One study found that employees who spent fifteen minutes at the end of each working day writing down what they had learned outperformed their peers by twenty-three percent. The number is striking. The mechanism is straightforward: experience only becomes learning when you extract it deliberately. Most practitioners survive their sessions without extracting anything. The reflection loop closes that gap.
When you are working with AI, the loop has an additional function. The AI contributes output, framing, and suggestions throughout the session. Without a closing loop, it is easy to end a session uncertain about what was yours and what was the model's — uncertain about where your judgement operated and where you deferred. The loop makes that visible.
This is not journalling. It is a five-question protocol that takes twelve to fifteen minutes. You run it at the end of any significant build session with an AI model.
Workflow
The Five Questions
Run these in sequence. Write or speak your answers — voice notes work well and can be transcribed automatically into your log. A complete round takes twelve to fifteen minutes. Short answers are fine. The point is active reflection, not eloquence.
1. What was the original intent? Did we achieve it?
Before you assess the output, restate what you were trying to accomplish. Then assess: did you achieve it, partially achieve it, or arrive somewhere else? If you arrived somewhere else, was that better or worse?
This question catches the most common failure mode of AI-augmented work: goal drift. A session that started as "draft a project brief" can end as "review three different structural approaches" without the shift being noticed. Naming it is the first step to controlling it.
2. Where did the AI's output diverge from what you expected — and what did that reveal?
Divergence is data. When the model produced something you did not anticipate — a framing you had not considered, an argument you disagreed with, a structure that surprised you — note it. Then ask: what did that divergence tell you about the problem, about the model's assumptions, or about your own?
The most valuable learning in AI-augmented work often comes from the moments of friction, not the moments of fluency.
3. Where did you override the model — and were you right to?
Every time you rejected AI output and substituted your own judgement, that was an override. List the overrides and assess them honestly. Some will be justified by domain knowledge, personal voice, or creed-level constraints. Some will be reflex — instinctive rejection that was not better-grounded than the model's suggestion.
Tracking your overrides builds your honest map of where your judgement is calibrated and where it is not.
4. What do you know now that you did not know at the start of this session?
Not what you produced — what you learned. This might be something about the domain (a reference you followed, a concept that became clearer). It might be something about the model (how it interprets a particular type of brief). It might be something about your own working pattern.
If you cannot name anything, the session may have been productive without being developmental. Both happen. Knowing which is which is worth knowing.
5. What is the one thing you need to remember from this session?
Distil. The previous four questions are designed to produce material; this one asks you to select the most important item. Write it as a single sentence. This is the entry that goes into your error-pattern library or your project log. Everything else is context.
The Log Format
Your reflection log entries do not need to be elaborate. A single Markdown file or a section in your project log works well. The minimum viable entry:
## [Date] — [Session brief description]
**Intent achieved:** Yes / Partially / No — [one sentence]
**Key divergence:** [what surprised me and what it revealed]
**Override I'm most unsure about:** [what I changed and why]
**What I learned:** [one concrete thing]
**Remember:** [single sentence]
If you use voice notes, end each session with a spoken version of these five questions and ask your AI to transcribe and format the result into this structure. The instruction is:
"Transcribe the following voice note into a reflection log entry using this structure: Intent achieved, Key divergence, Override, What I learned, Remember. Voice note: [paste or dictate]."
When to Use It
Run the reflection loop at the end of any session where:
- You used AI for more than thirty minutes of substantive work
- You produced something that other people will act on
- You overrode the model in ways you were not certain about
- Something surprised you and you do not yet understand why
You do not need to run it after every AI interaction. A brief search or a quick reformatting does not warrant a loop. Substantive build sessions do.
The Compounding Effect
The loop compounds. The first ten entries look thin. By thirty entries, patterns emerge: the types of divergence that signal genuine insight, the overrides that keep proving wrong, the domains where your judgement is strong and the ones where it is not yet reliable.
This is the Learning Rate made operational. Not a number to estimate — a practice that raises it.
From Chapter 7: Your Tree — The Practitioner (Phil Rust, 2026). Part of the Practitioner companion resources.