# Day 2 hackathon brief

## Objective

Produce a credible proof on an Oodle problem and leave behind a reusable improvement. A polished demo without evidence does not pass.

## Teams

- Mixed groups with an even technical/non-technical split — at least one engineer per team, two where possible.
- Every team names a **captain** before Wave 1: the captain breaks ties, cuts scope at checkpoints and owns the demo clock. The captain can hold any of the five role lenses.
- The five role lenses (product, engineering, domain, risk, proof) are worn by everyone but owned by someone.

## Mock data is the fast path

An MVP on synthetic data is cheap to build and green-zone by definition. Mock the data, prove the flow end to end, and treat swapping in real inputs as a follow-up ticket with an owner — not as today's risk. Teams should not lose Wave 2 time waiting on access or real datasets.

## Borrow skills — don't build everything

Ready-made Agent Skills cover most scaffolding (documents, spreadsheets, PDFs, frontend polish, mock data, MCP building). Browse before you build:

- **[anthropics/skills](https://github.com/anthropics/skills)** — Anthropic's official skills repo (installable as a plugin marketplace: `/plugin install document-skills@anthropic-agent-skills`)
- **[anthropics/claude-plugins-official](https://github.com/anthropics/claude-plugins-official)** — official plugin directory, includes `frontend-design` for demo-quality UI
- **[skillsmp.com](https://skillsmp.com/)** — searchable community skills directory
- **[awesome-claude-skills](https://github.com/travisvn/awesome-claude-skills)** — curated community list

Rule: **read a skill's SKILL.md before installing it** — a skill is instructions your agent will follow, so it gets the same source-scrutiny as any dependency. Official and curated sources first; anything unvetted is amber.

## Candidate challenge families

- Retrospective code-versus-business-logic audit
- Cross-language code comprehension
- Oodle coding-convention guidance or skill
- Business-context retrieval through approved Atlassian/Jira MCP
- Structured-data assistant with a fixed evaluation set

## Three waves

### Wave 1 — Specify (45 min)

Freeze WHO, WHAT, WHY, NOT and DONE. Include an authoritative source, happy path, edge case, explicit exclusion, proof plan and red line. The facilitator must approve the contract before generation.

### Wave 2 — Generate (75 min)

Cowork produces the business artefact and Claude Code produces the repository slice. Build only the frozen contract. At the 35-minute checkpoint show a working path or reduce scope. Capture commands, sources and representative input/output as you work.

### Wave 3 — Comprehend & Prove (45 min)

Stop adding features. Run deterministic, source-grounded, AI-assisted and human-only checks as applicable. Show one known failure or recovery path, then package a reusable rule, test, skill, Jira workflow or playbook with an owner.

## Demo format — five minutes

1. User and before-state — 30 seconds
2. Live proof — 90 seconds
3. Evaluation results and failure — 60 seconds
4. Risk boundary and human control — 45 seconds
5. Reusable asset, owner and next decision — 75 seconds

## Decision

The panel chooses: continue, reshape or stop. Novelty is not a criterion. Value, evidence, safety, technical quality and reuse are.