The Agent Skills Directory

[PROMPT_INJECTION]: The SKILL.md file includes a 'Prompt Architect Overlay' which uses role definition and persona instructions ('prompt-architect-enhanced specialist') to guide the agent's behavior. While this is a form of instruction override, it is intended to ensure deterministic execution of the skill's features.
[PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection because it processes untrusted data (source code) and provides summaries to the AI. If a codebase contains malicious instructions within comments or strings, the agent could potentially be influenced when reading the structured output.
Ingestion points: tldr warm, tldr extract, and tldr context commands which read all files in a project.
Boundary markers: None explicitly specified in the tool output descriptions; content is delivered as structured markdown/text summaries.
Capability inventory: The tool can read/write to a local cache directory (.tldr/), execute CLI commands, and run a background daemon.
Sanitization: No explicit sanitization or filtering of codebase strings/comments is mentioned to prevent LLM instruction obedience.
[EXTERNAL_DOWNLOADS]: The skill instructions require the installation of the llm-tldr package from the official PyPI registry, which also pulls in dependencies such as sentence-transformers and faiss-cpu for its semantic search functionality.
[COMMAND_EXECUTION]: The workflow involves running various local commands via the tldr CLI to warm indexes, generate call graphs, and perform semantic searches. It also includes starting and stopping a background daemon process.

llm-tldr