auto-compound AI skills
Your corrections compound into permanent skills
Temet detects repeated corrections across AI sessions and injects stable rules back into your agent. Each session makes the next one better, automatically.
The problem: agents keep making the same mistakes
Every developer who works with AI coding agents knows the pattern. You correct the agent once, it adjusts. Next session, same mistake. The correction you gave yesterday is gone. Your agent has no memory of what went wrong before. The standard fix is to write instructions manually: CLAUDE.md files, system prompts, skill definitions. This works, but it requires you to notice the pattern, write the rule, and maintain the file. Most corrections never make it into permanent instructions.
What Temet does after every session
Temet runs a lightweight audit after each AI session using a SessionEnd hook. It parses your conversation, detects moments where you corrected the agent, and stores them as structured decision traces. When the same correction appears across three or more sessions, Temet elevates it to a stable rule. This is not a guess. It is a pattern confirmed by repeated real behavior: you told the agent the same thing multiple times, so it matters.
The missing link: feeding rules back into the agent
Until now, Temet detected patterns and stored them locally. You could export them with temet rules, but the agent never read them automatically. The new auto-injection closes that loop. When a correction pattern reaches stable maturity, Temet writes it directly into your project's .claude/ directory. The next time your agent starts a session, it reads the rule. The correction you gave three sessions ago is now a permanent behavior change. No manual editing. No copy-pasting rules. The loop closes itself.
How the compounding loop works
Session N: you correct your agent. Session N+1: Temet detects the correction. Session N+2: the same correction appears again, pattern is now recurring. Session N+3: third occurrence, pattern becomes a stable rule. Temet injects it into your agent's skill files. Session N+4: your agent reads the rule at startup. It does not make the mistake again. You never had to write anything. Your past sessions wrote the rule for you.
Why this matters for agent-native workflows
This is the difference between an AI tool you use and an AI partner that learns from working with you. Most AI products reset to zero every session. Temet makes each session build on the previous one. The implications go beyond personal productivity. If your correction patterns are stable enough to become rules, they are stable enough to become operational knowledge. The same patterns that fix your agent today become the documented expertise that others can learn from tomorrow.
FAQ
Does Temet modify my CLAUDE.md directly?
No. Temet writes to a separate file (.claude/temet-rules.md) so your manual instructions stay untouched. You can review, edit, or delete any auto-injected rule.
How many corrections does it take to create a rule?
Three occurrences across at least two distinct sessions. This threshold filters noise and ensures the pattern is real, not a one-time preference.
Does this work with Claude Code and Codex?
Yes. Temet parses both Claude Code and Codex session formats. Any AI coding agent that writes .jsonl session logs can be analyzed.
Can I disable auto-injection?
Yes. Run temet audit without the --track flag and no rules are injected. The feature only activates during tracked audits, typically via the SessionEnd hook.
Next step
Use this guide in practice with Temet's audit, tracking, and profile workflow.
Connect your agentPublished March 19, 2026