The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local scripts (python evals/content/score_run.py) and git commands (git log, git commit) as part of its automated content scoring and reporting workflow. These operations are performed within the workspace context.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of 'signals', which include transcripts and PR descriptions from GitHub. This creates a surface for indirect prompt injection if the source data contains malicious instructions.
Ingestion points: Signal files located in signals/<YYYY-MM-DD>-*.md (Step 0 and Step 1).
Boundary markers: The instructions explicitly require the agent to remain 'Grounded, not fabricated' and carry citations from the source (Rail 2).
Capability inventory: The agent can write files to the content/03-drafts/ directory, perform git commit operations, and execute a Python scoring script (Step 8).
Sanitization: The skill implements a multi-stage review process, requiring both a 'copy-reviewer' (Category-specific ICP roleplay) and a 'copy-editor' to pass each text format before completion (Rails 4, Steps 3-6).
[DATA_EXFILTRATION]: While the skill primarily operates on internal drafts, it is designed to prepare content for the recoup-internal-consulting-linkedin-publisher skill. The instructions mitigate the risk of unauthorized external data transfer by explicitly stating 'Never auto-publish' and requiring owner review for all audience-facing content.

recoup-internal-consulting-nightly-content