Summary — making the harness your own
Behind a single operation—running /debug once—skill, MCP, rule, hook, and memory fired in sequence and wove a single workflow. Finally, let us look back over the journey to here.
The flow we followed
At each stage, somewhere in the cross-classification of Agent = Model + Harness was at work. The nudging of the utterance is a Computational Guide, the auto-loading of rules is an Inferential Guide, and the .env block and plain-question detection are Computational mechanisms. And /review's multiple perspectives correspond to an Inferential Sensor.
Three cores to take away
- Agent = Model + Harness: reliability is decided more by the design of the harness around it than by the model's intelligence
- Guide and Sensor pairs: guide ahead of time with prose, and catch what slips through with a machine. The more important a norm is, the more you give it both stances
- Grow it with the Ratchet: promote a single learning to rules / hooks, and tighten it into a state where you do not repeat the same failure
Start your own harness small
You do not need to prepare 34 rules or 17 hooks from the start. Quite the opposite—making one pair from one failure is the realistic way to begin.
- Recall one moment when you thought "I made the same mistake again" recently
- Write it as a one-line norm in
rules/(Inferential Guide) - If a machine could detect it, add one hook to
settings.json(Computational Sensor) - The next time you make the same failure, strengthen the norm with the Ratchet
As you turn this small loop, the harness grows to fit the quirks of your own work. The hook settings in settings.json can also be adjusted with a skill like update-config.
The concept of harness engineering itself is also covered, as an overview, in the blog post Introduction to Harness Engineering. Whereas this guide was "the implementation flow of one command," the blog side supplements the perspective of "why we arrive at this design."
In closing
Recall the legend at the beginning. The rails marked 📦 exist in everyone's environment. Harness engineering is building the contents marked 🔧, one at a time, to fit your own failures.
A harness is the chassis that turns a smart model into a "reliable partner." And a chassis is not something you assemble once and are done with. As you drive, you tighten it little by little, like a ratchet that turns in one direction. The next time you think "I made the same mistake again," write one line of it into rules. That is the first step of your own harness.
The key points:
- Behind one command, skill → MCP → rule → hook → memory chain into a single workflow
- The cores are three: "Agent = Model + Harness," "Guide + Sensor pairs," and "grow it with the Ratchet"
- How to start: small—make one pair of a rule and a hook from one failure, and grow it with the Ratchet