Promptfolio Summarize

promptfolio-summarize

You are building a portrait of the human user — the person running this command is the subject being analyzed. Their portrait will be displayed on the promptfolio platform. You are not summarizing projects, not evaluating AI usage skills, not writing a performance review. You are finding the moments where this person teaches AI how to think — their framework sentences — and using those to paint a picture of who they are.

To do this, analyze their AI conversation history (Claude Code, Cursor, Codex, ChatGPT, Gemini CLI, Trae, OpenCode, Antigravity, Windsurf, OpenClaw and any other coding agents found on the system) from the last 30 days, extract an activity heat map and framework sentences that reveal this person's thinking, then build a portrait around those sentences.

Fundamental principle: find where the user is TEACHING, not where they are COMMANDING. "Fix this bug" tells you nothing. "Don't think about it that way — this isn't a performance problem, it's a user psychology problem" / "你不要这样想，这不是性能问题，是用户心理问题" tells you everything.

Step 0: Auto-Update

Before anything else, run the auto-updater to ensure you have the latest skill files and data formats:

bash ~/.promptfolio/update-check.sh

• If output is UPDATED v... → tell the user: "Skills updated to v{version}." Then re-read this SKILL.md file since it may have changed, and continue from Step 1.
• If output is UP_TO_DATE v... → continue silently.
• If output is OFFLINE v... → tell the user: "Could not check for updates (offline). Running with local v{version}." Continue normally.

Then continue with Step 1 normally.

Step 1: Authentication

Ensure ~/.promptfolio/config.json exists and has a valid token.

1a. Validate existing token (if config exists):

if [ -f ~/.promptfolio/config.json ]; then
  API_TOKEN=$(python3 -c "import json; print(json.load(open('$HOME/.promptfolio/config.json'))['api_token'])")
  API_URL=$(python3 -c "import json; print(json.load(open('$HOME/.promptfolio/config.json')).get('api_url','https://promptfolio.club'))")
  HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" -H "Authorization: Bearer $API_TOKEN" "$API_URL/api/profile/me" || true)
  echo "HTTP_CODE=$HTTP_CODE"
fi

• If HTTP 200, proceed to Step 2.
• If config missing or HTTP is not 200, run the device auth script:

1b. Run device authorization:

bash "SKILL_DIR/../scripts/device-auth.sh"

Replace SKILL_DIR with the directory containing this SKILL.md file. The script handles everything: requesting a device code, opening the browser, polling until authorized, and saving ~/.promptfolio/config.json.

IMPORTANT: You MUST run the script as-is. Do NOT reimplement the auth flow yourself. Do NOT construct auth URLs manually — let the script handle everything.

If the script fails, tell the user authorization timed out and to try again.

Step 2: Discover Sessions

Before running any commands, tell the user:

All conversation analysis happens locally on your machine. No raw conversation content is sent to any server during analysis. Only structured results are uploaded after your explicit confirmation.

2a. Detect installed tools

Before scanning, check which AI coding tools the user actually has installed. Look for their config/data directories:

Run a quick check:

echo "=== Detected AI tools ==="
[ -d ~/.claude/projects ] && echo "claude-code"
[ -d ~/.cursor/projects ] && echo "cursor"
[ -d ~/.codex ] && echo "codex"
[ -d ~/.openclaw ] && echo "openclaw"
[ -d ~/.gemini/tmp ] && echo "gemini-cli"
[ -d "$HOME/Library/Application Support/Antigravity" ] || [ -d ~/.gemini/antigravity ] && echo "antigravity"
[ -d ~/.codeium/windsurf ] || [ -d ~/.windsurf ] || [ -d "$HOME/Library/Application Support/Windsurf" ] && echo "windsurf"
[ -d ~/Desktop/chatgpt_history ] && echo "chatgpt"
[ -d "$HOME/Library/Application Support/Trae CN/ModularData/ai-agent" ] || [ -d "$HOME/Library/Application Support/Trae/ModularData/ai-agent" ] && echo "trae"
[ -f "$HOME/.local/share/opencode/opencode.db" ] && echo "opencode"

Then scan for other coding agents not in the list above. Look for dot-directories in ~/ and app directories in ~/Library/Application Support/ that look like AI coding tools (e.g. Kiro, Aider, Continue, Copilot Chat, Trae, Roo Code, etc.):

echo "=== Checking for other coding agents ==="
ls -d ~/.kiro ~/.aider* ~/.continue ~/.roo* 2>/dev/null || true
ls -d "$HOME/Library/Application Support/Kiro" "$HOME/Library/Application Support/Continue" 2>/dev/null || true

If you find any unknown tool directories, peek inside to see if they contain conversation logs (.jsonl, .json, .txt). If they do, include them — no need to ask the user.

Present the results: "Detected: Claude Code, Cursor. Also found: Kiro (has session logs)."

Set PF_SOURCES to the known tools for the script, and handle unknown tools manually in 2b.

2b. Find session files

Run the discovery script, targeting only the detected tools:

export PF_SOURCES="claude-code,cursor"  # only the detected ones from 2a
SESSION_LIST=$(bash "SKILL_DIR/scripts/discover-sessions.sh")
export SESSION_LIST

The script scans known locations for each tool and filters to the last 30 days. After running it:

1. Check zero-session tools — if a detected tool has 0 sessions, explore its data directory yourself to find files the script missed.
2. Handle unknown tools — for any extra coding agents discovered in 2a, explore their data directories, find conversation log files (.jsonl, .json, .txt), and append them to $SESSION_LIST.
3. Ask the user if they have session data in non-standard locations or from tools you didn't detect.

If you find additional session files, append them to $SESSION_LIST.

2c. Compute statistics + activity data

Run the stats script to compute token estimates and extract activity heat map data in a single pass:

python3 "SKILL_DIR/scripts/compute-stats.py"

This reads $SESSION_LIST and produces:

• stdout: session/token summary per source
• _pf_parts/activity.json: per-day activity data for the heat map visualization

2d. Present summary and ask user

Present a summary to the user:

• Total sessions found (last 30 days), grouped by source
• Activity heat map highlights: most active day, latest night, longest session
• Estimated total tokens (labeled as "estimated")

Then use AskQuestion tool to ask:

• Question: "Proceed with analysis?"
• Options:
  1. "Yes, analyze these sessions" — proceed to Step 3
  2. "No, let me adjust" — ask the user what to change

Step 3: Analyze Conversations

Parsing different formats

Claude Code sessions (.jsonl in ~/.claude/projects/):

• Each line is a JSON object from the conversation transcript
• Read and analyze directly

Cursor sessions — Plain text (.txt in ~/.cursor/projects/*/agent-transcripts/):

• Alternating user: and assistant: blocks
• User messages are wrapped in <user_query>...</user_query> tags
• Assistant messages may include [Thinking] sections
• Parse by splitting on user: / assistant: markers

Cursor sessions — JSONL (.jsonl in ~/.cursor/projects/*/agent-transcripts/{UUID}/):

• Each line: {"role":"user|assistant","message":{"content":[{"type":"text","text":"..."}]}}
• User messages contain <user_query> tags in the text field
• Parse JSON and extract the text content

Codex prompt history (~/.codex/history.jsonl):

• Each line typically looks like: {"session_id":"...","ts":...,"text":"..."}
• Treat text as user message content
• Group by session_id to reconstruct threads

Codex session events (~/.codex/sessions/**/*.jsonl):

• Event stream with type + payload
• Extract user-originated content where role/type indicates user messages
• Use history.jsonl as the primary fallback if event parsing is ambiguous

OpenClaw sessions (~/.openclaw/sessions/*.jsonl and ~/.openclaw/agents/**/*.jsonl):

• OpenClaw stores session metadata in ~/.openclaw/sessions.json and per-session transcripts as JSONL.
• Parse each JSONL line as an event; extract entries where the speaker/role/source indicates the user.
• If multiple event shapes exist, prefer explicit role == "user" content, then fallback to user-tagged event payload text.
• Use sessions.json to map session IDs to project/workspace context when available.

ChatGPT export (conversations*.json):

• File is an array of conversations, each containing a mapping tree
• Extract nodes where message.author.role == "user"
• Get user text from message.content.parts[]
• Sort by create_time when reconstructing chronology

Gemini CLI sessions (.json in ~/.gemini/tmp/*/chats/):

• Each file is a full session: {"sessionId":"...","messages":[{"type":"user|model","content":[{"text":"..."}],"timestamp":"..."}]}
• type: "user" = user message, type: "model" = assistant response
• Content is an array of parts, each with a text field
• Timestamps are ISO 8601 strings
• Project name is the parent directory under tmp/ (e.g. ~/.gemini/tmp/myproject/chats/)

Trae sessions (VS Code fork by ByteDance — encrypted DB, requires user-assisted export):

• Trae stores AI chat in an encrypted SQLite DB (SQLCipher) at ~/Library/Application Support/Trae CN/ModularData/ai-agent/database.db — cannot be read directly
• Home directory: ~/.trae-cn/ (CN) or ~/.trae/ (international)
• The discover script checks for exported chat files at ~/.trae-cn/chat-export.json, ~/.trae/chat-export.json, ~/Desktop/trae-chat-export.json, ~/Downloads/trae-chat-export.json
• If Trae is detected but no export files are found (stderr contains TRAE_DETECTED_NO_EXPORT), ask the user with AskUserQuestion:
  • Explain: "Trae stores chat history in an encrypted database. To include Trae data, you can export it from within Trae."
  • Option 1: "Export from Trae" — Tell the user to open Trae, start a new chat, and send this message:

    Export all my chat history as a JSON file. Save it to ~/.trae-cn/chat-export.json (or ~/.trae/chat-export.json for international version).
    The JSON format should be: {"messages": [{"role": "user", "content": "...", "timestamp": "2026-01-01T12:00:00Z"}, {"role": "assistant", "content": "...", "timestamp": "..."}]}
    If there are multiple sessions, combine all messages from all sessions into one flat messages array, ordered by timestamp.

    Then re-run the discover script after the user confirms export is done.
  • Option 2: "Skip Trae" — Continue without Trae data.
• Do NOT attempt to decrypt or brute-force the database
• The exported JSON file, once created, will be picked up automatically on future runs

OpenCode sessions (SQLite → extracted JSON):

• The discover script extracts sessions from ~/.local/share/opencode/opencode.db into temp JSON files
• Each extracted file: {"messages":[{"role":"user|assistant","content":"...","timestamp":1234567890}]}
• Timestamps are Unix epoch integers (seconds)
• If the extracted JSON files are missing or empty, fall back to querying the DB directly:

  sqlite3 ~/.local/share/opencode/opencode.db "SELECT data FROM message WHERE session_id='...' ORDER BY time_created"

Antigravity / Windsurf (experimental sources):

• Prefer explicit transcript files (.jsonl, .json, .txt) when present
• If only state.vscdb / logs are available, treat them as metadata-only context
• If readable conversation content cannot be extracted safely, ask the user for an export path instead of guessing