AI Engineering

Instructions
that learn

An agent that repeats a mistake in a new session does not have a model problem, it has a memory problem. The answer is not a bigger context window. It is a curated set of operating rules that grows out of real failures, where every rule can be read, questioned and reversed.

System: Internal agent operations platform
Scale: Roughly 200 rules in four months
Role: Sole architect and engineer
Stack: Retrieval / Git / Prompt engineering

The challenge

Months of daily use kept surfacing the same classes of error, because nothing carried a lesson from one session into the next.

The approach

Distil real failures into proposed rules, gate every one behind human review, and keep the result as versioned plain text.

The result

Recurring errors stopped recurring, and the system's accumulated judgement became something a new engineer can read in an afternoon.

01SessionOrdinary daily work. The system is doing real tasks on real business material, which is the only place failures worth learning from actually appear.Claude Code / agent runtime 02Failure seenSomething went wrong and the reason is identified — a path that changed, a format that was assumed, a check that was skipped. What gets recorded is the cause, not the symptom, because the symptom will look different next time.Session logs / structured issue records 03Rule proposedA dedicated agent turns that cause into a proposed rule and nothing more. It cannot adopt its own suggestion, which keeps the system from quietly teaching itself something wrong.Dedicated curator agent 04Human reviewA person decides whether the rule is right before it takes effect. This is the step that makes the whole loop safe: the system improves deliberately rather than drifting, and a rule that turns out to be a mistake can simply be taken back out.Review gate before adoption 05Operating rulesThe accepted rules are plain text under version control, each traceable to the incident that produced it. That means the learning can be read, argued with, diffed and reverted — it is a file history rather than something buried in a model. Long documents are never pasted into the conversation either; they are referenced and edited in place, which is what keeps this affordable.Markdown / Git / semantic retrieval

Fig. 09 — Learning is a file history, not a model weight

Hover any step to read what it doesTap any step to read what it does

Instructionsthat learn

The challenge

The approach

The result

Instructions
that learn