The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from project files and conversation history during the knowledge update process.
Ingestion points: Git diff output, file content from docs/, research/, and README.md, and the session's dialog history in SKILL.md.
Boundary markers: Employs a manifest-based tracking system using expected_count vs verified_count to ensure no items are hallucinated or omitted.
Capability inventory: The skill utilizes Write, Agent, and Skill tools to modify the local environment and trigger sub-tasks.
Sanitization: Uses a fresh-context sub-agent for review and requires mandatory human confirmation via AskUserQuestion before committing any knowledge changes.
[PERSISTENCE]: The skill maintains technical decisions and learned patterns across sessions by writing to persistent memory paths.
Evidence: It writes to ~/.claude/projects/<project-hash>/memory/MEMORY.md and ~/.claude/skills/learned/.
Purpose: This is the primary stated purpose of the skill to ensure technical context is not lost between sessions.
[DYNAMIC_EXECUTION]: The skill supports tool composition by chaining other skills at the end of its lifecycle.
Evidence: Step 7 parses $ARGUMENTS to dynamically trigger a follow-up skill, such as a pull request utility, using the Skill tool.
Safety: Pipeline execution is conditional and only occurs after the user has confirmed the previous steps.
[COMMAND_EXECUTION]: The skill uses local shell tools to verify the integrity of the documentation process.
Evidence: Utilizes git diff, grep, and glob to audit sub-agent behavior and ensure that documentation changes accurately reflect the source code state.

update