CHAPTER 04 / 7 MIN READ
From mistakes to procedures
A revised answer, a remembered lesson, and an executable skill are different things.
START HERE
An agent solves the same setup problem every morning. It improves during each run, then starts from scratch the next day. What exactly was learned?
Ask what is written back: a better answer, a verbal lesson, or a reusable procedure.
1.Revision changes the current artifact.
Self-Refine uses a cycle of generation, feedback, and refinement to improve an output. It is useful to think of the object being edited: the draft answer or solution. That does not necessarily change the next task’s harness.
The source of the critique matters. Feedback generated by the same model can be helpful, but it is not an independent proof of correctness.
2.Reflection changes the next attempt’s context.
Reflexion turns feedback from a trial into a verbal reflection held in episodic memory. A later attempt can condition on that lesson. The persisted object is language about what went wrong and how to improve.
A good reflection identifies a cause and an actionable adjustment. “Try harder” preserves little. “Check whether the output file exists before reporting completion” can alter the next trajectory.
3.A skill packages behaviour for reuse.
Voyager builds a library of executable skills in Minecraft. Successfully developed behaviours can be retrieved and composed into later activities. Instead of remembering only advice, the system keeps a procedure it can run.
These mechanisms live at different levels. Fix this answer; remember this lesson; reuse this operation. None of those descriptions implies an online update to the model’s parameters.
A recurring data import
BEFORE
Each run rediscovers that dates use day/month ordering.
AFTER
Revision fixes today’s output. Reflection records the convention. A skill implements a checked parser for next time.
Choose the persisted object that matches the failure. A note and a function have different guarantees and maintenance needs.
PAUSE & PREDICT
Which example is a reusable executable skill?
Reusing a procedure also reuses its assumptions. A successful skill in one environment may fail after a schema, permission, or goal changes.
GO TO THE SOURCE
Self-Refine2023 ↗Reflexion2023 ↗Voyager2023 ↗This chapter is an original teaching synthesis. Open each research note for source links, findings, and classification caveats.
TAKE IT INTO YOUR OWN WORK
Take one repeated correction from your work and write it three ways: as an edit, as a lesson, and as a procedure.