decision traces AI
Decision Traces: How Temet Turns Your AI Corrections Into Reusable Rules
Temet CLI 0.3.6 introduces decision traces: structured capture of your AI corrections, pattern detection, and rule export. Turn your expertise into agent rules your AI tools can follow.
The Problem: Your Expertise Is Trapped in Your Head
If you run ads, do SEO, or consult on anything technical, you make thousands of micro-decisions every week. Which headline variant to keep. Which bid adjustment to override. Which recommendation to reject because the context is wrong. These decisions represent years of accumulated expertise, and none of it scales. You sell time because your judgment lives in your head. You are the bottleneck. Every correction you make to an AI assistant is a signal, but today those signals vanish the moment the session ends. The next session starts from zero. The next person who asks the same question gets the same wrong first draft. Your expertise stays locked inside you, unstructured and invisible.
Decision Traces: Capturing What You Actually Do
Temet CLI 0.3.6 introduces decision traces. Every time you correct an AI agent, reject a suggestion, or override a default, Temet captures a structured trace: what the agent proposed, what you chose instead, and why the correction matters. This is not a skill tag or a competency label. It is the decision itself, preserved with enough context to be useful later. Decision traces work on real Claude Code sessions (.jsonl files) already on your machine. There is no setup, no integration, no prompt engineering. Temet reads the conversation, identifies the moments where you steered the agent, and records the correction as a first-class object. Over time, your corrections become a structured record of how you actually work.
From Corrections to Patterns to Rules
A single correction is interesting. Ten similar corrections are a pattern. Twenty stable corrections are a rule. Temet's pipeline works in three stages. First, corrections accumulate as decision traces, each one tagged with the domain and the type of override. Second, when the same type of correction appears repeatedly across sessions, Temet flags it as a recurring pattern. Third, when a pattern is stable enough (consistent direction, no contradictions, observed across multiple projects), Temet promotes it to a stable rule. This is the compound effect of captured expertise. Each session you run makes the pattern detection better. Each pattern that stabilizes becomes a rule you can export. The more you work, the more valuable your rule library becomes.
temet rules: Export Your Expertise for Your Agents
The new temet rules command takes your stabilized patterns and exports them as rules your AI agents can follow directly. Run temet rules --format claude and you get a set of structured rules ready to paste into your CLAUDE.md, your project instructions, or any agent configuration file. Each exported rule includes: the original correction pattern, the domain it applies to, the direction (what to do instead), and a confidence level based on how many sessions contributed. These are not generic best practices. They are your practices, extracted from your real work, proven stable across sessions. The rules are portable. Paste them into CLAUDE.md for Claude Code. Add them to your agent's system prompt. Share them with a teammate. The format is plain text, readable by humans and machines alike.
Operational Loops: Proof That Your Expertise Is Encoded
Decision traces are not a CV feature. They are a proof mechanism. Each operational loop Temet builds shows the full chain: a problem was detected in your AI sessions, corrections were captured as decision traces, a pattern stabilized over time, rules were exported, and an improvement was observed in subsequent sessions. This is what separates encoded expertise from a list of skills. A skill says you know something. An operational loop proves it works. It shows the before and after. It shows the correction that triggered the pattern. It shows the rule that changed how your agent behaves. And it shows the measurable difference in output quality. For service professionals, this is the path from selling time to selling encoded judgment. Your rules become reusable assets. Your expertise becomes transferable without you being in the room.
Try It Now
Decision trace extraction runs locally, requires no API key, and works on sessions you already have. Step 1: Run npx @temet/cli audit --track to scan your Claude Code sessions and start building your correction history. Step 2: Run temet rules --format claude to export your stable patterns as agent rules. Step 3: Paste the output into your CLAUDE.md or agent configuration. Zero setup. Fully local. Your expertise, finally structured and portable.
FAQ
Do I need an API key?
No. Decision trace extraction is fully local, zero LLM required. Optional narration needs ANTHROPIC_API_KEY.
What AI tools does this work with?
Currently Claude Code sessions (.jsonl). Codex support is coming.
Is my data sent anywhere?
No. Traces, patterns, and rules stay in ~/.temet/ locally. Publishing is opt-in via temet share.
Next step
Use this guide in practice with Temet's audit, tracking, and profile workflow.
Connect your agentPublished March 16, 2026