audit-friction

Friction Audit Skill

Mine Claude Code conversation logs to identify and categorize friction — repeated failures, stuck loops, tool errors, and misunderstanding cycles that cost the most effort. Surfaces patterns causing the most resistance and recommends concrete mitigations.

When to Use

• User says "audit friction", "find friction", "friction audit", or "what keeps going wrong"
• User wants to know where workflows repeatedly stall or fail
• User wants to identify systemic tooling or workflow issues across sessions

Arguments

The user may provide a "since" date (e.g., 2/7, 2026-02-07, last month). If not specified, use AskUserQuestion to ask what the earliest lookback date should be before proceeding. If the resulting window contains no logs, fall back to the last 30 days and note the adjustment.

Critical Constraints

NEVER:

• Modify any source code files
• Create files outside {{AUTOSKILLIT_TEMP}}/audit-friction/ directory
• Have subagents write files — they return all findings in response text only
• Analyze subagent log subdirectories (*/subagents/) — top-level session files only
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Use subagents heavily for parallel log analysis
• All output goes under {{AUTOSKILLIT_TEMP}}/audit-friction/ (create if needed)
• Final report: {{AUTOSKILLIT_TEMP}}/audit-friction/friction_audit_{YYYY-MM-DD_HHMMSS}.md
• Report the file and line counts to the terminal before choosing analysis mode

Friction Categories

Friction is any pattern where repeated effort yields no progress:

1. Tool Failure Loops — Same tool called 2+ times with repeated errors (file not found, permission denied, command not found, network errors)
2. Approach Reversals — Multiple strategy pivots for the same goal: tries approach A, fails, tries B, fails, tries C
3. Test Fix Cycles — Repeated test-run → fail → fix → test-run → still fails sequences without resolution
4. Misunderstanding Cycles — User corrections indicating the model misread the task; repeated wrong-direction attempts before course correction
5. Search Failures — Repeated failed searches for relevant code (wrong file paths, grep returning nothing, symbol not found)
6. Build/Compile Errors — Recurring build or compilation failures that block progress across multiple attempts
7. Permission/Access Blockers — Tool calls denied, files inaccessible, repeated approval interruptions halting flow
8. Context Re-exploration — Model re-investigates already-covered ground within the same session, forgetting earlier findings