Compound code review skill update: fewer subagents and custom standards support
iamrobotbear · x · 2026-08-31
Compound Engineering's ce-code-review skill received a major overhaul. Updates include: 1. Reduced overlap between personas, cutting subagents to a maximum of 6 to save tokens and reduce duplication. 2. Added support for a CODINGSTANDARDS.md file in the repo root to enforce specific rules during review, which is more effective than stuffing them in AGENTS.md. 3. Avoids role-play directives like "review like Fowler" in favor of naming principles from Fowler, Google, Ousterhout, and OWASP that are proven to help. The update cites a paper noting that while expert personas improve alignment, they can damage accuracy.
More from coding & agent
- Agents Develop 'Classical Chinese' to Compress Long-Chain Task Context — wzenus · 2026-09-01
- Debugging multi-step LLM apps quickly turns into detective work — Sensitive-Parsnip-12 · 2026-08-31
- Lanyon AI generates 32k lines of formally verified MHD solver code — jfischoff · 2026-08-31
- LLMs Only Guess Next Tokens; Agency Is Just Scaffolding — sethjuarez · 2026-08-31
- Soleio Shares Grok Bot Template for Auto-Replying to Emails — soleio · 2026-08-31
- What should Codex verify before a change is ready for human review? — RocketSeven · 2026-08-31