The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes past conversation logs and brief artifacts to identify steering moments and propose code changes to other skills. This creates a surface for indirect prompt injection where malicious instructions embedded in a conversation could influence the agent's logic during the improvement phase.
Ingestion points: Reads conversation history from ~/.codex/sessions/ and ~/.claude/projects/, along with HTML visual briefs from project .briefs/ directories.
Boundary markers: The instructions do not define specific delimiters or instructions to ignore potential commands embedded within the transcripts being analyzed.
Capability inventory: The skill can perform file modifications to other skills, manage git branches, and open or merge GitHub Pull Requests.
Sanitization: No significant sanitization of the conversation content is performed beyond filtering out common system-generated shims.
[REMOTE_CODE_EXECUTION]: The skill performs routine updates of platform-provided meta-skills from the NPM registry.
Evidence: In Step 0b, it executes npx skills update to refresh external tools located in the .agents/skills/ directory.
[COMMAND_EXECUTION]: The skill relies on external CLI tools and local scripts to manage its workflow.
Evidence: It invokes git for repository synchronization, gh for Pull Request operations (including automated merging for state-only updates), and runs python3 for conversation discovery and artifact auditing (scripts/list_conversations.py and scripts/list_briefs.py).
[DYNAMIC_EXECUTION]: The use of npx allows for the dynamic loading and execution of packages from a remote registry at runtime.
[EXTERNAL_DOWNLOADS]: The skill fetches updates for its internal tools from the NPM registry as part of its pre-flight setup.

skill-improver