A learning rule in the context file turns every in-chat correction — formatting, terminology, assumptions, fiscal calendars — into a permanent lesson the next session inherits.
3 things to walk away with
Learning is a rule
Add a "learn from corrections" instruction once; the agent self-updates from then on.
Correct mid-flow
You do not need to prep context up front — correct the agent when it gets something wrong.
Lessons persist
Corrections are written back to the context file, so future sessions never make the same mistake.
Learning Rules — coach the coach
The crown jewel of this video: add a single rule to the context file that says "when a user corrects something you generated — whether formatting, convention, assumptions, or terminology — treat that correction as a permanent lesson for this site and add it to this context file."
That sentence alone changes the system's character. The agent now treats every correction as a teaching moment, not a one-off fix. It interprets the user's words, extracts the underlying rule, and writes it back to its own context.
A learning rule is meta-context — context that tells the agent how to update its context. One sentence converts the system from static to self-improving.
The agent learns from corrections automatically. It does not — without the learning rule, corrections live only in that one chat. The rule is what makes the lesson permanent.
Correction in the Flow of Work — no need to prep up front
The demo asks "what contracts expire in Q1?" The agent assumes calendar Q1 (January–March) because that is the model's default. The user corrects: "our fiscal year starts in July, so Q1 is July–September." The agent re-runs the query with the new definition and saves the lesson to the context file.
The user did not need to prep the agent with the fiscal calendar in advance. They corrected when the issue appeared — and from then on, every session uses the right calendar.
Correction-driven learning lowers the bar to deploy an agent. You do not need a perfect context file on day one; you just need a learning rule and the willingness to correct mid-flow.
Correcting mid-flow is wasted effort because the agent forgets after the session. With the learning rule in place, the correction becomes permanent on the next save — the agent does not forget.
Permanent Lessons — corrections become context entries
After the fiscal-year correction, the agent updates the context file with a new section: "Learned corrections — Zava's fiscal year is July to June." The next chat session loads that section as part of the rules. Q1 means July–September forever, until someone edits the file again.
This is how the agent gets opinionated about a specific organization. Generic LLMs assume calendar quarters; this site's agent now knows your quarters. Specificity compounds with every correction.
Permanent lessons are how a generic AI becomes your team's AI. Each correction nudges the behavior closer to how your organization actually works.
The agent stores lessons in a hidden memory you cannot inspect. Lessons are written into the context markdown file in plain text — you can read them, edit them, or delete them like any other line.
Flashcards — Advanced
03 · Advanced
Learning Rules
tap to reveal →
A meta-instruction in the context file telling the agent to treat user corrections as permanent lessons and write them back to context. One sentence converts the system from static to self-improving.
← tap to flip back
03 · Advanced
Correction-as-Lesson
tap to reveal →
The behavior pattern enabled by learning rules. Correct the agent in chat; it extracts the underlying rule, saves it to context, and re-runs the query with the new knowledge.
← tap to flip back
03 · Advanced
Permanent Lessons
tap to reveal →
Corrections written back into the context file as plain-text entries. Survive across sessions, inspectable by anyone, editable like any other markdown — your team's agent gets smarter without code.