The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains sophisticated instructions designed to override the AI's system prompt. Specifically, scripts/context.mjs includes an AUTONOMY_DIRECTIVE_CHECK directive that tells the AI to ignore harness-level assertions about its own autonomy or the user's presence, forcing the AI to continue with the skill's interactive steps regardless of the system context.
[PROMPT_INJECTION]: The skill uses instructions to conceal its actions from the user. reference/craft-floor.md explicitly tells the AI to "build without announcing the checklist."
[PROMPT_INJECTION]: Instructions for sub-agents (e.g., in agents/impeccable_asset_producer.toml) use role-play and overriding patterns: "Step fully out of the work you just finished, adopt only this file's instructions for the pass."
[PRIVILEGE_ESCALATION]: The skill attempts to bypass security sandboxes. In scripts/live-copy-edit-agent.mjs, the AI is instructed to run sub-agents using high-risk flags like --dangerously-bypass-approvals-and-sandbox for the codex tool and --permission-mode bypassPermissions for the claude tool.
[PERSISTENCE]: The skill implements a persistent "hook" mechanism via scripts/hook-admin.mjs. It installs configuration into project-level metadata files (e.g., .claude/settings.local.json, .cursor/hooks.json, .github/hooks/impeccable.json) that cause the skill's scripts to execute automatically whenever the user edits files or the agent session ends.
[INDIRECT_PROMPT_INJECTION]: The skill has a large attack surface for indirect prompt injection. It ingests and processes untrusted project data (HTML, CSS, and JS source files) using complex logic like Svelte AST parsing and regex-based design detection. This data is used to drive significant capabilities, including local command execution and file system modifications.

impeccable