Conversation Logger

Log AI coding agent sessions as timestamped markdown for project documentation.

Agent Detection

Detect the invoking agent to determine where history/timestamps live. Check in order:

Detection command:

if [ "$CLAUDECODE" = "1" ]; then echo "claude-code"
elif [ -d "$HOME/.codex" ]; then echo "codex"
else echo "unknown"; fi

Record the detected agent in the log header as **Agent**: {agent-name}.

Why This Skill Exists

The user's prompts are the primary value. When reviewing past sessions, the user wants to see what they were thinking and asking—their mental journey through a problem. Claude's responses are secondary context.

This creates an asymmetric log:

• User prompts: Captured in full detail (the main record)
• Claude responses: Compressed to 1-2 lines (just enough to know what happened)
• Summaries: Avoided entirely (prompts tell the story)

Think of it like a lab notebook: the scientist's observations and questions are the record; the equipment readings are noted briefly. Don't over-document Claude's side—it's noise that buries the user's thought process.

Quick Start

1. Determine the log folder from context or ask the user
2. Compute today's date and the next stage number {N} for that folder
3. Create a new file named prompt-log-YYYY-MM-DD-{N}-{topic-slug}.md
4. Use the template from assets/init-prompt-log-template.md

File Location and Naming

Filename pattern: prompt-log-YYYY-MM-DD-{N}-{topic-slug}.md

Where:

• prompt-log- — literal type prefix. Always present. Marks the file as a conversation/prompt log so it's distinguishable at a glance from other dated docs that may share the folder (plans, retros, migration records, decision records). It also groups all logs together in a folder listing.
• YYYY-MM-DD — today's date from date "+%Y-%m-%d". Day precision only — no hours/minutes in the filename. ISO 8601 day-level format is the one identifier that's meaningful across every project.
• {N} — zero-based stage number that orders logs within a single day. Always present, even if it's the only log on that date — it keeps folder listings deterministically chronological by pure alphanumeric sort. First log of the day is 0, next is 1, etc.
• {topic-slug} — kebab-case, 3–8 words, descriptive (e.g. init-eval-harness, building-batch-runner, refinement-flaky-tests). The slug may begin with a stage word (init, building, refinement, followup) as a human readability hint, but the numeric {N} is what enforces order.

New day = new file. If the user resumes work on a new calendar day, always start a fresh file with the new date and N=0. Never append to yesterday's log. This is the most important rule — it's what makes the log set readable as a timeline.

Picking {N}: List the target folder for existing prompt-log-YYYY-MM-DD-*.md files matching today's date, take the highest {N} you see, and add 1. If none, start at 0.

TODAY=$(date "+%Y-%m-%d")
ls "$LOG_DIR" 2>/dev/null | grep -E "^prompt-log-${TODAY}-[0-9]+-" | \
  sed -E "s/^prompt-log-${TODAY}-([0-9]+)-.*/\1/" | sort -n | tail -1
# Add 1 to that value (or use 0 if no match)

Examples:

• prompt-log-2026-05-17-0-init-eval-harness.md — first log of the day
• prompt-log-2026-05-17-1-building-batch-runner.md — second log, different phase
• prompt-log-2026-05-17-2-refinement-flaky-tests.md — third log
• prompt-log-2026-05-18-0-followup-batch-runner.md — next day, counter resets

Location priority (if clear from context):

1. ./ai-docs/{project-name}/ - if working on a specific project/agent
2. ./ai-docs/ - if exists
3. ./docs/ - if exists
4. ./ - project root as fallback

If location is not clear: Ask the user where they'd like the log saved before creating it. Example: "Where should I save the prompt log? I see you have ./ai-docs/ — want me to create it there as ./ai-docs/{project}/prompt-log-2026-05-17-0-{slug}.md, or somewhere specific?"

Existing single prompt-log.md files: Leave them alone. The new per-session naming applies to logs going forward; don't migrate or rewrite legacy single-file logs.

Logging Format

Each prompt gets a timestamp at the H3 level:

### Prompt 1: Context Gathering (10:34 AM)

> {Full user prompt - preserve as-is, only trim if extremely long}

→ Response: {1 line what was decided/done, DO NOT GO LONG}
→ Action: {1 line files/tools used, DO NOT GO LONG}

Timestamps — Use Real Data from History File

CRITICAL: Do NOT guess timestamps. Read them from the detected agent's history file.

Claude Code (~/.claude/history.jsonl)

One JSON object per line. Fields: timestamp (epoch ms), display (prompt text), sessionId, project.

tail -100 ~/.claude/history.jsonl | python3 -c "
import sys, json
from datetime import datetime
for line in sys.stdin:
    d = json.loads(line)
    if d.get('project','') == 'PROJECT_PATH_HERE':
        ts = datetime.fromtimestamp(d['timestamp']/1000)
        prompt = d['display'][:100].replace('\n',' ')
        print(f'{ts.strftime(\"%I:%M %p\")}  {prompt}')
"

Replace PROJECT_PATH_HERE with the actual project path. Adjust tail -100 if the session is longer.

OpenAI Codex CLI (~/.codex/history.jsonl)

One JSON object per line. Fields: ts (epoch seconds), session_id, text (prompt text).

tail -100 ~/.codex/history.jsonl | python3 -c "
import sys, json
from datetime import datetime
for line in sys.stdin:
    d = json.loads(line)
    ts = datetime.fromtimestamp(d['ts'])
    prompt = d.get('text','')[:100].replace('\n',' ')
    print(f'{ts.strftime(\"%I:%M %p\")}  {prompt}')
"

Note: history.jsonl is only written in interactive mode. For headless (codex exec) sessions, read timestamps from session transcripts instead:

# Find the latest session transcript for today
LATEST=$(ls -t ~/.codex/sessions/$(date +%Y/%m/%d)/rollout-*.jsonl 2>/dev/null | head -1)
# Each line has ISO timestamp: {"timestamp":"2026-03-28T22:59:25Z","type":"response_item",...}
# User prompts have payload.role == "user" and type == "response_item"

Codex session index: ~/.codex/session_index.jsonl.

GitHub Copilot

Copilot chat history is stored inside VS Code's internal state.vscdb SQLite databases under workspace storage paths. It is not exposed as standalone files. Extracting timestamps is possible but fragile — fall back to date for current time and note timestamps are approximate.

Format

Use 12-hour format with AM/PM: (10:07 AM), (2:15 PM)

Notes

• AskUserQuestion responses (multiple-choice answers) are NOT in Claude Code's history.jsonl — only text prompts are recorded
• Codex history may use different field names across versions — inspect the first line of the file if parsing fails
• If history file is unavailable or empty, fall back to calling date for the current time and note that earlier timestamps are approximate

What to Capture

PRIORITY: User input is the main value. Response/Actions = 1 LINE TOTAL.

User Prompts (PRIMARY FOCUS)

• Capture the full user prompt - this is the most important part
• Preserve file paths, code snippets, and specific instructions verbatim
• Only abbreviate if prompt exceeds ~500 words
• If prompt is from AI agent with generic instructions: Summarize the intent instead of logging full boilerplate (e.g., "Build phase 3 from plan.md" instead of full agent prompt)

Response and Actions (TWO LINES MAX)

CRITICAL: 1 line for response, 1 line for action. NEVER use bullet lists.

→ Response: Fixed blank screen bug in BatchEvaluationForm.
→ Action: Updated 2 files, created test helpers, committed.

BAD (too verbose - NEVER do this):

#### Response and Actions
- **Response**: Implemented Phase 31 using pragmatic stub approach
- **Actions**:
  - Stubbed deterministic.ts
  - Removed playwright dependency
  - Created handoff document

GOOD (correct format):

→ Response: Stubbed deterministic agents, removed Playwright dependency.
→ Action: Updated 4 files, simplified Dockerfile, created handoff. ✅

Rules:

• Use → Response: and → Action: format
• Each line 1-2 sentences max
• No bullet lists, no sub-items
• Include status emoji on action line if applicable (✅ ⚠️ ❌)

Action Shorthand

Session Breaks (new file vs. new header)

The default is one session per file. New-file triggers:

• New day — always start a fresh file with the new date and N=0. Never append to yesterday's log, even if the topic is identical.
• Phase shift mid-day — when the work clearly transitions (e.g. debugging → docs, building → refinement), bump {N} and create a new file. Reflect the shift in the slug (prompt-log-2026-05-17-0-debugging-batch-runner.md → prompt-log-2026-05-17-1-refactoring-batch-runner.md).
• Long gap (>2 hours) — usually warrants a new file with bumped {N}.

Within a single file you may use ## Session N: ... headers if a single session has multiple internal phases that aren't worth their own file. But prefer splitting into files — that's what makes the folder a timeline.

## Session 2: Implementation (May 17, 2026)

Session Summaries

DO NOT add verbose session status sections. The prompt log captures user prompts - that IS the record. No need for:

• ❌ "What Was Built" bullet lists
• ❌ Feature summaries or tables
• ❌ "Key Decisions" sections
• ❌ "Next Steps" lists

If a summary is truly needed (e.g., end of major milestone), keep it to 2-3 lines max:

---

**Session 1 Summary**: Fixed 6 bugs, implemented 4-scenario support. Kanban app Tasks 18-20 complete.

The prompts themselves tell the story. Don't duplicate.