json (fields: tier1_num, title, post_slug, post_date, description, rationale):
1. Read /home/a/Projects/craig/blog/raw/ (derive from post_slug).
2. Rate description_match: match | minor-drift | mismatch.
3. If drift/mismatch, name specifically what's wrong.
4. Rate rationale_honest: true/false (is the "Why Nth" an honest explanation of position in the reading order, or post-hoc?).
5. Record a short evidence quote from the post.
Write to audits/YYYY-MM-DD-tier1.json + .md.
wisdom-rank9plus
For each insight in audits/inputs-YYYY-MM-DD/wisdom-sample.json (fields: id, text, source.{title,date,slug,category}, pillar, impact_rank, type, consequences, tier, theme):
1. Read /home/a/Projects/craig/blog/raw/.
2. Rate text_match: verbatim | paraphrase | missing. Record a context snippet from the post.
3. Rate pillar_correct: correct | arguable | wrong. If wrong, suggest pillar.
4. Rate rank_deserved: deserves-9+ | borderline | inflated. Rank 10 = foundational thesis; rank 9 = major distinctive claim.
5. Rate shape: insight (claim/argument) | boilerplate (emphatic description).
Write to audits/YYYY-MM-DD-wisdom-rank9plus.json + .md.
blog-tags
For each post in audits/inputs-YYYY-MM-DD/blog-tags-sample.json (fields: slug, category, title, word_count, topics, ...):
1. Read /home/a/Projects/craig/blog/raw/.
2. For EACH assigned topic: is the post substantively about it (core / relevant / surface-match)?
3. Identify obvious missing tags.
4. Rate overall verdict: honest | over-tagged | under-tagged | mis-tagged | data-issue (empty stub).
5. Thesis in one line.
Watch for the L29 trap. Posts that got a tag from one keyword mention (e.g. bitcoin-protocol from a single "blockchain" reference) are drift. Pillars known to drift: bitcoin-protocol (via "blockchain"), identity-history (via "history" in title), information-theory (via "information" — S7 fix should have held).
Write to audits/YYYY-MM-DD-blog-tags.json + .md.
topic-crossrefs
For each post in audits/inputs-YYYY-MM-DD/topic-crossref-sample.json (nested: {topic_name: [{slug,category,title,raw_path,words,date}]}):
1. Read /home/a/Projects/craig/.
2. For EACH of the two audited topics: rate core-topic | major-theme | tangential | not-about-it.
3. If a post is listed under BOTH topics, it must score ≥ major-theme for each. Otherwise flag as drift.
4. Also note posts listed under only ONE topic that should be cross-referenced under the other (under-tagging).
Write to audits/YYYY-MM-DD-topic-crossrefs.json + .md.