oleg494 Production-first: decisions grounded in relevant evidence
Workflow
- Define the decision and acceptance criteria. Read the existing implementation, project constraints and known decisions before inventing a second convention.
- If those sources settle an ordinary local choice, reuse the existing pattern. Do not require web research for a local definition or a known stdlib operation.
- Research unresolved external facts through current primary documentation, source code and relevant production experience. Cross-check disputed or high-impact claims. No fixed source/query count or popularity gate.
- Choose by fit, correctness, maintenance and measured constraints. Widespread adoption is evidence, not proof that an approach fits this project.
Test-before-integration
For an unfamiliar dependency or mechanism, run a task-sized sandbox probe of the uncertain behavior before integrating. Benchmark only when performance motivates the choice; compare a simpler existing option when material. Installation, paid calls and outward actions still follow AGENTS.md authority. Record consequential decisions in the existing project record or authorized memory; an ADR is not a mandatory artifact for every implementation detail.
Research depth
- Narrow authoritative fact: read the owning source and stop when resolved.
- Consequential or uncertain choice: compare relevant alternatives, seek counterevidence and test the remaining risk. Broaden only when it could change the decision.
Three decision principles (filter before code)
- DRY — one piece of logic and one piece of knowledge in one place. Duplication = two places that must change together.
- KISS — the simpler option, if it closes the task. Complexity is justified when the simple one can't cope, not "for the future".
- YAGNI — don't build what wasn't asked for. "Might come in handy" is an insufficient reason.
A decision that violates at least one principle without a clear reason → reconsider.
Checklist before a decision
- Existing project constraints and patterns considered
- Material external uncertainty checked against appropriate primary sources
- Unfamiliar behavior exercised before integration; performance claims measured
- Chosen approach meets the full request without unnecessary machinery
- Remaining uncertainty and consequential tradeoffs reported