Contribute a run

Anyone can add a model. A run only counts if the model gets the exact prompt, works on its own, and the file it delivers goes up unedited.

Run the model

1. Start your agent harness in a new, empty folder. Don't run it inside this repo: the prompt tells the model not to look at other files, and other entries must not be within reach.
2. Paste prompts/v1.md exactly as written. Don't add hints, fix typos or change the wording.
3. Let it work for up to 2 hours, or 3 if the model asks for more time. Stop at 3 hours and take whatever file exists.
4. Note the exact model id, the harness and its version, the effort or thinking setting, the time it took, and the tokens and cost if your harness reports them.

Submit it

1. Name the run <model>-<effort>-<yyyymmdd>, for example opus-5-5-high-20261009. Leave out the effort if you don't know it. Add -b, -c for more runs on the same day.
2. Put the delivered file at runs/<id>/index.html. Only move or rename it, never edit it.
3. Get its hash:

   shasum -a 256 runs/<id>/index.html

4. Add an entry to manifest.json following manifest.schema.json. Write null for anything you didn't record. Don't estimate. For a v1 run, set:

   "params": {
     "seed": true,
     "coat": { "red": "red", "sable": "sable", "tri": "tri", "bht": "bht" }
   }

5. Open a pull request with the file and the manifest change.

What happens next

If the file fails to load or throws errors, it stays listed with first_try: false. Failures are results too. A second attempt is a new entry, not a replacement.

Full runner rules are in prompts/RULES.md.