Fact Lock

The problem

The most dangerous failure mode in AI-assisted factual writing isn't obvious errors — it's confident, fluent, entirely fabricated specifics: a statistic that sounds exactly right, a quote attributed to someone who never said it, a citation to a study that doesn't exist. Fluency reads as authority, but fluency is not evidence.

This is the mirror image of Fiction World Bible. In fiction, canon is something the author builds and Claude enforces alongside them. Here, the "canon" is the real world — the job is representing it honestly rather than filling gaps with invention.

What it does

Before producing content for a new project, the skill runs a setup interview: what's being written, what source material already exists, how sourcing should work, what citation style to use, and what happens when something needed isn't actually covered by any available source.

The default sourcing mode is hybrid: your own provided material — uploaded documents, a defined reading list, whatever corpus you bring — is treated as the primary authority. Web search fills gaps that material doesn't cover, and anything found that way gets flagged explicitly as externally verified, never blended in as if it came from your own sources.

From there, every piece of content follows the same rules:

• No specific fact without a real, checkable source. Statistics, dates, quotes, study findings, names — all of it needs a citation, every time.
• No fabricated citations, ever. If a source can't be found, the claim doesn't go in. A missing fact is always better than an invented one.
• Three tiers, kept visibly separate: what your own material directly states, what web search found to fill a gap (flagged as such), and Claude's own interpretation connecting sources (labeled as analysis, never presented with the confidence of a direct citation).
• A live source registry tracks every source actually used and what specific claims depend on it — an audit trail you or anyone else could use to verify the finished content claim by claim.
• Gaps get flagged, not filled. If something the content needs isn't in your material and isn't reliably verifiable, you hear about it directly rather than getting something plausible-sounding instead.

Use cases

• A non-fiction author working from a defined set of sources — every chapter checked against a running registry showing exactly which claims come from where.
• An educator building curriculum content — every factual claim treated the same regardless of how confident it sounds.
• A software team shipping domain-accurate content — health, legal, financial, or scientific claims embedded in a product get tightened discipline, not relaxed, with a recommended human-expert review pass before anything ships.
• A researcher synthesizing across sources — connections stay clearly separate from directly cited fact, so it's always obvious what's established versus your own interpretation still needing review.

What it won't do

It won't invent sources to make thin research look complete, and it won't relax its citation standard because a deadline is close or a section feels short. A shorter, fully-sourced piece of writing is always the right tradeoff against a longer one padded with unverified claims. It also won't decide what's true on its own — your provided material and verifiable outside sources do that.

Try it

Tell Claude you're writing non-fiction, a research report, educational material, or fact-based content for a product, and it'll run the setup interview automatically. Download the skill below and add it to Claude or Claude Code.