Five sources report the same detail, so you write it up as confirmed. Then you look closer. Two of them cite the third. The third cites a post that has since been deleted. The last two ran the same wire copy. You had one source, not five, and the one source is gone.
I spent 20+ years as an intelligence analyst, and this trap has a name: circular reporting. It catches smart people constantly, because agreement feels like proof. What surprised me when I started using chat assistants for research is that they fall for it too. Ask ChatGPT or Claude to research a company and you get a confident, fluent answer. Ask where each claim actually comes from and the confidence quietly falls apart.
You do not need a smarter model to fix this. You need a prompt that makes the assistant do the checking instead of just talking about it. Here is the one I use, in full. Paste it into whatever assistant you already have, then paste your sources where marked.
Screen the material below for circular reporting and signs of deliberate
disinformation. Be concrete, cite which source you mean, and do not
overclaim: label suspicions as suspicions.
Run these checks:
1. Origin tracing: for each major claim, trace it back as far as the sources
allow. Draw the chain (S3 cites S1, S1 cites an anonymous post, etc.).
Identify any single point of origin that multiple sources depend on.
2. Circular reporting: flag cases where apparent corroboration collapses
to one origin.
3. Seeding indicators: note any of these if present, and say which source:
- a claim that appears fully formed with no traceable first report
- coordinated or near-identical wording across accounts
- emotional or identity-triggering framing that discourages checking
- a source with a clear motive to plant this specific claim
- timing tied to an event the claim would conveniently influence
- images or documents presented without verifiable provenance
4. Absence check: what would we expect to see if the claim were true and
independently reported, that is missing here.
Return a short verdict per major claim: Independently supported / Suspected
circular / Possible seeded narrative / Insufficient to judge, with reasons.
Distinguish "I found evidence of manipulation" from "I cannot yet confirm
this," and never present the second as the first. If you find no sign of
seeding or circularity, say so plainly. Do not manufacture a suspicion to
look thorough.
SOURCES AND CLAIMS:
[PASTE SOURCES]
What you get back
Three things make this different from just asking "are these sources good?"
The assistant has to draw the chain. Not summarize it: draw it. "S4 cites S2, S2 cites a deleted post, origin unknown." Written out like that, five sources collapsing into one becomes something you can see, and check yourself.
A suspicion stays labeled a suspicion. The prompt makes the assistant keep "I found evidence of manipulation" separate from "I cannot confirm this yet," and forbids dressing the second up as the first. Without that line, assistants drift toward whichever version sounds more decisive.
"Nothing looks wrong" is an allowed answer. A checker that always has to find something will start inventing problems to look thorough. This one is told, in writing, not to.
Try it on your own work first
Before you trust it on anything new, run it on the last research summary you wrote. If every claim comes back independently supported on the first pass, either your sourcing is unusually clean or the assistant is going easy on you. Ask it to draw the chain for the two claims that matter most and see if they hold.
The other fifteen
This is prompt 3 of 16. The rest cover the habits around it: how much to trust each source (graded, not gut feel), whether your sources actually agree or just repeat each other, pressure-testing the explanation you like best before you commit to it, stripping overconfident wording out of a final draft, and writing the one-page summary a busy reader actually reads. Three report templates are included. Every prompt has a note on when to use it and what a good answer looks like, so you never need to know any method names.
The pack is pay what you want with a $0 minimum, and the files are identical at any price: OSINT Tradecraft Prompt Pack. Full disclosure: I wrote it, and this article is the sales pitch. The pitch is that one of the prompts is free, in full, above.
If you work in Claude Code, the same tradecraft also exists as agents instead of prompts: a free MIT-licensed verification subagent at github.com/beeswaxpat/osint-verifier, and the full four-agent research team in the OSINT Research-Agent Pack ($29). The prompts need nothing installed; the agents run the same checks end to end.
Top comments (1)
Circular reporting is exactly where prompts need to become workflow, not just wording. The model can help organize sources, but the important guardrail is forcing provenance checks before two recycled claims start looking like independent confirmation.