CRAFT for Cowork: A Structured QA Framework for Claude
Patch your Claude integration to use the RCP‑CWK‑024 gate logic and confidence scoring to enforce evidence‑based claims.
Patch your Claude integration to use the RCP‑CWK‑024 gate logic and confidence scoring to enforce evidence‑based claims.
Summary
The article introduces CRAFT for Cowork, a structured working environment for Claude that treats AI reliability as a measurable engineering problem. It details the four verification gates—File‑pointability, Read‑vs‑Reconstructed, Lessons‑Learned conflict, and Untested assumption—implemented in the reusable sub‑routine RCP‑CWK‑024, and explains how each gate flags claims that fail to meet the required evidence hierarchy. The framework assigns a 0‑100 confidence score based on source hierarchy, with a 10‑point penalty after 70 % token usage to model late‑session decay, and it has already caught nine Week‑1 content files that incorrectly labeled CRAFT as “open source” and ~40 % drift in tracking‑file state tables. The article also highlights the dual‑license structure (BSL 1.1 for the spec, proprietary content) and the free public beta on GitHub, encouraging developers to adopt the framework for more reliable Claude interactions.
Key changes
- Four gates in RCP‑CWK‑024 flag claims that fail evidence checks.
- Confidence scoring uses source hierarchy (80‑100, 50‑79, 30‑49, 0‑29) with a 10‑point late‑session penalty.
- Cross‑file audit recipe RCP‑CWK‑036 runs every 5‑10 sessions and caught ~40 % drift.
- License is dual‑licensed: BSL 1.1 for spec, proprietary for content, switching to Apache 2.0 on Jan 1 2029.
- Free public beta on GitHub at https://github.com/CRAFTFramework/craft-framework.
- Framework caught nine Week‑1 files incorrectly labeling CRAFT as open source.
- The framework can be used to validate claims before reporting results.