aliradid/citation-needed

**No claim without a receipt.**
An agent skill that makes AI-written text checkable. It pulls every factual claim out of
a draft, verifies each one against live sources with the agent's own web tools, and
returns an evidence ledger you can open link by link. It flags what is wrong, in place.
It never rewrites your text behind your back, and it never invents a source.
Works in Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Kimi, Qwen, Copilot, Zed and
Grok. One SKILL.md, ten thin installers.
Other harnesses: INSTALL.md.
Who this is for
- **Anyone publishing AI-written text:** blog posts, newsletters, YouTube scripts, docs, reports.
- **Teams with a content pipeline:** lint the report in CI and block a publish on
"Fix N claims" (recipe).
- **People who got burned once:** a wrong year, a misquote, a stat that was true in 2019.
Why not just ask the model "is this accurate?"
| | "Is this accurate?" | citation-needed |
| --- | --- | --- |
| Source of truth | The model's memory | Live searches from this session |
| Links | Often remembered, sometimes invented | Only URLs a tool returned |
| Coverage | Whatever it feels like mentioning | Every hard claim gets a ledger row |
| Opinions and predictions | Mixed in or "corrected" | Triaged out as SKIPPED |
| When it finds nothing | Often says "looks correct" | UNVERIFIED plus the queries tried |
| Your text | May be silently rewritten | Never touched; fixes are proposed |
| Checkable | By reading it | By opening the links, or validate_report.py |
FAQ
**Does it need an API key or a service?** No. It uses the web search and fetch tools
your agent already has. If the agent has none, it says so on the first line and marks
everything UNCHECKED rather than guessing.
**Is it a magic truth machine?** No. It is a discipline: extract, triage, verify, show
receipts. Sources can be wrong, and a wiki is treated as a pointer, not a source. The
evals report one real miss, openly (RESULTS.md).
**Will it rewrite my article?** Never. It annotates in place and proposes fixes; you apply.
**How much does it cost?** About 130 tokens per session idle, about 2,500 when it fires,
plus the web searches (capped at three per claim).
**Can I add a platform?** Yes. Adapters are thin wrappers around one SKILL.md. See
CONTRIBUTING.md.
The problem
A language model writes the most likely next words. A date that *sounds* right comes out
with the same confidence as one that *is* right. Nothing in the default loop forces the
claim to be checked before it reaches a reader. So agents ship articles with the wrong
year, scripts with a quote attributed to the wrong person, and reports with a figure
that was true in 2019.
What it does
See a full unedited run, scored against ground truth, in docs/example-report.md. (URLs above are illustrative.) In a real report, every URL in the ledger came back from a tool call in that session. If the agent could
not find a source, the row says UNVERIFIED and lists the queries it tried, so you can
pick up where it stopped.
Two modes
**Audit.** Ask it to fact-check, verify, source, or audit any text. You get the report
above. Your file is not touched.
**Guard.** When the agent itself writes an article, script, post, newsletter, report, or
anything else it expects you to publish, it checks its own draft first and delivers it
with problem markers inline and a *Receipts* section at the end. Turn guard on for every
session with a flag file: