kaggle

What this skill does

# Kaggle — Unified Skill

Complete Kaggle integration for any LLM or agentic coding system (Claude Code,
gemini-cli, Cursor, etc.): account setup, competition reports, dataset/model
downloads, notebook execution, competition submissions, hackathon writeup
retrieval, badge collection, and general Kaggle questions. Five integrated
modules working together.

**Network requirements:** outbound HTTPS to `api.kaggle.com`, `www.kaggle.com`,
and `storage.googleapis.com`.

## Modules

| Module | Purpose |
|--------|---------|
| **registration** | Account creation, API key generation, credential storage |
| **comp-report** | Competition landscape reports (Python API + optional Playwright via host agent) |
| **kllm** | Core Kaggle interaction (kagglehub, CLI, MCP) — includes the `hackathon/` submodule for writeup retrieval and overview/rubric extraction |
| **badge-collector** | Systematic badge earning across 5 phases |

## Credential Setup

**Always run the credential checker first:**

```bash
python3 shared/check_all_credentials.py
```

**Primary credential (recommended):**

| Variable | How to Get | Purpose |
|----------|------------|---------|
| `KAGGLE_API_TOKEN` | "Generate New Token" at kaggle.com/settings | Works with CLI (>= 1.8.0), kagglehub (>= 0.4.1), MCP |

**Legacy credentials (optional, for older tools):**

| Variable | How to Get | Purpose |
|----------|------------|---------|
| `KAGGLE_USERNAME` | Account creation | Identity (auto-detected from token) |
| `KAGGLE_KEY` | "Create Legacy API Key" at kaggle.com/settings | Legacy key for older CLI/kagglehub versions |

Store your API token in `~/.kaggle/access_token` (recommended) or as an env var.
If any are missing, follow the registration walkthrough:
`Read modules/registration/README.md` for the full step-by-step guide.

**Security:** Never echo, log, or commit actual credential values.

## Module: Registration

Walks users through creating a Kaggle account and generating API credentials
(API token as primary, legacy key as optional). Saves to `~/.kaggle/access_token`
and optionally `.env` and `~/.kaggle/kaggle.json`.

Key commands:
```bash
python3 modules/registration/scripts/check_registration.py
bash modules/registration/scripts/setup_env.sh
```

`Read modules/registration/README.md` for the complete walkthrough.

## Module: Competition Reports

Generates comprehensive landscape reports of recent Kaggle competition activity.
Uses Python API for metadata; SPA-only content (problem statement,
rendered evaluation details, winner writeup links) requires the host
agent to provide Playwright MCP tools — the skill itself does not bundle
them. For most overview content, prefer `list_competition_pages` in the
kllm module (no Playwright required).

6-step workflow:
1. Verify credentials
2. Gather competition list across all categories
3. Get structured details per competition (files, leaderboard, kernels)
4. Scrape problem statements, evaluation metrics, writeups via Playwright
5. Compose markdown report with Methods & Insights analysis
6. Present inline

```bash
python3 modules/comp-report/scripts/list_competitions.py --lookback-days 30 --output json
python3 modules/comp-report/scripts/competition_details.py --slug SLUG
```

`Read modules/comp-report/README.md` for full details including hackathon handling.

## Module: Kaggle Interaction (kllm)

Four methods to interact with kaggle.com:

| Method | Best For |
|--------|----------|
| **kagglehub** | Quick dataset/model download in Python |
| **kaggle-cli** | Full workflow scripting |
| **MCP Server** | AI agent integration |
| **Kaggle UI** | Account setup, verification |

Capability matrix:

| Task | kagglehub | kaggle-cli | MCP | UI |
|------|-----------|------------|-----|-----|
| Download dataset | `dataset_download()` | `datasets download` | Yes | Yes |
| Download model | `model_download()` | `models instances versions download` | Yes | Yes |
| Execute notebook | — | `kernels push/status/output` | Yes | Yes |
| Submit to competition | — | `competitions submit` | Yes | Yes |
| Publish dataset | `dataset_upload()` | `datasets create` | Yes | Yes |
| Publish model | `model_upload()` | `models create` | Yes | Yes |

**Known issues:**
- `dataset_load()` broken in kagglehub v0.4.3 — use `dataset_download()` + `pd.read_csv()`
- `competitions download` has no `--unzip` in CLI >= 1.8
- Competition-linked datasets return 403 — use standalone copies

`Read modules/kllm/README.md` for full details and all task workflows.

### Sub-module: kllm/hackathon

Retrieves hackathon writeups, rules, and judging rubrics from Kaggle's MCP
hackathon endpoints. Lives under kllm because it's a focused MCP-workflow
surface like the rest of kllm. Built around the endpoint order from the
2026-04-22 audit (retested 2026-05-04):

1. `get_hackathon_overview` — rules, eligibility, rubric, prizes
2. `list_hackathon_write_ups` — submission roster (paginated, with track ids)
3. `list_hackathon_tracks` — resolve numeric track ids to titles
4. `get_writeup` — preferred full-body fetch (simpler arg shape than
   `get_hackathon_write_up`)
5. `get_writeup_by_topic` / `get_writeup_by_slug` — fallbacks when id missing
6. `get_resolved_writeup_links` — host/judge-gated link enrichment

```bash
python3 modules/kllm/hackathon/scripts/hackathon_overview.py --competition kaggle-measuring-agi
python3 modules/kllm/hackathon/scripts/list_writeups.py --competition kaggle-measuring-agi
python3 modules/kllm/hackathon/scripts/fetch_writeup.py --writeup-id 123456
```

**Live-server status** (verified 2026-05-04):
- `get_hackathon_write_up` — was broken in the 2026-04-22 audit, **now works**.
- `get_benchmark_leaderboard` — was permission-blocked in 2026-04-22, **now PASS** for ordinary KGAT tokens.
- `get_competition` for classic competitions — **now PASS** (recovered upstream).
- `download_hackathon_write_ups` may return CSV header only in some host contexts.
- `get_resolved_writeup_links` is role-gated; participants get an explicit denial.

`Read modules/kllm/hackathon/README.md` for the full retrieval workflow,
role-specific guidance (host/judge vs. participant), and the bundle shape
returned to the agent.

## Module: Badge Collector

Systematically earns ~38 automatable Kaggle badges across 5 phases:

| Phase | Name | Badges | Time |
|-------|------|--------|------|
| 1 | Instant API | ~16 | 5-10 min |
| 2 | Competition | ~7 | 10-15 min |
| 3 | Pipeline | ~3 | 15-30 min |
| 4 | Browser | ~8 | 5-10 min |
| 5 | Streaks | ~4 | Setup only |

```bash
python3 modules/badge-collector/scripts/orchestrator.py --dry-run
python3 modules/badge-collector/scripts/orchestrator.py --phase 1
python3 modules/badge-collector/scripts/orchestrator.py --status
```

`Read modules/badge-collector/README.md` for full details.

## Orchestration Workflow

This skill is primarily a **reference** — use the modules and scripts as needed
based on the user's request. When explicitly asked to run the **full Kaggle
workflow**, follow these steps:

### Step 1: Check Credentials

```bash
python3 shared/check_all_credentials.py
```

If any credentials are missing, walk through the registration module. **Never
echo or log actual credential values.**

### Step 2: Generate Competition Landscape Report

Run the comp-report workflow: list competitions, get details, scrape with
Playwright, compose report. Output inline.

### Step 3: Summarize Kaggle Interaction Methods

Present a concise summary of the four ways to interact with Kaggle (kagglehub,
kaggle-cli, MCP Server, UI) with the capability matrix from the kllm module.

### Step 4: Present Interactive Menu

Ask the user what they'd like to do next:

- **Earn Kaggle badges** — Run the badge collector (5 phases, ~38 automatable badges)
- **Explore recent competitions** — Dive deeper into specific competitions from the report
- **Enter a Kaggle competition** — Register, download data, build a submission, submit
- **Download a Kaggle dataset** — Search for and download any public dataset