Fact checking explainers

Fact-Checking Explainers

Overview

An explainer's authority comes from being right. A reader who catches one confident, wrong claim stops trusting the rest of the piece, figures and all. This skill makes claim verification enforceable by tying every checkable claim to evidence you can point to, correcting it, or cutting it. It is not a proofreading pass and not advisory: it is the gate an explainer passes before it is delivered.

The Iron Law

NO EXPLAINER SHIPS WITH AN UNVERIFIED FACTUAL CLAIM.

Every checkable claim either traces to a source that supports it, gets corrected to match the source, or gets cut.

Apply the same resolution loop to every factual claim:

• Support it with an actual source passage you can point to.
• Correct it until it matches the source.
• Cut it when no source supports it.

Treat memory, plausibility, background knowledge, and deadline pressure as signals to run the loop, not reasons to skip it. "Verify" means checking an actual source in this session.

When to Use

Run this skill:

• At research-time, over the sources gathered for an explainer, before any drafting.
• Post-draft, over the finished article, before delivery. This is a blocking gate.
• On request, whenever asked to fact-check, verify, or audit an explainer or article.

creating-explainers and explaining-codebases both call this skill at the gates above. When they do, interactive explainer delivery waits for a PASS report.

When NOT to use: content that is not an explainer, or a pure opinion or editorial piece that makes no factual claims. (Most explainers make many factual claims. Default to running it.)

What Counts as a Checkable Claim

A checkable claim is a specific factual or empirical assertion:

• numbers, dates, measurements, percentages
• names, attributions, who-did-what, who-built-what
• direct quotes
• historical events and their order
• performance results and benchmarks
• mechanism claims: "X uses Y", "X causes Y", "X is faster than Y because Z"

Not checkable (do not flag these as factual claims):

• clearly-labeled interpretation or opinion ("the most elegant part is...")
• analogies and illustrative scenarios the reader knows are illustrative
• the author's framing and transitions

Heuristic: if you are unsure whether something is a checkable claim, treat it as one and verify it. The cost of over-checking is small; the cost of a wrong claim shipping is the whole article's credibility.

The Verification Process

1. Extract every checkable claim, with its location in the article (section, figure caption, or file:line for code).
2. Find the support. For each claim, locate the passage in the cited source that backs it, and read that passage. For codebase explainers, open the actual code at the exact path and line. Do not treat the existence of a citation as support; a cited source often does not say what the draft claims.
3. Assign a verdict (table below).
4. Resolve everything that is not supported.

Verdict	Meaning	Allowed at delivery?
supported	A source passage directly backs the claim	Yes
needs-source	Plausible but no citation yet	No - add a source or downgrade
unsupported	No source backs it	No - source it, soften, or cut
contradicted	A source says otherwise	No - correct it or cut it

The Two Gates

Research-time gate (inside creating-explainers research intake). Before drafting, go through the gathered sources and confirm each one exists and actually supports the points it will be used for. Fabricated, misremembered, or misread sources are cheapest to catch here, before they are baked into prose.

Post-draft gate (every explainer, before delivery). Audit every checkable claim in the finished article against its cited source. For codebase explainers, the source of truth is the real code. Every "this function does X", every quoted snippet, every architecture claim is checked against the actual implementation at a specific path and line. Code drifts; a snippet that was right yesterday may be wrong today.

Resolution

At delivery, every checkable claim must be supported. Resolve each other verdict with the same loop:

• Add support from a source passage (turns needs-source into supported).
• Correct the claim to match what the source actually says.
• Soften to clearly-labeled interpretation, if the statement is genuinely interpretive rather than factual.
• Cut it.

An unresolved needs-source, unsupported, or contradicted claim keeps the report at FAIL. Caveats do not make unsupported factual claims deliverable.

The Report

Produce a claim-by-claim report plus an overall PASS/FAIL verdict. The exact format is in references/verification-report-format.md. The verdict is FAIL until every checkable claim is supported.

When invoked as a gate by another skill, return the report inline and hold delivery until PASS. When invoked directly by a user, present the report and offer to apply the fixes.

Rationalization Table

Excuse	Reality
"I wrote it, so I know it's right"	You know what you intended. Verify what you actually wrote.
"It's a well-known fact"	Well-known facts are wrong often enough to check. It takes 30 seconds.
"The source probably says this"	Probably is not a verdict. Open the source and confirm.
"Close enough"	Numbers, dates, and names are exact or they are wrong.
"It's only one claim"	One confident wrong claim discredits the whole explainer.
"I'll flag it and let the reader decide"	The gate is your job, not the reader's.

Red Flags

These thoughts mean stop and verify before the claim goes in:

• "I'm fairly sure that..."
• "If I recall correctly..."
• "roughly..." / "about..." / "something like..." attached to a specific number
• "it's basically the same as..."
• pasting a statistic without re-opening the source it came from
• citing a URL you have not actually opened in this session

Common Mistakes

Mistake	Fix
Verifying that the citation exists, not that it supports the claim	Read the cited passage. A real source can still fail to back the claim.
Checking the easy claims, waving through the hard ones	The hard-to-verify claims are exactly where errors hide.
For code: trusting a comment, a function name, or a variable name	Read the implementation. Names and comments can lie; the code does not.
Treating needs-source as shippable	needs-source is a to-do, not a pass. Resolve it.
Stopping at the first source that agrees	If sources disagree, that disagreement belongs in the article, not buried.