← ClaudeAtlas

harness-engineeringlisted

Design the scaffolding around an LLM — the tool set, permission model, context-management hooks, verification loops, and orchestration — rather than just the prompt. Use when building or hardening an agentic coding tool, CLI agent, autonomous pipeline, or any system where the harness (not the model) decides what the agent can see, do, and recover from. Covers tool design, sandboxing, approval gates, retries/escalation, observability, and eval-driven iteration.
RaveenDi/Skilloadout · ★ 0 · AI & Automation · score 78
Install: claude install-skill RaveenDi/Skilloadout
# Harness Engineering for AI Agents A **harness** is everything around the model: the tools it can call, what context it's fed each turn, what it's allowed to do unsupervised, how failures are caught and retried, and how progress is shown to a human. Two agents running the *same model* with the *same prompt* behave completely differently if one has a good harness and one doesn't. This skill is about designing that scaffolding deliberately instead of letting it accrete. ## 0. Harness vs. prompt engineering Prompt engineering optimizes *what you say to the model*. Harness engineering optimizes *everything else*: what the model can act on, what it's shown, what's verified before and after each action, and what happens when it's wrong. A weak model in a strong harness often beats a strong model in a weak harness — the harness is where reliability actually comes from. ## 1. Tool design - **Small, composable, verifiable tools** beat one giant "do anything" tool. Each tool should do one thing, return a result the model can check, and fail loudly with an actionable message — not a silent no-op or a stack trace dump. - **Shape tools around the task's real primitives**, not around the underlying API. If the task is "edit a file," don't make the model regenerate the whole file through a raw write call — give it a diff/patch tool so mistakes are small and reviewable. - **Make illegal states unrepresentable.** If a tool call can't possibly succeed given current state (e.g.