structure-first（结构优先）

Skill: Structure First

Purpose

Primary Flow: the top-down readable main path that orchestrates the logic

When generating, refactoring, or reviewing code, prioritize a readable success path (Primary Flow) first. Keep boundaries minimal (only when needed), compose with Atoms (small units with one stable role and clear I/O), and secure stability through contract-driven tests rather than implementation-following tests.

Use / Do Not Use

• Use this skill when the code shape itself is changing or when you are doing a structure-focused code review.
• For planning, classification, or scope analysis, keep it as an internal lens rather than a response template.
• Use it when code does not read naturally from top to bottom.
• Use it when function/module splitting becomes excessive and utilities start to spread.
• Use it when tests are drifting toward implementation-following patterns.
• Do not use it for throwaway experiments or one-off exploratory code.
• Do not use it for tiny changes where structural intervention would be excessive.

Core Bias

• Readable flow > structural simplicity > reusability > abstraction
• Prefer clarity of the current code over speculative future needs.
• Splitting is not a goal; split only when readability clearly improves.

Operating Model

Pick one current unit and make that unit readable before moving outward. Local changes usually mean function/file. Feature work usually means module/use case. Larger refactors may mean capability/subsystem.

1. Fix Intent

• State the intent of the change/code in one sentence.
• State the current unit before restructuring: function/file, module/use case, or capability/subsystem.

1. Minimize Boundaries

• Classify steps as I/O, domain decision, or transform.
• Do not add more boundaries than necessary.

1. Primary Flow First

• Make the success path readable in one top-down pass.
• Keep branches/exceptions from breaking the main flow (early return or push them downward).
• A good result often still reads like normalize -> load -> decide -> return.
• When flattening local branch logic, preserve whether rules are cumulative or precedence-ordered.

1. Extract Atoms

• Split when a Primary Flow sentence becomes clearer as a function or child unit.
• Atoms do one stable job; their I/O should be describable in one line.
• Prefer pure functions when possible.
• At function/file scale, stop when splitting no longer clarifies the local flow.

1. Keep Composition in One Place

• Keep orchestration in one place at the current unit.
• Minimize direct dependencies/calls between Atoms.

1. Push Side Effects to Boundaries

• Gather side effects (I/O, state mutation) at boundaries.
• Keep inner logic focused on computation and decisions.

1. Align Read Order

• At file level, default to: export/public -> orchestrator -> atoms -> utils for top-down discoverability.
• If branch narration starts depending on extra intermediate data structures or helper chains, keep more inline or stop descending.

Growth and Ownership

• Start with the minimum public I/O/signature for the confirmed responsibility; grow it only when responsibility changes (new external input, mixed semantics, or boundary move).
• One policy, decision, or calculation rule, one owner.
• One externally visible write path, one owner. If coordinating multiple writers is itself the responsibility, make that coordinator explicit.
• At larger scales, make the entrypoint, main orchestrator, and decision owners visible.
• Async boundaries should have one owner for freshness and completion policy.
• If rule ownership changes or you introduce an equivalent new path, remove/disable the old one in the same change when possible. Otherwise include a staged migration plan (owner, exit condition).
• Decision rule: repeated predicate, weight, priority, policy, calculation, or key-generation logic that decides behavior.
• Equivalent path: an alternative execution path that yields the same externally observable result.

Testing

• Write sufficient tests at the most stable Atom level available whenever possible.
• Validate contracts (I/O, invariants, edge cases) between the current unit and its Atoms/boundaries, not internals.
• If orchestration or boundary integration is where the risk lives, test the current unit directly.
• For async or stateful boundaries, test ownership contracts such as stale-result handling, balanced completion, and equivalent-input no-op at the most stable unit that owns them.
• If tests cannot be added in the current change, say so explicitly and name the next stable Atom(s) plus the required contract cases.
• Keep test code readable: use each/table cases to reduce duplication, allow only small helpers that do not blur structure, and keep each test focused on one core assertion.

Anti-Patterns

• If splitting increases argument/state passing, roll it back.
• Do not split functions/files for appearance only (avoid utility sprawl).
• If names start turning into long explanations, re-check boundaries.
• Avoid adding abstractions/layers for assumed future reuse.
• Avoid over-abstracted tests and helper sprawl.
• Do not add parameters "for later."
• Do not keep the same policy, decision, calculation, or key-generation rule in multiple owner locations.
• Do not split the same externally observable result across multiple writer locations unless coordinating those writers is itself the explicit responsibility.
• Do not synchronously mirror upstream boundary state into local mutable state unless ownership and reset semantics are explicit.
• Do not create self-feedback loops where a unit reads from an input/update path and writes back into that same path.
• Do not keep new and legacy equivalent paths in parallel without a staged migration plan (owner, exit condition).

Final Gates

• Can the success path be seen in one top-down read?
• Is the current unit stated clearly and still the right unit for this change?
• At function/file scale, is the flow still shallow enough without extra intermediate data structures or helper chains narrating the branches?
• Does splitting reflect real responsibility/boundary changes?
• Can each Atom's I/O be explained in one line?
• Are side effects concentrated at boundaries?
• Are tests contract-focused and concise?
• At capability/subsystem scale, can you name the entrypoint, main orchestrator, and decision owners?
• If more structure work remains across units, is the next re-entry point clear before ending this pass?
• Are parameter growth and old-path handling (cleanup or staged migration) justified and complete?

Optional Completion Evidence

Use this template only when structure itself is the point; otherwise answer naturally.

When the format is useful, provide these four lines:

• Current Unit: function/file | module/use case | capability/subsystem
• Primary Flow: top-down in 3-6 lines
• Boundaries: list of I/O boundaries
• Tests: added... or deferred because...; next stable Atom(s):...; required contract cases:... (include freshness/completion contract when relevant)

For refactoring work where rule ownership changed, also provide:

• Decision Ownership: policy/decision/calculation/key-generation rule -> owner unit; duplicated owner removed? yes/no

For refactoring work where signatures/boundaries grew or an old path was replaced, also provide:

• Refactor Check: parameter growth reason / legacy path status (removed, disabled, migration plan)