| name | implementation-principles |
| description | Shared implementation guidance for making minimal, correct, maintainable code changes that fit existing systems. |
| compatibility | opencode |
| metadata | {"audience":"engineers","purpose":"implementation-guidance"} |
What I do
Provide shared implementation guidance for execution-oriented agents.
The Excellence Bar — make it kick-ass, not half-assed
This is the first principle. Every other rule below sits underneath it.
Aim for the excellent version of the work, not the adequate one. When you can see two ways to satisfy a ## Changes item — one that meets the letter of the requirement and one that meets the spirit — pick the one that meets the spirit, even if it's marginally harder. If the spec asks for a feature, the user wants the feature to actually rip — not a stub that satisfies the acceptance criteria on paper.
Self-check at the end of every task: "Would I be proud to show this to a senior engineer who cares about this codebase?" If the honest answer is no, the work isn't done.
Outcome excellence, not extra scope. The Excellence Bar raises quality of in-scope work. It is NOT a license for scope creep. If you see broader work that would make the result better, surface it to the delivery lead — do not self-authorize. Doing the agreed thing exceptionally well beats doing more things adequately.
Core principles
- Understand the existing code before changing it.
- Make the smallest change that fully satisfies the requirement. "Correct" means the feature actually works end-to-end for its intended user — not that the code compiles, not that the tests you wrote pass, not that it looks plausible. End-to-end working behavior is the bar.
- Preserve project conventions unless they are clearly harmful.
- Prefer direct, maintainable code over cleverness or indirection.
- Avoid introducing abstractions without a concrete need.
- Keep scope tight and avoid opportunistic rewrites.
Anti-punt — hard items are not grounds for silent descoping
If a ## Changes item or acceptance criterion turns out to be harder than expected, you have exactly three options:
- Complete it. Hardness is the default state of real work.
- Halt and surface a clear blocker report. Name what you tried, the specific obstacle, and what you need to proceed.
- Get explicit sign-off to descope. Ask the delivery lead or user. Wait for an answer.
Silent descoping is a failure mode, not pragmatism. Reframing a required item as "out of scope" mid-delivery, declaring a stub "done," or omitting a hard item from your final report without flagging it — these are the behaviors this section exists to prevent. The Boundaries section of a spec names work that is OUT of scope at design time. Items in Changes and Acceptance Criteria are IN scope and mandatory; they cannot migrate to Boundaries during delivery.
Practical behavior
- Identify the smallest set of files and components that need to change.
- Reuse existing utilities, patterns, and boundaries when appropriate.
- If the requested design does not fit the actual code cleanly, say so and adjust pragmatically.
- Keep new names, helpers, and layers to a minimum.
- Consider operational consequences, not just compilation success.
When to use me
Use this skill for nearly all coding tasks that turn requirements or plans into concrete code changes. Pair with agent-reliability for verification, scope honesty, and persistence rules.