# Day 1 labs

Every lab uses synthetic, sanitised or explicitly approved inputs. Never enter secrets, credentials or unapproved personal data.

## Lab 1: Source-grounded Jira/Confluence (45 min)

**Mission:** retrieve one pre-vetted Oodle issue and approved sources; find missing context without inventing it.

**Inputs:** approved Jira issue, Confluence/source pack, Jira/Confluence MCP, source checklist.

**Steps:** retrieve the issue and sources; identify at least two missing pieces; draft user, problem, evidence, assumptions and unresolved questions; preserve source links. A human approves before any write.

**Done when:** another pair can trace every claim and name the unresolved questions.

**Safety boundary:** read-only until facilitator approval; approved spaces only.

**Stretch:** propose acceptance checks, labelling each as sourced or inferred.

## Lab 6: Context & token economics (45 min)

**Mission:** make the token cost of your context visible, then load less and lose nothing.

**Inputs:** Claude Code with the approved Atlassian MCP available; any approved repo or working folder.

**Steps:** start Claude Code with the Atlassian MCP connected; before asking anything, run `/context` and write down the tokens consumed by MCP tools (expect roughly 20–40k from ~30 tools). Start a fresh session with the MCP off (or deselect its tools) and run `/context` again; write down the new number. Ask the same small question in both sessions and compare the answers. Name what you would actually load for tomorrow's hackathon task, and what you would leave out.

**Done when:** you can show a before/after token number and say one sentence about what you'll stop loading.

**Safety boundary:** read-only; approved Jira/Confluence spaces only; no writes.

**Stretch:** apply the same test to your own artefact: how small can a shared `oodle-context` skill stay while still being worth loading? Narrow beats broad: a long context file costs tokens on every question and degrades the answer.

## Lab 2: Role-based Cowork / Claude Code (45 min)

**Mission:** produce a useful artefact in the tool that matches your role.

**Inputs:** source-grounded brief from Lab 1, Cowork, approved repository slice, normal test command.

**Steps:** product/finance/data participants use Cowork to create a requirement, test matrix, analysis plan, decision paper, stakeholder brief or risk assessment; engineers use Claude Code read-only to explain behaviour, likely files and tests. Pair across roles and challenge the output.

**Done when:** the participant can explain the result in their own words, identify an uncertainty and defend every source.

**Safety boundary:** no repository edits; no external sends; no unsupported business rule.

**Stretch:** perform a context relay and record what was lost between artefacts.

## Lab 3: Grill the requirement (45 min)

**Mission:** freeze WHO, WHAT, WHY, NOT and DONE before generation.

**Inputs:** `templates/requirement.md`, approved sources, risk traffic light.

**Steps:** answer one question at a time; distinguish required behaviour from implementation; add a happy path, edge case and explicit exclusion in Given/When/Then form; name the authoritative verifier.

**Done when:** another team can repeat the contract, proof and red line without asking a question.

**Safety boundary:** unresolved policy or business truth stays an open question, never an agent assumption.

**Stretch:** ask an adversarial reviewer to find one ambiguity, then resolve or record it.

## Lab 4: Evaluation stack (30 min)

**Mission:** define confidence before seeing the final output.

**Inputs:** frozen requirement, `templates/evaluation-rubric.md`, fixed examples or test set.

**Steps:** add each applicable layer: deterministic, source-grounded, AI-assisted and human-only. Human-only includes regulatory judgement, customer impact, security risk and production decisions where relevant.

**Done when:** every acceptance signal has an owner, method and evidence location.

**Safety boundary:** AI may assist review but cannot own regulated, customer-impacting, security or release decisions.

**Stretch:** add one known-failure case and its recovery path.

## Lab 5: Day 2 scope lock (30 min)

**Mission:** prepare a buildable, safe brief and mixed team for Day 2.

**Inputs:** requirement, evaluation rubric, risk traffic light, proof packet and approved systems.

**Steps:** choose the thinnest slice; confirm product, engineering, domain, risk and proof roles; name non-goals; test access; draft the continue/reshape/stop decision rule.

**Done when:** the facilitator approves the draft contract, proof, red line, roles and fallback.

**Safety boundary:** a red use case does not enter the hackathon; amber requires named controls and reviewer.

**Stretch:** identify the reusable skill, rule, test, Jira flow or playbook the team hopes to leave behind.