Clean Code: Structure, Naming and Cohesion

Most defects are not clever. They are the ordinary consequence of code that is harder to read than it needed to be. A function that does three things hides which one broke. A name that lies sends the next reader to the wrong file. A module that depends on everything cannot be changed without changing everything.

So the work is not cleverness. It is applying a small number of well-established structural disciplines consistently, and being willing to keep applying them after the code already works.

The line that does most of the work:

Code is read far more often than it is written, and almost always by someone with less context than the author had. Optimise for the reader who arrives in six months knowing nothing. If your change makes sense only because you remember why, it is not finished.

The five disciplines

Every recommendation below reduces to one of these. Naming them keeps a review from collapsing into taste.

Single responsibility. A unit should have one reason to change. When you cannot describe a function without "and", it is doing two things, and the two things will need to change on different schedules. Split along the seam where the reasons differ, not where the line count is convenient.

Cohesion over proximity. Things that change together belong together. Code grouped by technical layer (all the controllers here, all the models there) scatters a single feature across the tree, so every change touches five directories. Group by what the code is about.

Explicit dependencies. A unit should declare what it needs rather than reach for it. Constructor parameters and function arguments are declarations; module-level singletons, global config and ambient state are not. The test for this is whether the unit can be exercised without standing up its whole world.

Names that survive being read alone. A name is the only documentation that cannot go stale, because changing the code without changing the name is visible. Prefer a longer name that is accurate to a short one that is approximately right. elapsed_ms beats t.

Duplication is cheaper than the wrong abstraction. Two similar blocks are a fact. One premature abstraction over them is a commitment, and unwinding it later costs more than the duplication ever did. Wait until the third occurrence, and until the three genuinely share a reason to change rather than a shape.

How to work

1. Read before you write. Understand what the current shape is for. Code that looks wrong is often load-bearing in a way the diff does not show.
2. Name the problem before proposing the fix. "This function is 200 lines" is an observation. "This function mixes request parsing, business rules and persistence, so a change to any one of them risks the other two" is a problem.
3. Prefer the smallest change that removes the problem. A rewrite that also improves five unrelated things cannot be reviewed, so it will be approved on trust rather than reading.
4. Say what you did not change and why. A review that only lists changes reads as though everything else was examined and approved.
5. Leave the reasoning, not just the result. The next reader needs to know which constraint drove the shape.

Routing

Read the sub-skill matching the work, then follow it. If more than one applies, read both; if none clearly applies, continue with this document.

Sub-skill	Use for
design	Designing a new interface, module, schema or type. What the shape should be before it has callers.
audit	Reviewing code that already exists for structural problems. Diffs, PRs, whole files.
retro	Something broke and you are deciding what to change so the class of problem is less likely.
ux	Forms, flows and screens. Structure and clarity of user-facing interaction code.
authz	Permission and access-control code. Structure, clarity and testability of authorisation logic.
data	Pipelines, transformations, queries and reporting code.
ops	Deployment, migration, configuration and infrastructure code.
guardrails	Lint configuration, CI setup, formatting rules and repository conventions.
agent-guardrails	Repository configuration for AI coding assistants.
llm	Code that calls a language model API and handles its output.

What good output looks like

Concrete, ordered by impact, and anchored to the code in front of you.

• Point at specific lines. "The process function" is reviewable. "The code" is not.
• Order by cost of leaving it. A misleading name in a widely-called function costs more than a long function nobody touches.
• Show the shape you mean. A two-line before-and-after communicates more than a paragraph describing it.
• Separate the structural from the stylistic. Formatting is not a finding; a tool should already own it. Spending review attention on brace placement is how the structural comments get skimmed.
• Do not pad. Three findings that matter beat eleven that include four restatements of the same point and three matters of preference.

What to avoid

Cargo-cult patterns. A factory that has one implementation, an interface with one implementer, a layer that only forwards calls. Each adds a hop the reader must follow and buys nothing until the second case actually exists.

Rewriting to taste. If the existing code is consistent and clear but not how you would have written it, that is not a finding. Consistency within a codebase is worth more than any individual improvement.

Advice that is only true in the abstract. "Reduce coupling" is not actionable. "The report builder imports the HTTP client directly, so it cannot be tested without a network stub; pass the fetched rows in instead" is.

Confusing length with complexity. A long function that does one thing in a straight line is easier to follow than three short ones that pass state between them. Count reasons to change, not lines.

What it looks like in practice

A worked example, because the disciplines above are easy to agree with and hard to apply.

Here is a function that works, passes its tests, and is still a problem:

def process(data, flag, config):
    if flag:
        rows = [r for r in data if r.get("active")]
    else:
        rows = data
    out = []
    for r in rows:
        v = r["amount"] * config["rate"]
        if config.get("round"):
            v = round(v, 2)
        out.append({"id": r["id"], "value": v})
    db.save(out)
    return len(out)

Four separate problems, in order of what they cost:

It has three reasons to change. Filtering policy, the arithmetic, and persistence all live here, so a change to any one risks the other two. That is the finding; the length is a symptom.

process, data, flag, v and out name nothing. A reader has to execute the function mentally to learn what it is for. flag is the worst of them: a boolean parameter at a call site reads process(rows, True, cfg) and communicates nothing at all.

It reaches for db rather than declaring it. The function cannot be tested without a database, so it will be tested less than the arithmetic deserves.

Its return value answers a question nobody asked. A count of rows saved is not what a caller of a transformation wants; they want the rows.

The restructured version separates the reasons to change and lets the names carry the meaning:

def active_only(rows):
    return [r for r in rows if r.get("active")]

def apply_rate(rows, *, rate, round_to_cents):
    for r in rows:
        value = r["amount"] * rate
        yield {"id": r["id"], "value": round(value, 2) if round_to_cents else value}

Persistence moves to the caller, which is where the decision to save belongs. The two functions can now be read, and tested, without each other.

Note what did not change: the arithmetic is identical, and no abstraction was introduced for a second case that does not exist yet. The goal is to remove the reasons the code was hard to read, not to demonstrate patterns.

Judgement calls worth making explicitly

When to stop. Refactoring has no natural end, and a review that keeps going becomes a rewrite nobody can check. Stop when the unit has one reason to change and its names are accurate. Further improvement is preference.

When consistency beats correctness. A codebase that does something suboptimally but uniformly is easier to work in than one where every module is locally optimal and globally inconsistent. Match the surrounding code unless the surrounding code is the problem you were asked about.

When to leave duplication alone. Two call sites doing similar work with different reasons to change should stay separate. Merging them creates a shared unit that both must now agree about forever.

When a comment is the right answer. Structure cannot express why a threshold is 30 seconds rather than 60, or which regulation requires a field. When the reason lives outside the code, write it down; renaming will not carry it.

When to say nothing. Not every file needs an opinion. A review that finds something in every file trains the reader to skim, and the one finding that mattered goes past unread.