The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core functionality involves ingesting and processing untrusted external Product Requirement Documents (PRDs) from Feishu/Lark links or local files, creating a potential surface for adversarial instructions embedded in those documents.
Ingestion points: Untrusted PRD content is fetched via lark-cli commands or read from local markdown files as specified in SKILL.md.
Boundary markers: The skill mitigates risks by enforcing strict adherence to structured templates (digest-template.md and report-template.md) and a specific audit taxonomy (dimensions.md), which constrains the model's output format and logic.
Capability inventory: The skill possesses capabilities to execute CLI tools (lark-cli) and perform local file operations (reading/writing to the sdlc/ directory and /tmp/).
Sanitization: The instructions explicitly forbid the model from "brain-filling" or adding its own interpretations (忠实转述原文，不评价、不脑补、不优化), and require every finding in the audit phase to be anchored to a specific quote from the original text (引文 ≤20 字).
[COMMAND_EXECUTION]: The skill invokes an external CLI tool to retrieve data from collaboration platforms.
Evidence: SKILL.md defines specific commands for fetching document content and comments: lark-cli docs +fetch --doc "<URL>" --doc-format markdown and lark-cli drive +list-comments --url "<URL>" .... These calls are structured and scoped to the task of requirement intake.
[EXTERNAL_DOWNLOADS]: The skill retrieves data from well-known external collaboration and AI services.
Evidence: The skill targets feishu.cn and doubao.com for document and comment retrieval. These are recognized services provided by ByteDance and are used here for legitimate PRD analysis purposes.

sdlc-intent