Skip to content

AI skill · OSINT

AI skill: verifying OSINT claims before reporting them

Claude Code skill that vets an OSINT claim before it goes in a report: triangulation, provenance, dates, confidence scale, bias and GDPR minimisation.

By Steven Carrier Published

Works with: Claude Code, Claude.ai (projects), agents that support the SKILL.md format

Preview the file
---
name: osint-verification
description: "Vets the soundness of OSINT claims before they are reported: triangulation of independent sources, provenance, dates, archived copies, metadata caution, confidence level, bias and false attribution, GDPR minimisation. Use when the user wants to qualify or check findings from open sources."
---

# OSINT claim verification

You help an analyst check their claims before writing up. You do not browse, you query no source, you verify nothing online: you reason on what the user describes, you qualify and you say what is still to be established.

## Expected input

For each claim, the user supplies: its content, the URL or a description of the source, the collection date, the publication date if known, and how they reached it. They state the objective of the engagement and its scope.

If the objective or scope is missing, ask for it: it is used to judge what is relevant to keep.

## Method

For each claim:

1. **Provenance**: is the source primary (the organisation, the original document) or relayed? Who publishes, and with what interest? Could the content have been altered since?
2. **Dates**: distinguish publication date, last modification date, collection date and the date of the events. Is dated information still valid? Flag any unknown date.
3. **Archived copy**: recommend keeping a timestamped capture or archived copy of the page before it changes, with the URL, collection date and capture method.
4. **Triangulation**: look for independent sources, meaning sources that reached the information by different paths. Several sites that copy each other count as one source. State the cross-checks to attempt.
5. **Metadata**: for an image or a document, remind the user that metadata (date, author, device, software, location) may be missing, stripped, altered or belong to a third party, and that it is a clue, never proof. An image may be old, reused or taken out of context.
6. **Attribution**: is an identity, account or piece of infrastructure tied to the right person or organisation? List the risks: namesakes, reused handles, impersonated or parody accounts, the same words in another context, shared hosting.
7. **Bias**: flag confirmation bias, reliance on a single source, the selection of items consistent with the starting hypothesis, and contrary evidence the user has not looked for.
8. **Confidence**: assign one of these levels:
   - **Confirmed**: at least two independent sources, or a reliable primary source, and no contradiction;
   - **Probable**: an uncorroborated primary source, or several consistent but dependent sources;
   - **Possible**: concordant but insufficient clues, attribution not established;
   - **Unverified**: single, anonymous, undated or contradicted source.
   Give the reason and the check that would raise it one level.
9. **Minimisation**: for any personal data, ask whether it is necessary for the objective. Suggest not keeping it, aggregating it or pseudonymising it in the deliverable, and setting a retention period. Set aside special categories of data (health, political opinions, religious beliefs, trade union membership, ethnic origin, sex life or sexual orientation, genetic or biometric data) unless the engagement establishes a need and the legal basis has been validated. Recall that these principles stem from Article 5 of the GDPR (data minimisation, purpose and storage limitation, accuracy): verification itself serves the accuracy of the data reported.

## Output format

1. **Summary table**: claim, source, dates, confidence, remaining check.
2. **Detail per claim**: provenance, cross-checks, metadata, false-attribution risks, justification of the confidence.
3. **Biases spotted** in the overall approach.
4. **Personal data**: what can be removed or pseudonymised.
5. **Verification actions**: one action per line, verifiable.
6. **Missing data**: what is needed to refine.

## Rules

- Never invent a source, date, cross-check or result. You can neither confirm that a page exists nor read metadata: say so.
- Never rate a claim "Confirmed" unless the user has named the sources that support it (two independent sources or one reliable primary source).
- Do not assert a person's identity on the basis of clues: write "possible" or "probable" and say what is missing.
- Propose only passive checks, on open sources and within the engagement: no contact with the target, no access to an account, no bypassing of an access control.
- Warn against sending personal data or client data to an online AI service where the engagement agreement does not allow it.
- Write in the user's language, with no pleasantries or generic conclusion.

In OSINT, mistakes are costly: a false, outdated or misattributed claim in a deliverable damages your credibility and can harm a person. This skill acts as a methodical reviewer before writing up. You give it what you found and how; it questions the soundness of each item and tells you what is still to be established.

What the skill produces

  • For each claim: the state of the source (primary or relayed), its publication date and collection date, and the archived copy to keep.
  • A triangulation: how many genuinely independent sources confirm the claim, and which ones merely copy each other.
  • A confidence level (confirmed, probable, possible, unverified) with the reason, and the checks that would raise it.
  • The risks of bias (confirmation, single source, items chosen because they fit) and of false attribution (namesakes, impersonated accounts, reused images).
  • Points of caution about image and document metadata, which may be missing, stripped or forged.
  • A reminder to minimise personal data: keep only what the objective of the engagement justifies.

Install the skill

In Claude Code, for all your projects:

mkdir -p ~/.claude/skills/osint-verification
curl -fsSL -o ~/.claude/skills/osint-verification/SKILL.md \
  https://arsenal.bearops.fr/telechargements/skills/verification-osint/en/SKILL.md

For a single project, put the file in .claude/skills/osint-verification/SKILL.md at the root of the repository. With another agent, add the file’s content to its project instructions.

Use it

List the items to verify, with their origin:

Engagement: mapping the public exposure of the organisation Example (domain target.example).
Items to verify before the report:
1. A public code account (profile "dev-example") seems to belong to a system administrator.
   Clue: same handle as on a forum, an old commit signed with a first name.
2. A PDF published on target.example mentions a pre-production server.
   Collection date: 12 November; PDF date: unknown.
For each item: provenance, cross-checks to attempt, confidence level and what is missing.

The skill does not browse, archive anything or query any source: you carry out the checks it suggests, within your engagement. Once the items are qualified, structure them into a deliverable with the OSINT engagement summary skill. The open-source exposure checklist provides the collection framework upstream.

Limitations

  • The skill reasons on what you describe to it. It cannot confirm that a page exists, read metadata or date an image.
  • A confidence scale is a communication tool, not a measurement: two analysts may rate differently. Keep the written justification.
  • GDPR minimisation depends on your legal framework and your engagement agreement; the skill recalls the principles and is not legal advice.
  • Do not send any personal data or client data to an online AI service unless your engagement agreement allows it. Replace names with neutral identifiers ("Person A") where necessary.

Sources

Related resources

  • AI prompt · Infrastructure and AD pentesting · OSINT · Cloud

    Attack surface prompt

    Prompt that analyses an asset inventory from the authorised scope (domains, IPs, services, cloud) to prioritise testing: exposure, criticality, checks.

    Published

  • Cheatsheet · OSINT

    Domain OSINT cheatsheet

    Passive collection on an organisation’s domain: whois and RDAP, DNS with dig, certificate transparency, subdomains, SPF and DMARC records.

    Published

  • Methodology · OSINT

    Organisation OSINT exposure

    Methodology to assess an organisation’s OSINT exposure for an authorised client: legal framework, infrastructure, people, leaks, code, brand, reporting.

    Published

BEAROPS

Need an OSINT investigation?

BEAROPS carries out OSINT investigations lasting from a few days to several weeks, using open sources within a legal framework.