Karpathy-inspired harness prompt turns Opus 5.5 into a self-verifying execution engine

Roger_M_Taylor · x · 2026-10-23

@0xPontiuz adapted ex-OpenAI founding member Andrej Karpathy's public agent-harness guidance into a full harness prompt for Opus 5.5, claiming it turns the model into an execution engine that actually ships, alongside a guide: "Grok Bot + Opus 5.5: Build a 24/7 AI Engineering Team in 12 Steps."

The prompt's structure (visible in the shared image):

Run tasks inside a harness, not one long chat: plan, context, tools, checks, a visible verification report, and an updated CLAUDE.md form a closed loop
Seven layers: plan (plan.md steps, parallelizable marked), context (load only needed files, long logs saved as paths), autonomy (three autonomy levels per step), permissions (confirm before publishing/spending, no two sessions editing the same file), verification (every step proves itself with real signals — tests, builds, screenshots — tracked in checks.md), feedback (failed checks bounce back with reason, capped at 3 attempts, corrections written to CLAUDE.md), completion (every output has an openable path, checked against the saved version)
Execution loop: read plan.md, run steps in parallel, verify against checks, rerun failures, run the full suite before merge until done or blocked
The Grok Bot + Opus 5.5 setup: one Chief plus four specialist agents, Cursor cloud agents, shared memory, automated testing and a review loop that fixes GitHub issues and prepares PRs while your laptop is closed

Original post →

Karpathy-inspired harness prompt turns Opus 5.5 into a self-verifying execution engine

More from coding & agent