Competition Onboarding Skill

Mechanical procedure for CLAUDE.md's "Competition Onboarding Phase (Do this before writing any training code)". Also supports Simulation competitions (submit an agent that plays matches — Lux AI / ConnectX / Halite etc.). For simulation there is no fixed train/test, so the task-definition, metric, and submission-format items each get a simulation note (below).

Steps

1. Identify the input:

   • Use $ARGUMENTS if it contains a competition URL
   • Otherwise infer from the "Competition name" line in KAGGLE_DIRECTION.md
   • WebFetch the overview / data / rules / timeline pages in parallel

2. Fill in the 7 items (ask the user for unknowns; mark inferences as "Inference:"):

   1. Platform: Kaggle / grand-challenge.org / CodaBench / custom
   2. Task definition: inputs, outputs, number of classes, evaluation unit (per-image / per-pixel / per-patient)
      • For simulation: not a supervised task but "play against opponents in an environment (game) and win". Record the observation space, action space, episode/turn structure, and number of players
   3. Data location: download URLs, size, format, license, local target path
   4. Metric: exact definition (average / threshold / background class treatment / per-class vs micro/macro)
      • For simulation: not a fixed metric but a rating (TrueSkill / ELO-like). Locally, use "win-rate against a fixed opponent pool" as the proxy metric
   5. Submission format: CSV / prediction-file zip / Docker container / Simulation agent
      • For simulation: a self-contained script implementing def agent(observation, configuration). Record the per-move timeout, memory limit, and agent I/F
   6. Timeline: validation phase / test phase / final deadline, submission quota
   7. Rules: team size, external data, pretrained models, commercial license

3. Save:

   • Create survey/competition/overview.md (merge with existing, never overwrite)
   • Use this template:

# Competition Overview — <name>

Last updated: YYYY-MM-DD

## 1. Platform
- ...

## 2. Task Definition
- Input: ...
- Output: ...
- Classes: ...
- Evaluation unit: ...

## 3. Data Location
- URL: ...
- Size: ...
- Format: ...
- License: ...
- Local path: `datasets/...`

## 4. Metric
- Name: ...
- Definition: ...
- Implementation notes: ...

## 5. Submission Format
- Type: (A) CSV / (B) Prediction-file zip / (C) Docker / (D) Simulation agent
- Filename / naming rules: ...
- Row count / file count: ...
- Value range / dtype: ...

## 6. Timeline
- Start: ...
- Deadline: ...
- Submission quota: ...

## 7. Rules
- Team size: ...
- External data: ...
- Pretrained models: ...
- License constraints: ...

<!-- Fill this in only for Simulation competitions (agent-submission type). Otherwise delete. -->
## 8. Simulation extras (only if applicable)
- Engine / environment: ... (e.g. kaggle_environments connectx)
- Observation space: ...
- Action space / legal vs illegal moves: ...
- Episode / turn structure: ... (turn count / simultaneous vs alternating / number of players)
- Rating system: ... (TrueSkill / ELO-like, matches per day)
- Agent I/F: return format of `def agent(observation, configuration)` / per-move timeout / memory limit / file-size cap
- Local proxy metric: win-rate against a fixed opponent pool (see `reference_sim/`)

## Unknowns / Awaiting Confirmation
- [ ] ...

4. Check:

   • Confirm all 7 items are filled
   • List any unknowns under "Unknowns"
   • Tell the user not to start implementation until all 7 are filled

5. Suggest next actions:

   • Metric implementation plan (which sklearn function? custom?)
   • Fold design plan (candidate grouping keys)
   • One safe + one bold approach to consider