Specify.Generate.Prove.

Create a global Claude Code skill for the Oodle AI-Native Engineering Kickstart.

Connect the remote MCP server if needed:
claude mcp add --transport http oodle-qa https://oodle.waimakersmcp.com/mcp

The server exposes exactly two tools:
- ask_question(text, name): collect my real name before the first call; never guess it. Reuse it during this workshop.
- list_questions(): read the public questions and presenter answers.

It also publishes Oodle workshop resources under oodle://workshop/. Consult those resources when I ask about participant preparation, Day 1 labs, the Day 2 hackathon, requirements, evaluation, proof packets or risk classification. Jira and Confluence are a separate approved MCP connection; this Q&A server cannot access them, GitHub, Oodle repositories or my credentials.

Write the skill to ~/.claude/skills/oodle-qa/SKILL.md with concise YAML frontmatter and instructions to use these tools whenever I ask the workshop presenter a question or check for answers. Remind me never to submit customer PII, confidential lending information, credentials or secrets. After writing it, show one example invocation.

Using the approved Jira/Confluence MCP connection, work read-only first. Retrieve the pre-vetted issue and approved sources, preserve every source link, state what you could not access, and do not change or create tickets until I explicitly approve the proposed write.

Interview me relentlessly about this Oodle problem, one decision at a time. Separate facts you can retrieve from decisions only I can make; retrieve the facts yourself from approved sources. For each decision, recommend an answer and explain the trade-off. Do not generate the solution until WHO, WHAT, WHY, NOT and DONE are explicit.

Create a global Claude Code Skill named grill-with-docs. It must interview one decision at a time, retrieve codebase facts instead of asking me, update CONTEXT.md only with agreed domain language, and offer an ADR only for a hard-to-reverse, surprising trade-off. It must not implement until I confirm the shared understanding.

Using only the approved sources, draft a requirement contract with WHO, WHAT, WHY, NOT and DONE. Add source links, assumptions, open questions, concrete examples, dependencies and one red line that the solution must never cross. Mark every unsupported claim; do not invent Oodle policy or customer facts.

Convert the frozen contract into a proposed Jira handoff: outcome, acceptance criteria, constraints, dependencies, source links and unresolved questions. Show a preview before any write. Do not change the contract; return discrepancies to the mixed team and wait for human approval before creating or updating Jira.

In Cowork, use the frozen contract and approved sources to create the agreed business artefact: requirement, decision paper, risk assessment, analysis plan or evaluation rubric. Preserve source links, separate evidence from inference, list unresolved questions, and include the named human review gate.

Explore the approved repository slice without changing anything. Trace the relevant business rule from entry point to tests, identify likely files, existing behaviour, discrepancies with the frozen contract, permissions, risks and the smallest testable vertical slice. Cite file paths and evidence, then wait for approval.

Implement only the approved vertical slice. Treat the frozen contract as immutable: if code or evidence conflicts, stop and return the discrepancy to the mixed team. For code, work test-first in thin red-green-refactor slices. Capture inputs, outputs, commands, test evidence and limitations for the proof packet. Never deploy or change production without explicit human approval.

Evaluate this result in four layers: deterministic checks, source-grounded checks, AI-assisted critique and human-only judgment. For every layer show the check, evidence, result and limitation. Regulatory judgment, customer impact, security risk and production decisions remain human-only.

Review the output against the frozen contract and approved sources. Look for correctness gaps, unsupported claims, security issues, customer impact, Consumer Duty or regulatory concerns, permission overreach and hidden assumptions. Classify each finding as evidence-backed risk, false alarm or human decision, and name the accountable reviewer.

Create the proof packet: frozen requirement, source links, generated artefact or code slice, deterministic evidence, source-grounded evidence, AI-assisted review, human-only decisions, one failure or known limitation, green/amber/red risk classification, named owner and a recommendation to continue, reshape or stop.

Create a compact handoff for the next person or session. Preserve the frozen contract, decisions and reasons, source links, current state, evidence, commands already run, known failures, risks, unresolved questions and the exact next action. Omit conversational history that does not change the work.

Extract the smallest reusable learning from this work. Recommend whether it belongs in a Skill, Jira flow, requirement template, automated test, playbook, CLAUDE.md or AGENTS.md. Draft only that change, explain its trigger and guardrails, name an owner, and link it to the evidence that justified it.