What problem does it solve? Agents infer user intent from phrasing alone, which is inconsistent under long or compacted context. This Skill gives users explicit, machine-detectable one-word prefixes (like h:, f:, q:) that deterministically tell the agent how to treat the prompt that follows. ## Core Features & Use Cases - Marker detection and context injection: A stdlib-only Python UserPromptSubmit hook reads the prompt, matches a leading prefix against markers-context.json, and appends the marker's instruction text via additionalContext — preserving prompt caching by never rewriting the original prompt. - Six built-in markers: h:/hint: (validate before acting), f:/feedback: (high-priority correction with cited evidence), e:/extension: (scope widening), q:/queue: (run after current work), i:/important: (high priority), and i!: (immediate interrupt); every marker also accepts a ! importance token for interrupt-grade handling. - Audit trail and drift detection: Detected markers are recorded to a SQLite marker_state table (capped at 100 entries) when a .ai-badger directory exists, and two consecutive feedback turns trigger a consolidated-restart advisory. - Use Case: A user types f: the login test still fails with TimeoutError and the agent immediately receives injected instructions to treat it as high-priority feedback and cite the failing evidence before proposing a fix. ## Quick Start Register the user_prompt_hook.py script as a UserPromptSubmit hook in your Claude Code settings, then start any prompt with a marker like "h:" to have the matching behavior instruction injected automatically.
