The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user-edited Markdown files to triage research ideas and track findings.
Ingestion points: Scripts such as sweep-inbox.py and retriage.py read content from INBOX.md, LEDGER.md, and ROADMAP.md to update project state.
Boundary markers: The skill lacks explicit sanitization or delimiters to prevent the agent from being influenced by instructions embedded within the research text.
Capability inventory: The skill is capable of modifying files and executing local system commands (git) for metadata collection.
Sanitization: The scripts use regular expressions to parse structured data but do not sanitize natural language content against malicious prompt instructions.
[COMMAND_EXECUTION]: The script scripts/snapshot-provenance.py executes local system commands to capture the state of the codebase for reproducibility.
Evidence: The script uses subprocess.run to call git rev-parse and git diff to record commit hashes and detect uncommitted changes.
Context: These executions are restricted to read-only git operations for metadata collection, which is consistent with the skill's stated purpose.
[DYNAMIC_EXECUTION]: The skill dynamically generates project structures and report files based on internal templates.
Evidence: The init.sh and new-experiment.py scripts use template substitution to create new experiment folders and documentation.
Context: This behavior is used for project scaffolding and initialization based on safe, bundled templates.

experiment-knowledge-harness