correcting-ai-generated-content

Audits AI-generated documents, NotebookLM exports, and Google Drive research files for factual gaps, missing production constraints, outdated information, and unverified claims. Runs web search to fil

SKILL.md

GitHub

Correcting AI-Generated Content

Three-phase workflow: Audit → Search → Rebuild.

For voice rebuild targets (lb-essay, FlashFusion B2B) and detailed search protocol, see references/.


Phase 1 — Audit

Read the full document. Produce this structured audit before touching content:

CLAIMS INVENTORY
- [Claim] → [Verified / Needs verification / Likely outdated / Unsubstantiated]

TOP GAPS (prioritized, max 5)
1. [Gap] — why it matters
2. ...

STRUCTURAL ISSUES
- [Missing section / wrong framing / incomplete decision framework]

VOICE ISSUES
- [Drift patterns, register mismatches, banned phrases]

SEARCH QUEUE
1. [Query for gap 1]
2. [Query for gap 2]
3. [Query for gap 3]

Do not start Phase 3 until audit is complete. Skip confirmation only if user explicitly says to proceed.


Phase 2 — Search

Run web searches from the Search Queue. For each gap:

  • Search with a specific, current query (include year if recency matters)
  • Fetch the most authoritative source (official docs > primary research > aggregators)
  • Tag results [SOURCED] with source name + date, or [GENERATED] with reasoning
  • Do not proceed to Phase 3 until top 3 gaps have search results

See references/search-protocol.md for priority ordering and source rules.


Phase 3 — Rebuild

Reconstruct in the target surface format with sourced claims replacing assertions.

Surface options: lb-essay · ff-b2b · plain-markdown

See references/rebuild-targets.md for voice rules, CSS tokens, and structure per surface.

Every rebuild must include CLAIMS and GAPS footers with [SOURCED]/[GENERATED] tagging on every factual claim.


Gotchas

  • 04_Knowledge NLM deposits are unsafe. Kyle's Drive 04_Knowledge folder contains confirmed fabricated product names, invented events, and unverified case studies. Never use as a source without independent verification.
  • NLM exports strip production constraints. The audit almost always surfaces missing failure modes, cost tradeoffs, and "what this doesn't tell you" content.
  • Voice rebuilds require restructuring, not editing. Announcement openers and summarizing closers in AI-generated content are structural. Edit in place and you get a hybrid that satisfies neither register. Rewrite from the opener.
  • Model-generated precision is a red flag. "60-80% improvement," "99% of developers" — always verify specific numbers before repeating.