Welcome to Project Turquoise

End-to-end flow

From blank agents to a Drive submission

Step 1 — Set up two blank-slate agents

Each session runs the agent with a throwaway config home — your own setup is untouched, and the agent starts with no MCP servers, no memory, no custom instructions, and a fresh working directory.

Option A Codex

Paste this whole block into your terminal (it launches at the end):

blank_codex() {
  local root codex_bin

  codex_bin="$(command -v codex 2>/dev/null)"
  if [ -z "$codex_bin" ] && [ -x "/Applications/ChatGPT.app/Contents/Resources/codex" ]; then
    codex_bin="/Applications/ChatGPT.app/Contents/Resources/codex"
  fi

  if [ -z "$codex_bin" ]; then
    echo "Codex CLI was not found. Install it or add it to PATH."
    return 1
  fi

  root="$(mktemp -d "${TMPDIR:-/tmp}/codex-blank.XXXXXX")" || return 1
  mkdir -p "$root/home" "$root/work"
  git -C "$root/work" init -q

  CODEX_HOME="$root/home" "$codex_bin" \
    -C "$root/work" \
    -m gpt-5.6-sol \
    -c 'model_reasoning_effort="high"'
}

blank_codex

Option B Claude Code

Recommended: paste this into any coding agent (Claude Code, Codex, Cursor, etc.) and let it launch a blank-slate Claude session for you.

Set up a Project Turquoise blank-slate Claude Code session on this Mac.

Requirements:
1. Create two temp directories (config + work), e.g. with mktemp -d.
2. Write settings.json into the config dir with exactly:
   {"model": "claude-fable-5", "effort": "high"}
3. Launch Claude Code from the empty work dir using:
   CLAUDE_CONFIG_DIR="" claude --strict-mcp-config --mcp-config '{"mcpServers":{}}'
4. Do not add MCP servers, plugins, memory, or custom instructions.
5. Print the config dir, work dir, and the exact launch command used.
6. If `claude` is not on PATH, locate it or tell me how to install Claude Code.
7. Remind me to sign in with Claude Max if prompted, and to confirm the UI shows Fable 5 at high effort before I send my first task prompt.

Run the commands for me (or print them clearly if you cannot launch interactively).

cfg="$(mktemp -d)"; work="$(mktemp -d)"
echo '{"model": "claude-fable-5", "effort": "high"}' > "$cfg/settings.json"
cd "$work" && CLAUDE_CONFIG_DIR="$cfg" claude --strict-mcp-config --mcp-config '{"mcpServers":{}}'

Before you begin — confirm setup

Check every box. If any of them doesn’t hold yet, fix your setup first.

• I have read the instructions above in full.
• Both agents are blank-slate: fresh installs with no MCP servers, custom tools, or special configurations.
• Codex is running GPT-5.6-Sol and Claude Code is running Fable 5, both at maximum / high reasoning.

Step 2 — Run the same task in both agents

The exchange protocol

1. Give the exact same initial prompt to both agents, word for word.
2. Work the problem back and forth with each agent as you naturally would.
3. Stop when solved, or after 15 back-and-forth exchanges — whichever comes first.

When you finish each session

Send the following prompt to the model:

Please save a faithful transcript of this conversation as a .md file titled "claude/codex-transcript-'first four characters of task id'.md". Do not include any reference to the model or harness used.

Once the file is created, ask Claude Code / Codex to save the file to a permanent folder on your device (for example Desktop or Documents).

What to write up for the submission form

Save transcripts (two useful paths)

Project Turquoise wants a faithful Markdown transcript. tourqmd can also rebuild one from the raw session log.

Path A — Agent-written transcript (protocol default)

1. Use the “save a faithful transcript…” prompt at the end of each session.
2. Ask the agent to move the .md to a permanent folder.
3. Keep both files clearly named, e.g. codex-transcript-ab12.md and claude-transcript-ab12.md.

Option / Path B Raw session logs → tourqmd

If you need a mechanical export of everything that happened (or the agent transcript is incomplete), recover the raw .jsonl and convert it in tourqmd.

Recommended: paste this into a coding agent and let it locate and stage the files for you.

Help me recover Project Turquoise session logs for Codex and/or Claude Code on this Mac and stage them for tourqmd.

Do the following:
1. Find the newest relevant session files:
   - Codex: ~/.codex/sessions/**/rollout-*.jsonl (and any CODEX_HOME temp homes I still have open)
   - Claude Code: ~/.claude/projects/**/*.jsonl (and any CLAUDE_CONFIG_DIR temp configs)
2. List the newest few with full paths, sizes, and modified times.
3. Copy the ones I choose (or the newest for each tool) to a permanent folder on my Desktop, e.g.:
   ~/Desktop/turquoise-sessions/
   Name them clearly, e.g. codex-<taskid>.jsonl and claude-<taskid>.jsonl
4. Remind me to open https://tourqmd.traphic.dev/ , drop each .jsonl, review Reader mode, then Submit with my alias.
5. If blank-slate sessions used temp CODEX_HOME / CLAUDE_CONFIG_DIR, search /tmp and $TMPDIR for recent codex-blank.* / claude config dirs and check those too.
6. Do not upload anywhere yourself — only copy files locally and print the final paths.

If a path is missing or empty, explain what to re-run and how to save a transcript before closing the agent next time.

~/.codex/sessions

ls -lt ~/.codex/sessions/*/*/*/*.jsonl | head
open ~/.codex/sessions

~/.claude/projects

open ~/.claude/projects
find ~/.claude/projects -name "*.jsonl" | head

Convert with tourqmd

Clean back-and-forth transcripts for upload and rating

1. Open the converter.
2. Drop a Claude or Codex .jsonl (or use a sample to learn the UI).
3. Review in Reader mode (formatted). Switch to Source for raw Markdown.
4. Copy or Download .md for your records.

Output shape

# Transcript

**User:**
your message

**Assistant:**
agent reply and readable actions

**User:**
…

**Assistant:**
…

Submit to Google Drive

After both transcripts look good

1. Convert each session (Codex + Claude) in tourqmd.
2. Click Submit.
3. Enter your alias only (no email).
4. The file uploads as alias-1__….md, then alias-2__….md, and so on.

Prefer clear names before upload when possible (e.g. include codex / claude in the source filename). Your alias still drives the unique ID.

Task write-up (for the submission form)

Fill these two blocks carefully — raters rely on them

Block 1 — What you set out to do

Include:

• What you set out to do and why
• What area of coding the task sits in
• Required expertise to perform the task
• How long an expert would take to do it themselves

Block 2 — Initial prompt (identical for both agents)

Copy and paste the exact initial prompt you gave to both Codex and Claude Code.

Creating · Vetting · Blind rating

How your work is used in the project

Troubleshooting

Useful links

https://tourqmd.vercel.app/

https://tourqmd.vercel.app/instructions.html

https://tourqmd.traphic.dev/admin.html

You’re ready

Set up blank agents → run the same task twice → save both transcripts → convert in tourqmd if needed → submit with your alias.