Code Design Review

Code Design Review

Judge whether a design holds up against software-engineering principles, from the one perspective that matters: a future maintainer doing real tasks. This is consulting, not gatekeeping — the deliverable is a verdict with ranked concerns, not a merge decision. Focused examination is allowed; the focus is stated. Report only; no edits without separate authorization.

Subject

Fix what is being judged before judging it: the design surface — module boundary, API, data flow, state ownership, design doc, or the shape of a change — and the maintainer tasks it must serve (extending it, debugging it, changing it safely, understanding it cold). Every concern anchors to one of those tasks; nothing is judged against taste, symmetry, or fashion. When the artifact is readable, read it — don't judge a design from its author's description alone.

Core Model

Every concern maps a concrete surface to a reader task plus at least one symptom and one cause:

• Symptoms — change amplification (one concept, many edits) · cognitive load (too much non-local knowledge) · unknown-unknown risk (needed information not discoverable, or the maintainer can't know it's needed).
• Causes — dependency (the surface can't be understood independently) · obscurity (important behavior, owner, invariant, or coupling isn't obvious).

Style preferences, naming opinions, and unmapped "cleaner" advice do not qualify as concerns. Two further guards:

• Working code is not design evidence. Tests passing prove correctness, not maintainability. Ask whether the surface has the shape it would have if the requirement had been known from the start.
• Principles never override harder contracts: explicit user scope and authorization, public API / schema / persistence / security / compatibility contracts, regression evidence for behavior changes, and domain constraints (auditability, incident response, distributed consistency, data integrity). When a principle conflicts with a harder contract, report the contract risk first; the principle is supporting context.

Lenses

Four axes. Select the ones the subject activates and name the ones skipped, with reason. The axes overlap on purpose — a leaked invariant is a complexity fact and a change-cost fact; report the concern once, under the axis with the sharpest vocabulary for it.

Complexity — can a maintainer understand this, and how much must they hold in their head?

Sources: John Ousterhout, A Philosophy of Software Design — https://web.stanford.edu/~ouster/cgi-bin/aposd.php · Artem Zakirullin, Cognitive load is what matters — https://github.com/zakirullin/cognitive-load · Dan North, CUPID — https://dannorth.net/cupid-for-joyful-coding/

• Apparent complexity is what a maintainer experiences doing a real task, not a property of the text. Line count, file count, and function length measure nothing by themselves.
• Working memory is small — roughly four non-local facts at once. Every fact a reader must import from elsewhere (a flag's meaning, an ordering constraint, another module's internals) occupies a slot; past the limit, reading becomes re-deriving.
• Complexity is incremental. Each tolerated branch / flag / fallback / nullable / wrapper / undocumented invariant compounds. Ask whether a change reduces underlying complexity or adds another tolerated shortcut.
• Deep modules are about ratio. Small caller-facing interface relative to hidden useful behavior. Large ≠ shallow; short ≠ deep. Ask what callers must know about ordering, retries, cache, validation timing, storage format, fallback rules.
• Beware many shallow modules. Decomposition has a cost: each extra module adds its responsibility and its interactions to the reader's working set. A cohesive large module often reads cheaper than five fragments; "smaller classes and shorter functions" is not a simplification argument.
• Make the common case simple. Common callers shouldn't pass options, know internal defaults, or handle rare states. Don't make all callers pay for flexibility only one needs.
• General-purpose, not speculative. General enough for known nearby needs; not contaminated by product-specific or one-off policy; not justified by imagined future extension.
• Information leakage is the primary red flag. Knowledge that should belong to one owner appears elsewhere — duplicated rules, mirrored validation, exposed data structures, ordering assumptions, generated-artifact coupling, error semantics, comments describing another module's internals. Ask which decision leaks, where it should live, and how many places must change if it changes.
• Different layer, different abstraction. A layer that only forwards calls, renames parameters, or wraps lower-layer concepts adds cognitive load without changing the model. Removing it should make the system harder — not easier — to understand.
• Temporal decomposition is suspicious. Splitting code by execution order scatters one concept across phases. Group knowledge by ownership and abstraction, not "first do this, then do that".
• Pull complexity downward. The owner absorbs complexity so callers stay simple — without hiding important errors or state. The right owner centralizes unavoidable knowledge.
• Predictable: code does what it looks like it does. No surprising side effects behind an innocent name, no action at a distance, no behavior depending on invisible state or call order. A deviation from the least-surprise reading is a defect in the interface, not in the reader.
• Comments are interface tools. Capture what code can't say: contract, units, invariants, rationale, side effects, concurrency, cross-module coupling. A long comment compensating for a weak abstraction is a smell, not a fix.
• Names and consistency are cognitive contracts. Same concept → same name; different concept → different name. Deviation from local convention needs a visible semantic reason.
• Performance needs measurement. Optimization-motivated complexity needs evidence of a real bottleneck and should stay behind a simple interface.
• Reject: "fewer lines" or "fewer files" as simplicity proof · "short function" / "small class" as goals in themselves · "large module" as a proxy for shallow · "it follows a pattern" when the pattern isn't local or semantically justified · "add a comment" instead of fixing a weak interface.

Change economics — what does the next change cost, and is now the time to pay for structure?

Sources: Kent Beck, Tidy First? — https://tidyfirst.substack.com/ · Sandi Metz, The Wrong Abstraction — https://sandimetz.com/blog/2016/1/20/the-wrong-abstraction · Carson Gross, Locality of Behaviour — https://htmx.org/essays/locality-of-behaviour/ · tef, Write code that is easy to delete — https://programmingisterrible.com/post/139222674273/write-code-that-is-easy-to-delete-not-easy-to

• Software design is an investment decision, not a virtue. Coupling is the cost of a change spreading across elements; cohesion is the cost of a change within one element. Structure is good when it makes the changes that actually arrive cheaper — not when it matches a shape catalog.
• An element can lack cohesion by being too large or too small: a fragment solving part of a problem must be coupled to the elements solving the rest, and changing the solution means changing them all.
• Structure changes and behavior changes are different activities with different risk. A design where behavior can't be changed without restructuring, or restructured without behavior change, has already coupled the two — that is itself a finding.
• Tidy first only when it pays. Restructuring is justified by a change it makes cheaper — one that is planned or demonstrably recurring — not by discomfort. "Tidy after" and "don't tidy" are legitimate outcomes of the same calculation.
• Cohesion follows the domain. Things that change together belong together; structure should mirror the domain's change patterns, not technical layering. When answering one domain question requires touching N technical layers, the architecture itself is change amplification.
• Duplication is cheaper than the wrong abstraction. Abstract when the pattern is understood, not when text repeats — two occurrences form a line that may point the wrong way. A shared abstraction that accumulates parameters and conditional paths for each new caller has already failed; the recovery is re-inlining the duplication and letting it show the right shape.
• Weigh DRY against locality of behavior. Extracting shared behavior trades local readability for a new coupling. The behavior of a surface should be understandable at its point of use; ask how far a reader must travel from trigger to behavior before treating deduplication as free.
• Design for deletion. The measure of a boundary is the blast radius of removing what's behind it. Code that is easy to delete — isolated, unshared state, countable callers — is cheap to be wrong about; code woven into everything must be right forever. Ease of deletion, not ease of extension, is the default to optimize for.
• Reject: "more DRY" when it creates the wrong shared dependency · "we might need it later" ahead of a demonstrated change pattern · "refactor while we're here" without naming the change it makes cheaper · preserving an abstraction because it exists while it accretes parameters and modes — sunk cost is not cohesion.

Boundary correctness — can invalid state exist past this boundary, and where do effects live?

Sources: Alexis King, Parse, don't validate — https://lexi-lambda.github.io/blog/2019/11/05/parse-don-t-validate/ · Gary Bernhardt, Functional Core, Imperative Shell — https://www.destroyallsoftware.com/screencasts/catalog/functional-core-imperative-shell

• The strongest correctness is structural: a state that cannot be represented cannot occur, and no downstream code needs to defend against it. Prefer a boundary that transforms less-structured input into a more-structured type once over checks that verify and then forget.
• Stratify the program into a parse phase (all input rejection happens here) and an execute phase (illegal input is impossible here). Failure concentrates where it can be handled; the core is freed from defensive noise.
• Parse, don't validate. A check whose result isn't captured in the type or structure will be repeated downstream — or worse, assumed. Ask: after this validation, what stops an unvalidated value from reaching the same code path?
• Make illegal states unrepresentable. Nullable-but-never-null fields, paired fields with an implicit invariant ("when status is X, payload is set"), stringly-typed enums, boolean combinations of which only three of four occur. Each representable-but-illegal state is a latent branch every reader must rule out and every agent must rediscover.
• Functional core, imperative shell. Decision logic stays pure; effects — IO, clock, randomness, network, persistence — live at the shell. The test: can the interesting logic be exercised without mocks? A design that needs test doubles for its own domain rules keeps effects in the wrong place.
• Mirrored validation marks a failed boundary. The same rule enforced in UI, API, and storage layers means the boundary never captured its guarantee in a type or contract; the copies will diverge silently.
• Define errors out of existence — carefully. Eliminating an error case is valuable only when the new contract is explicit, safe, observable, and authorized. Not a license to hide security, data integrity, auditability, observability, compatibility, or recovery information.
• Reject: defensive checks scattered "just in case" instead of a boundary that makes them unnecessary · a doc comment promising an invariant a type could enforce · mocking your own domain logic to reach testability — a boundary defect wearing a test harness.

Agent legibility — can the newest maintainer, human or agent, operate here and verify its own work?

Sources: emerging practice, 2024–2026; e.g. Stack Overflow, Building shared coding guidelines for AI (and people too) — https://stackoverflow.blog/2026/03/26/coding-guidelines-for-ai-agents-and-people-too/

• An agent is a maintainer with extreme versions of ordinary human limits: zero memory of past decisions, a hard working-set budget, and reliable obedience only to what is machine-checked. Design that serves it serves cold human readers too — this axis contradicts none of the others; it re-weights them toward explicitness and verifiability.
• What an agent actually obeys is what fails a check: types, tests, linters, build gates. A prose convention without enforcement is a suggestion.
• Machine-checkable contracts over prose conventions. Every convention that matters should have a type, test, or lint rule that fails when violated. Ask which of the repo's stated rules would actually stop a wrong change.
• Tests are the executable spec. To the newest maintainer, a behavior without a test is undefined behavior — nothing marks it as intended, so nothing prevents its loss. Tests assert public behavior, not implementation details, or the spec they encode is the wrong one.
• Self-verification is a design property. After a change, one discoverable command should answer pass or fail. A repo whose correctness is judged by tribal knowledge or manual poking cannot be safely maintained by an agent — or by a new hire.
• Greppable, boring, explicit. Searchable names (no dynamically constructed identifiers), one concept one name, behavior visible at the point of use, no action at a distance via reflection, metaprogramming, or implicit registration. Cleverness that compresses text expands the working set.
• Context is a budget. The knowledge needed to change a surface should fit in a bounded neighborhood: the file, its imports, and written-down local rules (AGENTS.md, CLAUDE.md). Ask what must have been read to make this change without breaking an unstated rule — every unwritten rule is an unknown-unknown planted for the next maintainer.
• Reject: "it's documented in the wiki" — off-repo knowledge is invisible to the newest maintainer · "the team knows" — institutional memory is not a contract · conventions enforced by reviewer vigilance instead of tooling.

Judgment

Depth follows risk: concentrate where complexity concentrates — the interfaces most callers touch, the knowledge duplicated across owners, the layer whose removal would change nothing. Sampling is legitimate here; what makes it honest is naming what was examined and what was not. For each concern, state the smallest correction that resolves the mapped symptom. When a material design choice is questioned, compare a real alternative owner, representation, or boundary — no strawmen, and no ceremonial alternatives on trivial choices.

Output

User's primary language for prose; keep code symbols and paths original. Return:

1. Verdict — holds up / holds with concerns / does not hold — with the load-bearing reasons.
2. Concerns, ranked by impact on maintainer tasks, each naming surface, reader task, symptom, cause, and smallest correction.
3. Lenses checked and passed, briefly.
4. Boundary — lenses not applicable or not assessed, surfaces not examined, each with its reason.

Stop Rules

Finish only when the verdict is stated with its reasons, every concern carries the full surface → reader task → symptom → cause mapping, and the examined/unexamined boundary is named. If the subject or its maintainer tasks are too unclear to judge, stop with the clarifying question rather than substituting taste. No edits, commits, or pushes.