The Duel skill

Make Claude defend its answer against another model. The Duel, also called Face Off, sends an answer to an OpenAI model for a critique. Claude defends what holds up, revises what does not, and returns the answer with a record of what changed.

Say “Duel this with ChatGPT” or run /duel. Up to three rounds. Two actual models. One clear account of the claims they challenged.

“ChatGPT” is the friendly name here: the critic runs through the OpenAI API, not your ChatGPT app or subscription. Claude Code is the defender and writes the final synthesis. There is no neutral judge and no guarantee the resulting answer is correct.

Paste this into Claude Code:

https://github.com/Jakeschincariol/duel-skill

Install this skill, then confirm /duel is available. Do not run a paid duel yet.

Or install manually:

git clone https://github.com/Jakeschincariol/duel-skill.git
mkdir -p ~/.claude/skills
cp -R duel-skill/skills/duel ~/.claude/skills/duel

For a project-only install, copy skills/duel into that project's .claude/skills/ instead.

Or install as a Claude Code plugin:

/plugin marketplace add Jakeschincariol/duel-skill
/plugin install duel-skill@duel-skill

Plugin installs use /duel-skill:duel. The copied skill uses /duel.

Requires Claude Code and Python 3.9 or newer. The helper uses Python's standard library; there is nothing to pip install.

Configure OPENAI_API_KEY securely in the environment that launches Claude Code. Do this yourself through your preferred secret manager or shell setup; do not paste the key into chat or put it in this repository. No Anthropic API key is needed: the current Claude Code session is the defender.

The default critic is gpt-6-astra. Set DUEL_OPENAI_MODEL to use a different Responses API model available to your account, or pass --model to /duel or the Python helper.

The skill is free under the MIT license. Claude Code usage and OpenAI API usage are billed under their respective plans. The normal run makes up to three critic calls; --quick makes one. Actual cost depends on the model and the amount of text sent and returned.

/duel
/duel --quick
/duel --manual --quick
/duel --rounds 2 Should I raise my coaching price from $500 to $700?
Duel this with ChatGPT.
Face off this answer.

On its own, /duel uses the answer you want challenged and its original task. With a new task, Claude first writes an answer. Three rounds is the default; --quick selects one, and --rounds accepts 1, 2, or 3. A run can end early if the critique adds nothing substantive.

1. Prepare. Claude captures the task, answer, relevant context, and claims in a focused packet.
2. Challenge. The OpenAI critic identifies specific weaknesses, counterexamples, missing evidence, and conditions that could change the answer.
3. Defend and revise. Claude addresses the objections, updates its answer, and carries the debate forward to the next round.
4. Return the result. You get the revised answer and a claim ledger: retained, revised, dropped, or unresolved. Claims keep stable IDs so concessions remain visible.

Claude makes the final synthesis, and disagreement may remain unresolved. Debate is a way to surface weaknesses; it is not independent verification of every fact.

Without an API key, use manual mode: have an external model critique the prepared packet yourself, then paste its actual response back into Claude. The skill must label this manual mode and cannot invent an external model's participation.

The helper sends only the prepared packet and its critic instructions to https://api.openai.com/v1/responses. It does not upload the repository or conversation wholesale. Review the packet before using sensitive material, and include only the context needed for the task.

Requests use store: false. This disables Responses API application-state storage; it does not promise zero retention. See OpenAI's API data controls. The critic receives no tools. Each helper invocation makes one request, with no automatic retries, an input limit of 128 KiB (and 100,000 text characters) and a default output budget of 8,192 tokens. An explicit --max-output-tokens override can raise that budget to 32,768 tokens.

Local run records live under .duel/, which this repository ignores. Keep real run packets and responses out of commits, including when installing the skill into another project.

Run these from the cloned repository:

python3 skills/duel/scripts/duel.py check
mkdir -p .duel
python3 skills/duel/scripts/duel.py critique \
  --packet examples/pricing.packet.json \
  --output .duel/example.request.json --dry-run
python3 -m unittest discover -s tests -v

The example packet is illustrative offline input, not a live duel or a model verdict. A dry run prepares the request without contacting OpenAI. check also runs offline; it exits with status 1 when no API key is configured. To make one paid critic call:

mkdir -p .duel/example
python3 skills/duel/scripts/duel.py critique \
  --packet examples/pricing.packet.json \
  --output .duel/example/critique.json

The helper also accepts --model, --max-output-tokens, and --timeout. It handles one critique; the skill orchestrates the rounds and final answer.

CI runs the offline unit tests on Python 3.9, 3.12, and 3.14, without API credentials.

MIT. Made by Jake Schincariol.