edram A set of constraints an agent applies while producing code, biased toward less code and tighter scope. Distilled from Andrej Karpathy's coding guidelines and Sentry's code-simplifier, plus an explicit comment discipline. The aim is the opposite of most AI failure modes: not more features, more abstraction, more lines — but the smallest change that provably solves what was asked.
- Think before coding — surface assumptions and the simpler option before writing a line.
- Write the minimum code that solves the stated problem; nothing speculative.
- Keep changes surgical — every changed line traces directly to the request.
- Comment deliberately — important declarations, external sources, and hard-to-understand logic must be documented; explain why, constraints, and contracts, never narrate how.
- Define success criteria and verify before claiming done.
Core
| Topic | Description | Reference |
|---|---|---|
| Simplicity | Minimum code, no speculation, no over-simplification into cleverness | core-simplicity |
| Scope | Surgical edits, match existing conventions, preserve behavior | core-scope |
| Comments | Required for important intent, contracts, sources, and opaque logic; explain why, not how | core-comments |
Features
| Topic | Description | Reference |
|---|---|---|
| Think First | Clarify intent, surface assumptions and tradeoffs before coding | features-think-first |
| Verify | Define verifiable success criteria; loop until proven, not just written | features-verify |