Breaking

Karpathy Releases Task-Completion Prompt Framework for Claude Opus 5.5

October 8, 2026 at 03:16 EDT

Foundation Models
AI Agents
Software Dev & Coding

Andrej Karpathy, the AI researcher who joined Anthropic in May to lead pre-training research, has published a prompt framework and operational guide designed to make Claude Opus 5.5 finish work autonomously rather than explain processes. The approach, termed "harness engineering," structures the environment around the model with loops, verification gates, and clear termination conditions so the AI completes tasks end-to-end without human re-prompting.

AI RESEARCH · OCTOBER 2026

Task-Completion Harness for Opus 5.5

Karpathy's prompt framework and loop engineering discipline enabling Claude to finish work autonomously.

3–5×

Harness design task completion lift vs model alone

41% → 11%

Coding error reduction (Karpathy CLAUDE.md)

Medium

Opus 5.5 effort ≈ Opus 5 high

Core Principles

Specify End State

Define exactly what "done" means (passing tests, complete artifact) rather than "improve this."

Build Verification Loop

Run tests, lint, review to close the gap. Make the loop repeat until success criteria are met.

Strip Reasoning Phrases

Remove "think carefully" cues. Opus 5.5 thinks continuously; effort controls tune depth.

Tune Effort Strategically

Start at medium (cost-efficient); reserve xhigh/max for tasks where reasoning depth matters.

Developer Reception: Strong adoption for repeated/long-running work; requires upfront investment in clear metrics. Creative or exploratory tasks may not fit the binary success-criteria model.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.