The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a 'Prompt Self-Improvement Loop' allowing the agent to modify its own local workspace instructions and contracts (references/prompt-self-improvement.md). While this introduces risks associated with self-modifying code, the skill provides extensive mitigation strategies:
Mandatory 'Layer 3 Blocking Critic' using a different model family to provide independent review.
Automated 'Invariant Checks' to prevent the deletion of safety rules and approval buckets.
A required 'Smoke Test' before changes are committed.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from external, untrusted sources such as app reviews, Reddit, and web forums to discover user needs (references/battle-tested-patterns.md). This creates an attack surface for indirect prompt injection. Mandatory Evidence Chain:
Ingestion points: The Scout role mines reviews and social posts for unmet needs (referenced in references/battle-tested-patterns.md).
Boundary markers: The worker prompt (assets/prompts/worker.md) enforces structured artifact contracts and strictly defined JSON output blocks.
Capability inventory: The skill can write artifacts, update local state, and perform tool-based research (browser/search).
Sanitization: Relies on independent 'Critic' roles (references/rubber-duck-review.md) to evaluate evidence and provide an 'Independence Verdict' before results are accepted.
[COMMAND_EXECUTION]: The skill includes Python scripts for workspace initialization (scripts/init_factory_workspace.py) and smoke testing (scripts/smoke_test_initializers.py) which execute shell commands via subprocess.run. These scripts are intended for local environment management and use controlled, hardcoded arguments rather than external user input.

opportunity-factory