The Agent Skills Directory

[PROMPT_INJECTION]: The scripts/prompt_hook.py script generates standing instructions intended to be injected into every agent prompt. These instructions force the agent to use specific routing and memory skills, overriding default behaviors and imposing a mandatory framework.
[PERSISTENCE]: The skill maintains task state across sessions by writing to a local ledger file in ~/.agent/tmp/ as defined in references/multi-goal-ledger.md. Additionally, the prompt hook script acts as a persistent instruction layer for the agent environment.
[INDIRECT_PROMPT_INJECTION]: The multi-goal management system ingests and stores raw user strings in a ledger, which the agent is then instructed to treat as a source of truth, creating a surface for indirect instructions to influence behavior.
Ingestion points: The references/multi-goal-ledger.md protocol parses and stores original user command lines and goal descriptions.
Boundary markers: Markdown delimiters are used for organization, but the agent is told the ledger content is authoritative, potentially bypassable by instructions embedded in the user-provided goal text.
Capability inventory: The framework includes action and implementation routes (nerd-execute) and local validation tooling (scripts/validate_skills.py).
Sanitization: The skill instructions focus on redacting credentials but lack specific sanitization or escaping for natural language instruction boundaries within the ledger content.

nerd-smart