claims

name

claims

description

This skill should be used when the user asks to "verify claims", "check claims against sources", "claims dashboard", "show claims", "resolve claim", "submit claim", "claim verification", "fact-check claims", "re-verify claims", "review deviations", "inspect claim", "what claims are pending", or when any plugin needs to submit claims for source verification. Orchestrates the full claim verification lifecycle: ingestion, source-based verification, deviation detection, dashboard presentation, and user-guided resolution.

Claims Verification Orchestrator

Purpose

Manage the full lifecycle of sourced claims: accept claims (individually or in batch), verify them against their cited source URLs, detect deviations, present findings to the user, and guide resolution. This skill orchestrates the pipeline; verification work is delegated to the claim-verifier agent.

Not for:

• Generating claims from scratch (that is the submitting plugin's responsibility)
• Making editorial decisions (the user is the authority)
• Presenting deviation findings as definitive facts (they are LLM-based assessments)

Operating Modes

Support five operating modes, determined by the mode parameter or inferred from user intent:

Mode	Trigger	Description
submit	"submit claim", batch from plugin	Ingest new claims into the registry
verify	"verify claims", "check claims"	Run verification pipeline on unverified claims
dashboard	"claims dashboard", "show claims"	Display status overview
inspect	"inspect claim "	Show detailed deviation evidence for one claim
resolve	"resolve claim "	Present resolution options for a deviated claim

Workspace Initialization

Before any operation, ensure the workspace exists:

1. Determine working_dir — passed as parameter, or use the current working directory
2. Check for {working_dir}/claims/claims.json
3. If missing, initialize:
   • Create claims/, claims/sources/, claims/history/
   • Write initial claims.json with empty claims array
4. Read claims.json into working context

Mode: Submit

Accept one or more claims and add to the registry.

Required input:

• statement — the claim text
• source_url — URL of cited source
• source_title — human-readable source title
• submitted_by — plugin name or "user" for direct submission

Process:

1. For each claim, generate a unique ID: claim-<uuid>
2. Create a ClaimRecord with status unverified
3. Append to claims.json
4. Write submission event to history/{claim-id}.json
5. Report count of submitted claims to user

Batch submission: Accept an array of {statement, source_url, source_title} objects. Process all in a single registry update.

Mode: Verify

Run the verification pipeline on unverified claims (or a specific claim by ID).

Process:

Step 1: Select Claims

• If --id <id> specified, verify that single claim (even if previously verified — this is re-verification)
• Otherwise, select all claims with status unverified
• If no claims to verify, inform user and exit

Step 2: Group by Source URL

Group selected claims by source_url. Count unique URLs (K) and total claims (N). Report: "Verifying {N} claims against {K} unique sources."

Step 3: Dispatch Verification

For each unique URL group, launch a cogni-claims:claim-verifier agent via the Task tool:

Task parameters:
  subagent_type: "cogni-claims:claim-verifier"
  prompt: Include working_dir, source URL, claim IDs, claim statements

Launch agents in parallel when multiple URLs exist. Each agent:

1. Fetches the source URL (WebFetch, browser fallback)
2. Caches source content in claims/sources/{hash}.json
3. Compares each claim against source content
4. Returns verification results as JSON

Step 4: Collect Results

Read verification results from each agent. For each claim:

• Update status in claims.json (verified, deviated, or source_unavailable)
• Attach DeviationRecords if deviations detected
• Record source excerpt
• Write verification event to history/{claim-id}.json

Step 5: Present Summary

Show a brief verification summary:

Verification complete:
- {n} verified (no deviations)
- {n} deviations detected ({n} critical, {n} high, {n} medium, {n} low)
- {n} sources unavailable

If deviations with severity medium or higher exist, suggest: "Run /claims dashboard to review findings."

Mode: Dashboard

Display the claims dashboard grouped by status. Consult references/dashboard-format.md for the complete layout specification.

Process:

1. Read claims.json
2. Group claims by status
3. Sort within groups (see dashboard-format.md for sort rules)
4. Render markdown tables per section
5. Include action hints for deviated claims

Mode: Inspect

Show detailed evidence for a specific claim.

Process:

1. Look up claim by ID in claims.json
2. If claim has deviations, display each with:
   • Deviation type and severity
   • Source excerpt (verbatim)
   • Explanation
3. If claim is verified, display supporting excerpt
4. Offer to open source in browser for in-context inspection:
   • Launch cogni-claims:source-inspector agent with the source URL and relevant excerpt
   • The agent navigates to the page and highlights the passage

Mode: Resolve

Guide the user through resolving a deviated claim.

Process:

1. Look up claim by ID — must have status deviated
2. Display the claim, deviation details, and source excerpt
3. Present resolution options using AskUserQuestion:
   • Correct — prompt for corrected statement text
   • Dispute — prompt for rationale
   • Alternative source — prompt for new URL and title
   • Discard — prompt for rationale
   • Accept as-is — prompt for rationale
4. Record ResolutionRecord with user's choice
5. Update claim status to resolved
6. Write resolution event to history/{claim-id}.json
7. If action is alternative_source, optionally trigger re-verification against new source

Quality Constraints

• Conservative detection: Prefer "uncertain" findings over false positives
• Evidence-first: Always show the source excerpt — never present a deviation without evidence
• Explicit ambiguity: When the comparison is genuinely unclear, say so
• User authority: The system provides assessments; the user makes decisions
• No silent failures: Every unverifiable source is explicitly reported

Additional Resources

Reference Files

• references/verification-protocol.md — Detailed verification methodology, deviation type definitions, severity criteria, epistemic humility guidelines
• references/dashboard-format.md — Complete dashboard layout, resolution prompt format, truncation and sorting rules

Scripts

• scripts/claims-store.sh — JSON state management utility for reading/writing claims registry, generating IDs and URL hashes. Invoke via: bash "${CLAUDE_PLUGIN_ROOT}/skills/claims/scripts/claims-store.sh" <command> [args...]

Examples

• examples/claims-sample.json — Sample claims.json with claims in all statuses (unverified, verified, deviated, resolved) showing complete field structures

Cross-Plugin Contract

For the ClaimEntity data model, see the cogni-claims:claim-entity skill and its reference files.

Agents (via Task tool)

• cogni-claims:claim-verifier — Worker agent: fetches one source URL, verifies all claims against it, returns results as JSON. Launch in parallel for multiple URLs.
• cogni-claims:source-inspector — Browser agent: navigates to source URL, highlights relevant passage for user inspection.