Ai audit log

AI Audit Log Skill

Logs AI interactions into default path docs/audit/ai/<fullname>-YYYY-MM.log.md with structured entries and monthly statistics.

Invoke Syntax

/ai-audit-log                                      # log last 1 interaction, save to default path
/ai-audit-log --last=N                             # log last N interactions, save to default path (N: 1–5)
/ai-audit-log --file="path/to/output.md"           # log last 1 interaction, save to custom file path
/ai-audit-log --last=N --file="path/to/output.md"  # log last N interactions, save to custom file path (N: 1–5)

If N > 5, respond: "Max --last is 5. Running with --last=5." and proceed.

Fullname Resolution

The log filename is <fullname>-YYYY-MM.log.md. Resolve fullname as follows:

1. Context available — if human has provided their name earlier in the conversation, use it directly (no need to ask again)
2. No context — ask: "What is your full name? (Used for the log filename)" and wait for the answer
3. Normalization rules (always applied):
   • Convert to lowercase
   • Remove diacritics/accents
   • Replace spaces with -
   • Example: Nguyễn Tấn Phát → nguyen-tan-phat

Output Path Resolution

Invocation	Output path
No --file flag	docs/audit/ai/<fullname>-YYYY-MM.log.md
--file="path/to/output.md"	Exactly as specified by human

If --file is provided, the <fullname> and default directory are ignored entirely — use the given path as-is.

What AI Self-Detects

From conversation context, AI automatically extracts:

• Prompt — exact text of the human's message
• Model — current model in use
• Timestamp — current datetime (ISO 8601, UTC+7)
• Output — AI's response (text/code inline, or file summary as bullets)

Interaction Flow

1. AI resolves fullname from context or asks human (see Fullname Resolution above).
2. AI resolves output path from --file flag or constructs default path.
3. AI reads context → identifies the last N interactions (excluding the /ai-audit-log invoke itself).
4. AI displays a numbered summary of detected interactions for human confirmation.
5. AI outputs a structured evaluation form — one section per interaction.
6. Human fills in the form and pastes it back.
7. AI translates any non-English responses to English, then writes log entries.
8. AI updates (or creates) the monthly log file with new entries + updated stats.

Evaluation Form Template

For each interaction, AI outputs:

─── Interaction [N] ───
Prompt excerpt: "[first 80 chars of prompt]..."
Output type: [text | code | file(s) | mixed]

Tags (pick 1–3):
  Domain:  backend · frontend · mobile · data · devops · infra · security · performance · design
  Task:    code-gen · refactor · debugging · testing · planning · review · documentation · analysis · research · prompt-engineering
  Output:  file-output · script · query · config · diagram · report
>

Status:
  ✅ VALID       — Output correct, used as-is
  ⚠️ PARTIAL     — Partially correct, minor fixes needed
  🔄 REVISED     — Output sufficient but required significant edits before use
  ❌ INVALID     — Incorrect, not used
  🔲 INCOMPLETE  — AI did not finish the task (cut off, timeout, etc.)
  ⏳ PENDING     — Not yet evaluated
>

Reason (why this status?):
>

Revision (what did you change? Leave blank if VALID/INVALID/INCOMPLETE/PENDING):
>

Log File Structure

See resources/log-format.md for full format spec.
See resources/status-definitions.md for status guidance.
See resources/tag-taxonomy.md for full tag list.

Writing Rules

• All log content must be in English regardless of input language.
• Prompt goes in a fenced ```prompt ``` block.
• Text/code output goes in a fenced ```markdown ``` or language-specific block.
• File output → bullet list: filename — detailed description.
• Entry IDs reset to 001 each month: [AI-AUDIT-001], [AI-AUDIT-002], ...
• Append new entries at the bottom of the file; update stats block at top.
• If the monthly log file does not exist, create it with the stats template first.
• fullname in filename is always normalized (lowercase, no diacritics, spaces → -).

Edge Cases

Situation	Handling
--last > 5	Cap at 5, notify user
Not enough interactions in context	Log what's available, note the gap
Output was a file Claude created	Summarize as bullet list of files + descriptions
Human answers in Vietnamese (or any language)	Translate to English before writing log
Entry ID conflict (file exists, IDs overlap)	Read existing file, continue from last ID
fullname not in context	Ask once before showing the evaluation form
fullname contains diacritics or spaces	Normalize: lowercase, strip accents, spaces → -
--file path directory does not exist	Create intermediate directories as needed
--file conflicts with default path	--file always takes precedence

Quality Checklist

See resources/quality-checklist.md before finalizing.

Example Session

See examples/example-session.md for a full walkthrough.
See examples/sample-tan-phat-2026-07.log.md for a populated log file.