RCP-CWK-013 Cowork Factual Claim Validator

Validates factual claims in CRAFT deliverables using Cowork’s WebSearch tool 

Validates factual claims in CRAFT deliverables using Cowork’s WebSearch tool for live verification, cross-references against project files, and assigns confidence levels with source citations. Replaces the Core recipe’s manual Deep Research prompt workflow with inline automated search.

Cowork Factual Claim Validator with WebSearch

Cowork | QA & Validation | Difficulty: Medium | Time: 5–15 min
Tags: Introduced in Beta, Claim, Validation, WebSearch, Confidence, Source, Hallucination, QA, Cowork

TL;DR

What It Does
Validates factual claims in CRAFT deliverables using Cowork’s WebSearch tool for live verification, cross-references against project files, and assigns confidence levels with source citations. Replaces the Core recipe’s manual Deep Research prompt workflow with inline automated search.
How It Works
The AI scans your content for factual claims across five categories: statistics and numbers, temporal claims, attributed statements, technical specifications, and CRAFT-specific claims. Each claim is verified using a five-tier source hierarchy — project files (highest), git history, live web search, general knowledge, and no source (lowest). The recipe checks for internal contradictions and runs hallucination gates, then produces a validation report with confidence scores and recommended actions.
What To Expect
A structured validation report in 5–15 minutes listing every identified claim with its source, confidence level, and any flags. Contradictions and hallucination risks are highlighted. Unsourced claims get four resolution options.
Best Results When You
Run this recipe on content before it leaves the project — blog posts, website copy, documentation, or any deliverable that makes factual assertions. Specific claims validate more effectively than generic statements.
Typical Time
5–15 minutes depending on the number and complexity of claims. Content with many specific statistics or dates takes longer due to cross-referencing and web search.
Difficulty
Medium — you provide the content to validate. The AI handles claim identification, source checking, cross-referencing, hallucination detection, and report generation. You make decisions on how to handle unsourced claims.

How To Start

Provide the content you want validated. The AI scans for factual claims and verifies each one.

STEP 1Provide Content for Validation

Paste or reference the content to validate. Optionally specify the validation mode and which claim types to focus on.
Available parameters
  • content · string · required
    The content to validate for factual claims.
  • validation_mode · string · optional · default: “automatic”
    Validation aggressiveness. Options: “automatic,” “strict,” “lenient,” “inline_search.”
  • claim_types · list · optional · default: all types
    Types of claims to validate: “statistics,” “dates,” “quotes,” “technical_specs,” “craft_claims.”
  • search_enabled · boolean · optional · default: true
    Allow WebSearch for live verification.
  • project_file_check · boolean · optional · default: true
    Cross-reference against CRAFT project files.
Invocations
Standard Validation
#H->AI::Directive: (Validate factual claims — Cowork mode)
Strict Mode
#H->AI::Directive: (Validate factual claims — Cowork mode with validation_mode=strict)
CRAFT-Only Claims Check
#H->AI::Directive: (Validate factual claims — Cowork mode with claim_types=[craft_claims])

Usage Examples

Two scenarios — validating a blog post draft and catching a contradiction in project documentation.
Example 1 · Blog post validation
User: Validate this draft blog post about CRAFT’s recipe system.
AI: Scanning content for factual claims. Identified 8 claims across 3 categories. Claim 1: “CRAFT has 128 active recipes” — Sourced: cookbook file (Tier 1), confidence 95%. Claim 2: “Cowork was introduced in March 2026” — Sourced: chat history (Tier 1), confidence 90%. Validation complete. 7 sourced, 1 unsourced. The unsourced claim has no supporting evidence. Options: remove, qualify, investigate, or accept with risk.
Example 2 · Contradiction detection
User: Check the claims in our updated framework overview.
AI: Contradiction detected. The overview says “P067 has completed 15 sessions” but the chat history shows 12 handoff entries. Which is correct? Resolve before proceeding.

How AI Reads This Recipe

When this recipe is triggered, the AI runs a 6-step factual claim validation pipeline with multi-source verification and hallucination safeguards. The AI should:
  1. IDENTIFY factual claims by scanning content for five categories: statistics and numbers, temporal claims, attributed statements, technical specifications, and CRAFT-specific claims (recipe counts, session numbers, tool behavior assertions).
  2. CHECK each claim against the five-tier source hierarchy: Tier 1 (project files, 90–100%), Tier 2 (git history, 85–95%), Tier 3 (WebSearch, 70–90%), Tier 4 (general knowledge, 60–80%), Tier 5 (unsourced, 0%). Stop at the first tier that provides a source.
  3. VERIFY each claim with its source, recording the file, section, entry ID, URL, or domain. Assign confidence based on source tier.
  4. CROSS-REFERENCE claims for internal contradictions — check against Lessons Learned, handoff records, and roadmap status. Flag contradictions and require resolution before proceeding.
  5. RUN hallucination gates on every claim: (1) Was the file actually read this session? (2) Is the claim from the file or reconstructed from memory? (3) Does the claim match what the file says? (4) Are there untested assumptions about Cowork behavior? Apply 20% confidence penalty for hallucination risk.
  6. PRODUCE a validation report with total claims, breakdown by source tier, contradictions, and hallucination flags. For unsourced claims, offer four options: remove, add qualifying language, delegate to CWK-008 sub-agent, or accept with noted risk.
Never fabricate sources. If WebSearch fails or a domain is blocked by the egress proxy, note the domain and fall back to project files plus general knowledge. The hallucination check is mandatory for every claim regardless of initial confidence.

When to Use This Recipe

Use this recipe when you:
  • Are preparing content for publication and need to verify factual claims.
  • Want to catch contradictions between different project files before they propagate.
  • Need confidence-scored verification of CRAFT-specific claims.
  • Want automated fact-checking that leverages both project files and live web search.
Do not use this recipe when:
You are drafting creative content, brainstorming ideas, or writing opinion pieces where factual claims are not the focus. Running claim validation on intentionally speculative or exploratory content wastes time and generates false flags.

Recipe FAQ

Q.What are the five source tiers?

Tier 1: CRAFT project files (confidence 90–100%). Tier 2: git history (85–95%). Tier 3: live web search (70–90%). Tier 4: general knowledge (60–80%). Tier 5: no source (0%). The AI checks each tier in order, stopping when a claim is sourced.

Q.What happens when WebSearch is blocked by the egress proxy?

The AI notes the blocked domain and falls back to project files and general knowledge. It never fabricates a source to replace a blocked search result.

Q.How does the hallucination check work?

Four gates: (1) Was the source file actually read in this session? (2) Is the claim based on the file or reconstructed from memory? (3) Does the claim match what the file actually says? (4) Are there untested assumptions about Cowork behavior? Failing any gate triggers a 20% confidence penalty.

Q.What are the four options for unsourced claims?

Remove the claim from content, add qualifying language like “reportedly” or “approximately,” delegate to a CWK-008 sub-agent for deeper verification research, or accept the claim as-is with a noted risk annotation.

Q.Can I validate only CRAFT-specific claims?

Yes. Set claim_types to [“craft_claims”] to focus validation on recipe counts, session numbers, tool behavior assertions, and other framework-specific facts. This is faster than full validation and useful for internal documentation.

Q.How does this recipe differ from the Core Source Validator?

The Core recipe uses a manual Deep Research prompt workflow — you export claims, run them through external AI tools, and import results. This Cowork version replaces that with inline automated search using WebSearch and direct project file access via the Read tool. Same validation rigor, faster execution.

Version History

Changes to this recipe over time. Most recent first.
v1.00a 2026-02-28
Initial release. Extends Core Source Validator with Cowork automation: 6-step validation pipeline, five claim categories, five-tier source hierarchy, cross-reference contradiction detection, hallucination gates with confidence penalty, structured validation report with four resolution options. Audit trail logging per R-12.

Get this recipe with CRAFT for Claude Cowork

Cowork recipes ship bundled with CRAFT for Claude Cowork — there’s no separate download. Clone the framework once, and your AI runs every recipe automatically when invoked.

Pull anytime to stay on the latest version — free to clone, no login or email required.

Then start your session

Once CRAFT is in your project folder, open a new Cowork session and ask Claude to initialize. For example:

You

Please initialize my CRAFT session.

Claude

CRAFT session ready. Your project is loaded, your persona is active, and your recipes are available. What would you like to work on?

What is CRAFT for Claude Cowork?

Not familiar with Git? Download as a ZIP

No command line needed. Just download, move, and unzip:

  1. Open the CRAFT framework repo on GitHub.
  2. Click the green Code button, then choose Download ZIP.
  3. Move the downloaded ZIP into your Claude Cowork project folder.
  4. Unzip it: double-click on Mac, or right-click → Extract All on Windows.

Similar Posts