A GitHub repo nearing 200,000 stars reduces coding-agent behavior to four rules
import_jmr · x · 2026-07-24
A GitHub repo with nearly 200k stars distills coding-agent behavior into four rules
A developer turned Karpathy’s observations about where coding agents fail into a tiny repo: a single markdown prompt file with four rules.
The rules are not technical tricks:
- Don’t guess quietly.
- Don’t overbuild.
- Don’t touch what wasn’t asked.
- Verify before calling it done.
The post’s point is that much of a modern agent product spec now looks like senior-engineer advice written down as policy.
More from coding & agent
- ChatGPT Voice tips warn users to offload work into new threads and expect context resets — jxnlco · 2026-07-24
- 30 open-source AI repos covering agents, local inference, and workflow automation — Shruti_0810 · 2026-07-24
- Context Engineering 2.0 argues that smarter systems need shared memory and subagents — mdancho84 · 2026-07-24
- Context Engineering 2.0 says the next AI shift is toward context-cooperative systems — mdancho84 · 2026-07-24
- A new Context Engineering 2.0 thread says prompt engineering is over — mdancho84 · 2026-07-24
- OpenAI feels nice for coding, but Gemini is still faster — burhop · 2026-07-24