The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a maintenance script sync-classify-prompt-from-core-product.mjs that uses execSync to run a Go-based tool (go run ./tools/classify-prompt-export/). This command is used to synchronize prompt definitions from a local developer repository.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to ingest and classify untrusted human inputs and assistant messages, creating a vulnerability surface where malicious instructions in the data could influence the agent's reasoning.
Ingestion points: Data is received via the content, assistant_message, and prev_message arguments as defined in SKILL.md.
Boundary markers: The instructions explicitly direct the agent to ignore "ambient auto-injected blocks" like browser context tags and focused exclusively on the "My request" header.
Capability inventory: The skill is granted Bash access and contains a script capable of executing system shell commands.
Sanitization: The skill implements text normalization (stripping redactions and markdown noise) and aggressive length capping (rune-based truncation) for assistant and previous message context.
[DYNAMIC_EXECUTION]: The sync-classify-prompt-from-core-product.mjs script performs dynamic execution by identifying a local repository path (potentially via environment variables) and running code within that directory using execSync.

cawplan-internal-qa-coding-humaninputs