---
name: prism-evidence-auditor
description: "Use when an AI client must check claim support, Project knowledge health, weak evidence, or published-post evidence quality."
---

# Evidence auditor

Find weak support, stale coverage, contradictions, and next verification steps for important claims.

## Before you start

Copy this guidance into ChatGPT, Claude, Codex, Grok, or another client that accepts custom instructions. It changes the client's instructions only; it does not grant Prism access. Direct MCP calls require the client's current MCP support, a configured connection, and separately authorized scopes.

Scopes this workflow may need:

- `knowledge:read`
- `companies:read`
- `claims:read`
- `posts:read`

The installed skill does not grant these permissions. Use only the scopes and Projects authorized on the current Prism MCP connection.

## Workflow

1. Identify the target Project, company, claim set, or published post.
2. Check Project knowledge health when indexing coverage affects confidence.
3. Retrieve company context and claims, then inspect exact evidence for the claims that matter most.
4. Use the weak-claims or published-post health tools for a bounded audit instead of making a general judgment from snippets.
5. Report evidence strength, age, contradictions, and the next verification action separately.

## Tool routing

- `get_project_knowledge_health`: Measure current-revision coverage and indexing health before judging retrieval quality.
- `get_company_claims`, `get_evidence_for_claim`: Move from claim inventory to exact supporting or contradicting evidence.
- `list_weakly_sourced_claims`: Find the claims that most need sourcing work.
- `get_published_posts`, `get_published_post_evidence_health`: Select a post, then inspect its source and claim health.

## Expected result

- A ranked list of well-supported, weak, stale, and contradicted claims.
- Evidence gaps include the missing source type or verification step needed next.
- Published-post checks identify stale sources and unsupported claims without editing the post.

## Boundaries

- A skill does not grant Prism access. The connection must separately authorize every required scope and Project.
- Call only tools verified callable by the current AI host. Use discover_prism_actions for granular actions. If a default workflow tool is missing, follow the capability-handshake recovery guidance instead of inventing a substitute tool.
- Read before writing. Mutate Prism only when the user asked for a durable change and the connected tool exposes that exact action.
- Keep retrieved evidence separate from analysis. Cite Prism source ids, titles, snippets, locations, and chunk ids when available.
- Do not mark a claim false merely because indexed evidence is incomplete.
- Do not revise a thesis or published post from an audit. Return findings or create a separate authorized proposal workflow.