The Agent Skills Directory

[SAFE]: The skill processes project artefacts from local directories to identify technical learnings. It follows a structured workflow that enforces read-only access to source artefacts and requires explicit user approval before writing any updates to the knowledge base or standards files.- [COMMAND_EXECUTION]: The skill invokes the internal /estimate tool (mode E3) to compare actual work time against forecasts. This execution is limited to internal state processing and local file appending within the workspace, with no exposure to external command shells or network operations.- [PROMPT_INJECTION]: Analysis for indirect prompt injection surface identifies that the skill ingests content from user-influenced artefact files in the artefacts/ directory. This potential vulnerability is mitigated by a mandatory evidence chain: 1. Ingestion points (various markdown artefacts), 2. Boundary markers (process segmentation into steps), 3. Capability inventory (local file writes to .github and project folders), and 4. Sanitization (previewing proposed text and requiring manual 'yes' confirmation from the operator).- [SAFE]: The skill includes security-conscious design elements, such as explicitly stating it cannot edit its own or other SKILL.md files directly. Any proposed changes to agent instructions are written as diff files to a workspace folder for manual PR-based review, ensuring the supply chain integrity of the agent skills.

improve