Write Maintainable Code

Write Maintainable Code

Use the fewest concepts that satisfy the already-approved result and its acceptance evidence. Minimize current maintenance burden, not raw line count.

Keep the ownership boundary

This skill is a subordinate implementation lens inside the public method:

• During CONVERGE, generate code/no-code route alternatives and possible code

ownership seams. Let reasoning-doctrine own the stage, recommendation, rejected alternatives, hard stops, safe outcome slice, and transition to execution.

• During EXECUTE, apply this lens to the selected route. Do not redefine the

selected outcome, acceptance evidence, safe slice, or mutation method.

• During an authorized repair, apply only this execution lens.

run-review-repair-loop continues to own review, scoring, and repair iteration.

• When governance is active, let governed-operator own authority, seats,

scope, the formal contract, verdicts, publication boundaries, and Done.

If the outcome or acceptance evidence is unresolved, return to reasoning-doctrine. If authority, seat responsibility, scope, or publication authority is unresolved, return to governed-operator. If a finished change needs review-only grading, use run-review-repair-loop and keep its Builder self-review separate from the independent-review and Done boundaries owned by governed-operator; do not score or approve under this lens. Re-enter this skill only for an authorized repair.

Compare routes before adding code

1. Echo the fixed outcome and acceptance evidence without reopening them.
2. Inspect the current behavior and evidence. If they already satisfy the

acceptance check, choose no change and stop.

1. Compare existing behavior, configuration, reuse, deletion, documentation,

and no-code routes before proposing new code. A route is sufficient only when the fixed acceptance evidence can prove it.

1. Under CONVERGE, return the route and ownership-seam alternatives to the

public method for selection. Once a route is selected, execute only that route.

When different routes would satisfy different acceptance checks and the current evidence does not identify which check applies, do not invent one. Return that gap to reasoning-doctrine. Otherwise, let the public method select only the route proved necessary by the fixed check.

Do not turn this comparison into a second implementation brief or a competing return schema.

Bound the implementation

Before changing code, name:

• the smallest code ownership seam that can own the current behavior;
• the concepts required by the selected result; and
• the tempting concepts declined because the current result does not need

them.

Decline a dependency, layer, abstraction, state or status flag, endpoint, migration, generic helper, or compatibility path when its only justification is hypothetical reuse or an unselected future. Prefer readable, explicit, testable code over dense code golf. A slightly longer implementation can be more minimal when it carries fewer concepts or less maintenance risk.

Treat file size as a signal, not a diagnosis. Extract a bounded seam only when the extraction lowers verified current risk; do not split a file merely to make it shorter.

Apply this lens inside the already-selected safe outcome slice. Remove only code made obsolete by the current change. Report unrelated cleanup instead of performing it.

Respect operator priorities

Absent a contrary approved predicate, interpret minimum as minimizing concepts and maintenance burden. If the operator explicitly prioritizes raw line count, surface the readability and testability tradeoff. Follow that preference only when it remains compatible with correctness, safety, repository rules, and authority.

If a line-count preference or new evidence changes the fixed acceptance predicate, selected result, route, seam, scope, or authority, stop this lens and return to the appropriate public reconvergence owner before acting.

Stop and return evidence

Stop as soon as the selected acceptance evidence passes. Do not add cleanup or future-proofing after that point.

Without governed mode, return only a short minimum-sufficiency rationale and a residual-risk note. With governed-operator active, map the route, seam, declined concepts, verification, and residual risk into its existing Builder judgment record. Do not define another brief, score, verdict, approval rule, or Done condition.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Ezra144israel
• Source: Ezra144israel/governed-agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.