The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of local scripts and command-line utilities to manage its pipeline. It invokes Python scripts for specification filling, frame running, and status building (e.g., br-fill.py, br-run.py, br-queue.py). It also utilizes node to execute route-shots.mjs for visual QA tasks and the gh (GitHub CLI) for synchronizing frames with external issues.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process raw, untrusted user documents and test failure logs, which are then interpolated into prompts for specification generation and code revision.
Ingestion points: Raw files located in the llmwiki/raw/ directory and test outputs (via the {{verify_output}} placeholder) are the primary entry points for external data.
Boundary markers: The revise-prompt.md template utilizes clear Markdown headers and instructional blocks to separate context from instructions.
Capability inventory: The revision sub-agent is limited to a specific set of tools (Edit, Write, Read, Grep, Glob) and is explicitly denied shell and network access.
Sanitization: The skill architecture includes a mandatory harness described as providing diff-jail (reverting out-of-scope changes), test-hash (detecting unauthorized test modifications), and scope-clean (enforcing file system boundaries) to mitigate malicious code modifications.
[DYNAMIC_EXECUTION]: Code revision is performed dynamically by a sub-agent using claude -p with a restricted toolset. This sub-agent follows instructions generated from a template (revise-prompt.md) that is populated at runtime with information from specific work frames and test results.
[EXTERNAL_DOWNLOADS]: The skill references and adopts conventions from external sources, including github.com/github/spec-kit for specification templates and the automazeio/ccpm methodology for issue synchronization. These references are documented neutrally as part of the skill's operational framework.

br