architect

Files (4)

• SKILL.md
• design-red-flags.md · references/
• rationale-template.md · references/
• runner-prompt.md · references/

SKILL.md (first lines)

# Architect

Sketch types, function signatures, class shapes, and module boundaries with `not implemented` bodies and pseudocode. For a GPU kernel, the sketch is the op contract instead: the op signature with dtypes, the tensor layouts (including any preshuffle), the output buffer and whether the op writes in place, the dispatch gate (which arch, dtype, and shape classes reach which variant, and the fallback), the tuned-config key, and the tile and grid plan as pseudocode. Synthesize across multiple model perspectives, then fill in code against the chosen sketch.

## Start

Open a todolist with one entry per phase before starting.

1. Ground
2. Sketch
3. Agree
4. Implement
5. Scrap

## Phase A: Ground the problem

Build a real mental model of every system the new code touches. Run the **how** skill over the relevant subsystems. Naming a file isn't grounding. Produce the architectural explanation `how` prescribes. If the design redefines ownership or layering, also run the **why** skill on the existing shape so the rationale becomes a constraint, not a guess.

For a kernel, also bound the ceiling first. Find the roofline limit for the target shapes per the **explain-the-number** principle skill, or run a **Bound probe** when the limit is unclear. Then list the levers that the `lever.menu` skill function (`../ateam-mode/references/skill-functions.md`) offers for that bottleneck class.

Skip Phase A only when the work is greenfield with no surrounding system to integrate.

## Phase B: Sketch

Run the **arena** skill with the design-sketch task and the Phase A grounding artifacts. Pass `references/runner-prompt.md` as each runner's prompt.

Take the runners from the `architect runners` line in the `mstack-models.mdc` rule, in place of the `arena runners` line. If the rule or that line is missing, use `claude-opus-5-5-xhigh` and `claude-sonnet-5-5-high`. Alias and rejected entries follow the runner rules in the **arena** skill's Phase A.

Design it twice. Require at least two structurally distinct candidates before synthesis, even when the first looks sufficient. This is the **exhaust-the-design-space** principle skill made concrete. The alternatives must be whole shapes, not point fixes inside one shape. For a kernel, whole shapes differ in data layout, in how work splits into kernels (one fused kernel or two stages), or in how dispatch splits the shape space, not in one tile size.

Screen every candidate against [`references/design-red-flags.md`](references/design-red-flags.md) before synthesis. Assume the next contributor is an agent that sees only the files it opened, copies the nearest example, and takes the shortest path that compiles. Prefer the design where a change that looks right from one file is right for the whole repo.

Compare viable candidates on interface depth. Prefer the design that hides more complexity behind a smaller, simpler public interface. A rich interface can keep call chains short by concentrating capability instead of scattering it across layers.

Arena returns one synthesized design package. Its synthesis decision fills the rationale's "Synthesis decision" section.

## Phase C: Agree (opt-in)

Default: proceed directly to implementation with the synthesized design. No human checkpoint.

Opt in to a checkpoint when the invoker explicitly asks: "/architect with checkpoint," "stop and show me before implementing," or similar. Then show the synthesized design and pause for sign-off.

The synthesis can ship as its own commit either way, as the "scaffold first" mode of the **redesign-from-first-principles** principle skill. Planned and scoped breakage during fill-in is fine, per the **migrate-callers-then-delete-legacy-apis** principle skill. For adversarial pressure on the design before implementing, run the **interrogate** skill on the synthesized sketch.

If the human pushes back on the shape (in a checkpoint or after the fact), treat that as Phase A evidence. Re-ground and re-run Phase B before writing more code.

## Phase D: Implement against the sketch

Replace `not implemented` bodies with code, pseudocode with logic. The synthesized sketch is the contract. Report each deviation from it instead of absorbing it silently. If a function needs a parameter the sketch didn't anticipate, ask whether the sketch was wrong, the requirement was missed, or the implementation is overreaching.

## Phase E: Scrap when the architecture is wrong

If implementation keeps producing friction the sketch can't absorb, throw the sketch out. Don't bolt fixes onto a wrong design, per the **redesign-from-first-principles** and **fix-root-causes** principle skills.

The signal is a *pattern*, not single instances. Tells:

- The same shape of workaround appearing repeatedly across unrelated code.
- Multiple unrelated edge cases that all need special-case branches.
- Types that need escape hatches (`any`, casts, optional fields always set in practice) to compile.
- The "we need a lock" reflex when the sketch said the state wasn't shared.
- Callers having to know the abstraction's internal rules to use it.
- Two or more independent Phase D deviations of the same shape across the implementation.
- In a kernel: per-shape branches inside the kernel body, an arch special case added for each new path, or a tuned-config table that grows a row for every shape instead of a rule.

Use judgment. A few edge cases don't condemn an architecture. Some problems are legitimately complex. Complexity in the data is not complexity in the design.

When you scrap:

1. Re-run the **how** skill over what's been built.
2. Redesign as if the new constraints had been day-one assumptions, per redesign-from-first-principles.
3. Subtract before adding, per the **subtract-before-you-add** principle skill. The new sketch should be smaller than the old one before it grows.
4. Return to Phase B and re-run arena.

## Outputs

Write the caller's usage first and derive the type sketch from it. Use one file with new types and signatures for small changes, and a module map plus type definitions for larger work. For a kernel, write the model's call site first, then the op contract above as a stub with a torch reference body, so the op test can run against the stub before the kernel exists. Ship the rationale alongside, shaped per `references/rationale-template.md`, including the usage sketch and the synthesis decision.

From anhminhnguyenhoang/mstack, MIT.

More from anhminhnguyenhoang/mstack