Starter kit v0.1 - free
Context compaction and retry logic dissolve into better models. Permissions, spend limits, audit trails and a good instruction file do not - they encode what you will allow, not what the model can do. This kit is those parts, ready to adapt. Three files, plain text, no signup required.
Instruction-file skeleton that works with Claude Code, Codex, OpenCode, Qwen Code, Goose and OpenHands. The important section is the ratchet log: every rule traceable to a failure that actually happened, none added on prediction. That is Addy Osmani's ratchet - a harness built that way is a record of what went wrong, and it survives a model upgrade better than one built from predictions.
Sandboxing, credential scoping, hard budget caps, loop protection, and an audit trail you can replay. Run it once per deployment, revisit quarterly. This is the checklist that would have caught OpenHands looping on an ambiguous problem before the bill arrived.
OpenCode pointed at local models via Ollama, with the one privacy setting people miss: session titles go to a hosted model by default - this config keeps them on your machine. On Claude Code or Codex? Files 1 and 2 apply as-is; skip this one.
Team kit (shared instruction files, spend attribution) and the self-hosted audit kit are next. One email when they ship - no newsletter, no drip sequence.