Start with the exact failure.
Capture the original alert, summary, tool action, or generated output. Keep the context attached so the rule does not become generic.
How to turn user corrections into AI agent rules
A correction is only useful if it becomes future behavior. This workflow turns user feedback into specific, testable rules for personal AI agents that text, browse, summarize, and build.
Write rules from failures, not preferences from vibes.
The agent needs replacement behavior: what it did, why that was wrong, what source check was missing, and what it should always do next time.
Name the failure type.
False positive, false negative, wrong timing, duplicate, bad source, or wrong action.
Define the replacement.
Tell the agent what should always happen instead, in operational language.
Add source gates.
Specify evidence required before the agent interrupts or acts again.
Retest the pattern.
Watch the next similar event before expanding autonomy.
The correction-to-rule workflow.
Use this loop weekly, or immediately after a high-cost mistake.
Capture the correction receipt.
Save the original output, source context, user correction, and whether it happened in SMS, browser work, summary, or generated deliverable.
Write the replacement rule.
Use this shape: when similar conditions appear, do not repeat the bad behavior; instead check these sources, choose this channel, and take this smaller action.
Retest before trust expands.
For a text message AI assistant, test the next similar event before allowing more proactive alerts.
Rules should follow the agent across channels.
A correction from one workflow often reveals a broader operating policy.
SMS
Turn bad proactive texts into timing, evidence, and quiet-hour rules.
Browser work
Apply corrections to computer-use cache workflows where repeated actions should improve.
Web builds
Preserve design and source corrections when an AI agent builds websites.
Escalations
Make human handoff rules stricter when the agent lacks evidence.
Rule-writing checklist
- Write down the exact wrong output or action.
- Classify the failure before editing the prompt.
- Attach the source context the agent used and the source context it missed.
- Write the replacement behavior with "always" or "never" language only where justified.
- Add a verification step for the next similar event.
- Retire rules that create new false positives.
Example rule
When a delivery update appears after quiet hours, do not text unless the delivery requires a signature, is marked failed, or creates a same-night action. Otherwise batch it into the morning digest with the source link.
Sources
- Microsoft Research on interrupted work for attention-cost framing.
- Nielsen Norman Group on notification design for alert control and clarity principles.
- Supers for personal AI agent workflows across SMS, browser work, and generated outputs.
FAQ
Should every correction become a rule?
No. One-off taste feedback can stay as a note. Repeated or high-cost failures deserve rules.
Where should rules live?
In the system prompt, policy layer, or correction log, depending on how the agent is deployed.
How soon should rules be reviewed?
Review weekly, or immediately after any correction involving money, people waiting, or irreversible actions.
What proves a rule worked?
The next similar event is handled differently, with the right source proof and channel choice.
Make corrections operational instead of conversational.
Supers can help teams test personal agents that text, browse, and build while carrying corrected behavior into the next real workflow.