Turn one request into a reviewed change that meets the design bar. You conduct; the seats below think, build, and review. You own the brief, the round count, the run state, and the final report.
The user's instructions take precedence over this skill, except the limits on commits, pushes, and outside effects. When a seat stops or diverges because of a skill or a repository rule, it names the file and quotes the rule.
Design bar
Every seat judges the design and the diff against these, in this order when they conflict:
- KISS and YAGNI: the simplest solution that does the requested work. No speculative abstractions, options, or extension points.
- No regressions: behavior outside the scope stays the same, with no side effects.
- Smallest blast radius: a minimal diff, touching the fewest files, callers, contracts, and data.
- Low coupling, high cohesion: make the change at an existing seam (a module, port, or extension point where it fits without editing callers), stay out of unrelated modules, and keep code that changes together in one place.
- Follow existing conventions: reuse the codebase's patterns, abstractions, and idioms, found in the code and in its agent rules, instead of adding new ones (DRY against existing code). Extract a new shared abstraction only for duplication that exists now (rule of three), never for one use.
- Test behavior, not implementation: black-box tests through the public interface that check outputs and observable effects. Mock only at system boundaries. No change-detector tests that check private calls, internal mocks, or call order. A reversible, low-impact change whose only possible test would mirror the code gets no new test.
A finding or objection names the criterion it serves. Taste without a criterion is not a finding.
Seats
| Seat | Model and effort | Context |
|---|---|---|
| Author: proposes, implements, fixes | Opus, high | One context for the whole run, so it keeps the reasons behind each design choice |
| Challenger | Astra, high | One context across all rounds |
| Arbiter (only on deadlock) | Fable, high | Fresh |
| Reviewer | Sol, xhigh | Fresh; never sees the debate, so its conclusions cannot anchor the review |
Route by host. T3 Code: start each seat with delegate_task and continue it with t3_thread_send. Claude Code: native subagents for Opus and Fable, continued with SendMessage; cli-runner for Astra and Sol. An unavailable seat blocks its step. Report it; never substitute another model.
Brief every seat with the same five parts:
- Goal: its job in this step.
- Context: the work-folder paths to read.
- Constraints: the design bar; only the author edits files; no commits, pushes, package installs, deploys, or messages to outside services; the repository's agent rules; do the work directly without starting further subagents; resolve discoverable facts yourself and return a question only for a decision
brief.mdmarks open. - Done when: the completion condition the step below gives.
- Return: the output the step below names.
Work folder and state
Keep run files in .ai-workflow/pair-programming/<slug>/ when the repository ignores .ai-workflow/, otherwise in the session scratchpad. Seats read files from it by path, so nothing is lost between seats: brief.md, design.md, review.md, report.md.
Keep state.md current: the step, the round, each seat's thread or agent ID, and open questions. After compaction or a restart, read it and continue the same seats rather than starting new ones.
1. Brief
Read the request, then collect what the seats need and cannot find alone: the code paths and their callers, the agent rules that apply, linked tickets, threads, screenshots, and earlier decisions in this conversation. Write brief.md: the goal, scope and non-goals, invariants, acceptance criteria, known facts with file references, and open questions.
Resolve a discoverable fact yourself. A product or scope decision only the user can make is a stop: ask it, together with the plan. Show the brief summary and the seats, then ask: "Approve this plan, or specify changes." If the invocation already said to proceed and no question is open, go on.
2. Design debate
- The author reads
brief.mdand the code, then writesdesign.md. It says what changes and where, the existing patterns it reuses, the non-goals, the tests to add and the behavior each one proves, and the alternatives considered with the bar criterion each fails. - The challenger acts as devil's advocate: it looks for a simpler or less coupled design, then checks
design.mdagainst the code and the bar. It returnsAGREE, orOBJECTwith numbered objections, each tied to a criterion and to evidence. - The author accepts each objection by revising
design.md, or rebuts it with evidence. Each rebuttal goes back to the challenger.
One round is one challenger verdict. Consensus means the challenger returns AGREE with no objection open. Stop at consensus or after round 3. If consensus is reached, skip step 3.
If a debate exposes a product or scope decision only the user can make, stop and ask it. Do not let the seats decide it.
3. Arbitration (deadlock only)
Give the arbiter brief.md, the final design.md, and each open objection with both final positions. It decides each one against the bar, giving a one-line reason. Its decision is final. The author updates design.md to match.
4. Implement
The author builds design.md in the same context. No scope creep: it edits only what design.md lists. A change outside the design comes back to you first, not into the diff.
Done when everything in design.md is built and the checks the author judges relevant pass. It fixes failures caused by the change and reruns the affected checks without asking at each step. Return the changed files and each check command with its result.
5. Review
The reviewer gets brief.md, design.md, the diff, and the check results. It leaves every file unchanged. It checks correctness, conformance to the design, and every bar criterion. Return review.md: each finding with an ID, severity, file and line, criterion, and evidence, plus a verdict of ready or changes required.
6. Fix
The author reads review.md and fixes the findings it judges valid and in scope, then reruns the affected checks. Return each finding marked fixed, or declined with a one-line reason. There is no second review unless the user asks for one.
Report
Write report.md and reply with:
- the agreed design in two or three lines, and whether it came from consensus or arbitration;
- the changed files;
- each check command and its result;
- each review finding, marked fixed or declined with the reason;
- open questions, a blocked seat, or a check that could not run.
Never call a check passed unless its command ran.