Architecture Critic
ReviewAudited by ClawScan on May 6, 2026.
Overview
Architecture Critic is a coherent review helper that calls Anthropic with a scoped repository snapshot and stores a local verdict, but users should notice the external data sharing, API key use, and build-gating behavior.
Before installing, confirm you are comfortable sending task briefs and repository metadata to Anthropic and paying for those calls. Configure the Anthropic key securely, avoid secrets in briefs or commit messages, review the installed script and dependency versions, and keep human override available for the build gate.
Findings (4)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
Your task brief and some repository context leave your machine and are processed by Anthropic.
This documents the external data flow to Anthropic. It is purpose-aligned and disclosed, but may expose proprietary task and repository metadata to the configured LLM provider.
sends_to_external: true ... data_sent: task brief, repository file tree, dependency list, recent git commits
Use only on projects where sending this metadata to Anthropic is allowed; avoid putting secrets in briefs, commit messages, or config files.
Running the skill can use your Anthropic account and incur provider charges.
The skill uses a local Anthropic credential to make provider API calls. This is disclosed and bounded to the stated LLM review purpose.
reads your configured LLM provider API key from: 1. ANTHROPIC_API_KEY environment variable ... 2. Your local ~/.openclaw/openclaw.json config file
Provide a dedicated, least-privilege Anthropic key where possible, keep it out of briefs and logs, and monitor usage costs.
A model-generated verdict may stop or delay a build until the brief is revised or a human overrides.
The skill intentionally changes workflow stopping conditions based on the critic verdict. This is the advertised behavior, but it can materially affect an agent-driven build process.
REVISE → Build does NOT start ... REJECT → Build STOPS ... Critic infrastructure down → build is BLOCKED
Use this as a gate only where that workflow is desired, and keep the documented human override path available.
Users need to install and trust the local helper environment and Python dependency.
The skill relies on local runtimes and the Anthropic Python package. This is expected for the purpose, but the package is not version-pinned and the registry summary says no install spec.
runtime:\n requires:\n - python3 >= 3.8\n - bash >= 4.0\n - pip: anthropic
Install dependencies from trusted sources, consider pinning the Anthropic SDK version, and review the installed script before using it in sensitive workflows.