| name | hai-architecture |
| description | Investigates architecture and software change complexity, from whole-repository runtime paths to a bounded module/design decision. Use for 系统太绕、复杂度审计、调用链/配置太散、架构审查、模块边界、拆分合并, or why a repo is hard to change. Select global investigation or bounded design mode; return traced evidence, the complexity center, alternatives, and a first proof. Use hai-debug for an unexplained malfunction and code-review-and-quality for reviewing a change or local code smells.
|
Hai Architecture
For Chinese readers, see SKILL.zh_CN.md. The English SKILL.md is the execution source of truth.
Purpose
Review existing architecture or guide one design decision through the lens of managing complexity.
Find the boundary where change amplification, cognitive load, or unknown unknowns concentrate, then
recommend the smallest structural move that makes the system easier to understand and change.
Select the work mode
- Global investigation: the cause or boundary of change friction is unclear, or the request
spans a repository/large subsystem. Read
references/global-audit.md and follow its entrypoint,
call-chain, state/config, and test tracing procedure. Converge on important paths after mapping
the entrypoint families; do not substitute a local smell review.
- Bounded review/design: a module, package boundary, or design choice is already identified.
Follow the workflow below, inspecting its direct callers, contracts, and tests.
- A request can move from global investigation into a bounded decision. Reuse collected evidence;
do not rerun two full reviews or ask the user to choose an internal mode.
Evidence gate
Before making an architecture claim:
- Read every cited file in this session and cite real
file:line locations.
- Trace callers and implementations before asserting a call chain or blast radius; name the search.
- Read ADRs, dependency rules, architecture tests, and public contracts that declare intended
boundaries.
- Compare a module's public surface with the complexity it actually hides before calling it deep
or shallow.
- Label inference as unverified instead of presenting it as evidence.
A fabricated finding costs more trust than several missed findings. Prefer one verified painful
center over a checklist of plausible smells.
Core judgments
Find the painful center
Start with the boundary or call chain under review. Ask where a common change crosses owners,
requires several layers of knowledge, mixes lifecycle/persistence/execution concerns, or makes it
unclear what else must change. Local naming or style issues matter only when they reveal that deeper
force.
Prefer deep modules
A useful module hides significant, related complexity behind a small, stable interface. Do not
split because a file is long or merge because files are adjacent. Split when the new boundary hides
information; merge when separate pieces share hidden knowledge and must be understood together.
For each layer or new shape, ask:
- Does the caller know less?
- Does this layer own a rule or invariant?
- Does it make illegal states harder to express?
- Can downstream change stop here?
- Does a conversion add meaning, constraints, or ownership?
If every answer is no, the layer is probably a pass-through tax. Do not create a new shape merely
because data crossed a function.
Compare real alternatives
For a nontrivial decision, compare at least two materially different options. Explain why the
current design, a further split, a merge, or a generic abstraction is rejected when each is
plausible. For key recommendations, include a concise red-team misuse, blue-team defense, and
residual risk.
Read references/principles.md when the decision needs the full lens catalog, error-boundary
guidance, complexity vocabulary, or adversarial-review detail.
Workflow
Review existing code
- State the reviewed boundary and the request's scale.
- Pass the evidence gate: read implementation, callers, contracts, tests, and relevant decisions.
- Draw a simple current map of actors/modules, dependency or state flow, and the intended hiding
boundary.
- Locate the highest-leverage complexity center.
- Select only the relevant lenses from
references/principles.md; do not score every lens by
default.
- Apply the deep-module and layer-cost tests.
- Check relevant red flags in
references/red-flags.md; use
references/worked-example.md when calibrating a difficult finding or false positive.
- Write a proportional report with evidence-backed findings ordered by impact.
Guide a design decision
- Clarify the goal, boundary, current constraints, and contracts that must survive.
- Map the current chain before proposing changes.
- Design at least two fundamentally different options.
- Compare business fit, boundary clarity, module depth, migration cost, operational risk, and
rejected alternatives.
- Recommend one option, state residual risk, and name the first proof point.
Scale
- Quick: one bounded question. Return the map, conclusion, evidence, main tradeoff, and next step.
- Standard: one package or call chain. Use the full markdown template.
- Full: multiple boundaries explicitly requested by the user. Map broadly, then converge on the
painful center; increase evidence breadth without multiplying boilerplate.
The evidence gate applies at every scale. Do not make a quick question perform a full audit, and do
not call a repository-wide sweep complete after reading one attractive file.
Output
Match the user's language and keep code identifiers unchanged. For standard/full Markdown reviews,
fill references/output-template.md in bounded mode; use references/global-output-template.md in global mode; for a quick answer, preserve its decision fields without
emitting empty sections. When the user explicitly asks for HTML, also read
references/html-report.md. Language-specific guidance is available in
references/go-patterns.md and references/typescript-patterns.md.
Use a different skill when
- An unexplained malfunction needing reproduction and causal evidence →
hai-debug.
- One React component's API, data flow, effects, or rerenders →
react-component-diagnosis.
- Local naming, duplication, function, or style smells →
code-review-and-quality or hai-naming.
- Product requirements rather than technical design →
hai-prd.
Not this skill
Do not use it as a formatter, feature-completeness audit, generic test advisor, or performance
review unless performance architecture is the explicit decision. Keep the focus on design quality
as it affects complexity.