kaggle-writeup-skill

writing-kaggle-writeups

A Claude Code / Agent Skills plugin for drafting, structuring, and reviewing Kaggle competition solution write-ups.

The skill is not a style opinion. It was distilled from every public solution write-up linked from a Kaggle leaderboard — 5,774 write-ups across 546 competitions, harvested via the Kaggle API — by measuring what gold-medal posts contain that lower-ranked posts do not, and turning those differences into a section contract an agent can follow.

Install

/plugin marketplace add will-rice/kaggle-writeup-skill
/plugin install writing-kaggle-writeups@kaggle-writeup-skill

Or drop the skill straight into your personal skills directory:

git clone https://github.com/will-rice/kaggle-writeup-skill
cp -r kaggle-writeup-skill/skills/writing-kaggle-writeups ~/.claude/skills/

~/.agents/skills/ works as a cross-runtime location for Codex, Copilot CLI, and Gemini CLI.

Use

Ask for a write-up in your own words — the skill triggers on its own:

I finished 3rd in BirdCLEF+ 2025, here are my notes: … write the solution write-up I should post.

It will pull the competition's facts and the other teams' write-ups (using the nvidia-kaggle skill's fetch scripts when they are available), ask for any score deltas you left out, and produce a draft that follows the contract.

What the skill enforces

• A section contract — the sections real top-10 write-ups use, in the order they actually use them (measured median ordering, not invented).
• Numbers on every claim — CV, public LB, and private LB stated explicitly; a delta attached to each technique, or the technique marked unmeasured.
• No invented scores. The most common failure when an agent drafts from incomplete notes is back-computing a baseline from final − delta and presenting it as a measurement. The skill's results-table row schema blocks this; unmeasured cells read not measured.
• A real "what didn't work" section — with reasons, so a reader can act on it.
• Credit and artifact links — @-handles for the public kernels and discussions you built on, absolute links to code, notebook, dataset, model.

It also knows Kaggle's own pre-filled write-up template (Context / Overview of the Approach / Details of the submission / Sources) and when to use it instead — prize-eligible winners' documentation is expected to follow it.

What's inside

skills/writing-kaggle-writeups/
  SKILL.md                        the contract, the rules, the checklist
  references/corpus-evidence.md   the measurements, by finishing position
  references/exemplars.md         worked examples from real write-ups, with links

corpus-evidence.md is the receipts: how median length, section presence, diagram use, and code-linking vary from rank 51+ up to rank 1–3. Every recommendation in SKILL.md traces to a row in it.

Method

1. Enumerate all Kaggle competitions via the internal competition service.
2. For each competition, read the leaderboard and collect every team's linked solution write-up.
3. Fetch each write-up's source and convert it to markdown.
4. Measure structure — headings, section topics and their ordering, length, tables, images, links, quantified deltas — and cut those measurements by finishing position and by community votes.
5. Baseline-test an agent writing a write-up without guidance, write the skill against the failures observed, then re-test until the failures stop.

The harvest and analysis scripts are not shipped here; the corpus is other people's work, and references/exemplars.md links to the originals.

License

MIT