Everyone compares models. Almost nobody configures the thing around the model — the instructions it always reads, the permissions it never has to ask about, the hooks that run whether or not it remembers, and the sub-agents that keep one context from carrying every job.
That layer is the harness, and it is the part you keep when the model changes.
Download the packFree · no sign-up · 11 files
The instructions loaded on every single turn
This is the most expensive text in your project — read on every request. Keep it short enough to stay cheap and specific enough to change behavior. The template marks exactly which parts to replace and which to delete.
Permissions you grant once instead of every session
Every command you approve by hand is a turn spent on ceremony. Read-only commands are allowlisted; nothing that writes, deletes, pushes or spends money is pre-approved.
The parts that do not depend on the model remembering
An instruction is a request. A hook is a guarantee. One formats every file that gets written; the other inspects edits before they land and refuses anything holding a credential.
So one context does not carry every job
One agent can technically do everything, and quality falls off long before the window fills. Planner, reviewer and build-fixer each get their own context and one job.
cp CLAUDE.md your-project/ mkdir -p your-project/.claude cp settings.json your-project/.claude/settings.json cp -r hooks agents your-project/.claude/ chmod +x your-project/.claude/hooks/*.sh
To apply it everywhere instead, copy the same files into ~/.claude/. Restart Claude Code, then run /hooks to confirm both hooks registered.
Your real build and test commands in CLAUDE.md, the read-only commands you approve most often in settings.json, and the formatter your project actually uses. Everything else can wait until it annoys you.
The full config layer — instructions, permissions, hooks and sub-agents — ready to copy into your next project. Free, editable, no sign-up.
Download the packBuilding AI at your company? Eight Labs consults for enterprise, SMEs and startups. Details at theaihow.com.
No. Start with CLAUDE.md — filling in your real build and test commands stops the model guessing them every session, which is the fastest payoff in the pack. Add the permissions allowlist next, once you notice yourself approving the same command repeatedly.
No, and that is the point. The harness is the layer you keep when the model changes. Switching models is the smallest change you can make; the config around the model is what carries over to the next one.
It should not. The patterns are deliberately narrow and match live credential shapes — real key prefixes and minimum lengths — rather than any string containing the word key. A noisy hook gets disabled within a day, which protects nothing.
Per project, CLAUDE.md sits in the repo root and the rest in .claude/ beside it. To apply everywhere, put the same files in ~/.claude/ instead. Make the hook scripts executable, restart Claude Code, and run /hooks to confirm they registered.
It is the transferable part of it. Our own CLAUDE.md carries project-specific rules that would be noise in your repo, so the template here keeps the structure and the rule-writing guidance and leaves the specifics to you. The hooks and sub-agents are as we run them.
No. Free to download, use, and change. No sign-up.