Meli AI Mentoring

1.10

They keep what they learn so the next run starts higher

Chapter Progress: Early Draft

Chapter Progress

Agent instructionsPaid book

Saving turns one good run into a higher starting line

Every serious interaction produces two things: the output you asked for and evidence about the system that produced it. Beginners use the output and discard the evidence. Power users capture the evidence and turn it into a reusable standard, brief, or example. The save stage of the compounding loop is where that capture happens: the distinction that made one run good gets named, written down once, and loaded by the next run instead of found again from scratch. You can spot that distinction yourself, and you can also ask the assistant to read back over the session and propose it, which is often faster.

Saving takes a minute that feels skippable, and that is exactly why most people skip it. The minute pays back fast, because a distinction you keep in a written standard never has to be re-explained from memory again. Judgment compounds when it is explicit enough to travel from one run to the next.

Picture what a single AI session leaves behind. As of 2026 the common version of this looks like typing into a chat model such as Claude Fable: you open it, describe what you wanted, correct it twice, and get something good. The interface keeps changing, to voice, to glasses that carry world knowledge, to earbuds, someday to a brain-computer interface and to systems more capable than any person, but the shape of the loss is the same in all of them. The output is in front of you, and the reason it turned out well is not. The specific instruction that fixed the second draft lives only in the conversation and in your head. Close the session and that reason is gone. Next week you start the same kind of task from a blank page and rebuild the same correction by hand.

This loss is not a discipline problem. It is how human memory is built. The part of your mind that holds a fresh idea while you work, what the psychologist Alan Baddeley spent decades mapping as working memory, holds only a few items at once and fades quickly. A distinction you discovered on Tuesday is not waiting for you on Friday unless you put it somewhere you can retrieve it. Telling yourself to try harder to remember does not change the size of the store. The save stage works because it stops relying on memory at all: it gives your judgment a permanent place to live outside your head.

A fading store is why the save stage cannot be willpower

Working memory is the small, short-lived where you hold what you are thinking about right now. Baddeley's research describes it as a system with a few separate parts and a sharply limited capacity, holding only a little for seconds unless you keep refreshing it. A separate line of work by Nelson Cowan put the usable capacity at around four chunks for many kinds of material, fewer when the material is unfamiliar. The exact count is not the point. The point is that the store is small and leaky by design, so a distinction you noticed in the middle of a session is competing with everything else you are holding, and it loses.

This is why no amount of resolve fixes the loss. You are not forgetting because you are careless. You are forgetting because the store was never built to hold a week of fine-grained corrections. Writing the distinction down moves it from a store that fades into one that does not, which is the whole mechanism behind the save stage. You can skip this box and lose nothing from the practice: just write down the keeper before you close the session.

You can keep three kinds of thing, and naming them keeps the save small. Not everything from a session is worth storing, and a library that tries to keep all of it becomes a pile nobody opens. The keepers sort into three families, and almost everything reusable is one of them:

Standards: what good output for this task must clear. A standard is a rule the next output has to pass, like name a source for every claim, or lead with the decision and put the background after it.
Briefs: the context and constraints a task needs supplied up front. A brief is the short description of your situation that the model would otherwise make you re-explain every time, like who the audience is, what you are optimizing for, and what is off-limits.
Examples: outputs worth imitating, kept with one line on why they worked. An example is a finished piece you point the model at and say, match the tone and shape of this one.

These three families are also the shape of the library you are about to start. You now have a felt picture of it: a small store, sorted into standards, briefs, and examples, that any session can load. Here is the name for it.

Save the distinction, not the transcript

A saved chat log is a record of one conversation; a saved distinction is the lesson pulled out of it. Keep the whole transcript and you still have to reread it later and reconstruct what mattered, which is the same work you are trying to avoid. Keep the distinction and the lesson is stated plainly on its own: name a source for every claim, lead with the decision. The distinction is what the next run can load and apply without you back in the loop reading old conversations.

Keep the library separate from the task so any task can reach it

A standard buried inside one project's notes helps that project and dies with it. When your standards live inside the folder for one client email, they are out of reach the day you start a research summary. When they live in one place of their own, every task can pull from them. Separation is what lets a standard you learned writing a client email sharpen a research summary next month, and it is what lets the library outlast any single project.

Stay tool-neutral so the library outlives the tool

Tie the library to the principle, not to today's app. As of 2026 a library can be a folder of text files, a notes app, or a saved your AI loads on every session. The interface will change, the way the chat box is already becoming voice and glasses and earbuds. The principle of one retrievable place for your standards, briefs, and examples will not. Pick whatever you will reliably open, and treat the specific layout as one example of the idea rather than the idea itself, so a tool change becomes a move rather than a rebuild.

Use enough system to protect the work, not so much that the system becomes the work. A one-off task you will never repeat does not earn a saved standard. The save stage is for the distinctions that recur, the ones you have re-explained twice and would re-explain a third time. Saving everything turns the library into a second job; saving the recurring distinction keeps it small and worth opening.

Let the assistant pull the keepers out of a session for you

You do not have to spot the reusable distinction by hand. When the session is available, the same assistant that produced the output can often read back over it and propose what is worth keeping. After a run goes well, ask it to review the conversation, name the standards, briefs, and examples that reproduce the result, and phrase each one so a future session could load it cold. This is the save stage done by the system: it proposes, you review and approve, and the keepers land in your library in minutes instead of after a careful reread.

This step also collects what two earlier moves in the loop already produced. Rejecting AI gives you standards worth saving, because each named rejection is a distinction your output has to clear. Reverse , asking the model to extract the behind a result you admire, gives you briefs and examples worth saving. The save stage is where those keepers stop living in one conversation and start living in a place every conversation can reach.

Watch the save stage decide whether your setup compounds

Take a recurring example: building your own personalized setup, software shaped to how you live and work. You spend an afternoon with an AI getting it to draft the pieces you want, including how your day gets summarized into a short morning brief, which inputs it should pull from, and what it should never touch without asking. By the end it fits you well, because you corrected it a dozen times, and each correction taught it something true about your life.

The don't-apply move is to rebuild that setup from memory. You keep the output, the working setup, and discard the evidence, the dozen corrections that made it fit. Three months later a more capable model arrives, or your needs shift, and you sit down to redo it. You remember that it worked, but not exactly why, so you re-explain your travel schedule, your approval rules, and your tone preferences from scratch, and you rediscover the same corrections one frustrating round at a time. The setup never starts higher than it did on the first day.

The do-apply move is to keep a small library as you go. The correction never schedule anything without asking me first becomes a saved standard. The description of how you really live, traveling on no fixed schedule, thinking in writing, becomes a saved brief. The one morning brief that nailed your tone becomes a saved example. When the more capable model arrives, you hand it the library, and the new setup starts from everything the old one learned. The output is temporary; the system compounds. The library is the thread that carries your judgment from one version to the next, across every change of interface and every change of model.

When this method goes wrong, it usually goes wrong by hoarding. A library that saves every transcript and every draft becomes a pile you stop opening, and an unopened library compounds nothing. The fix is the same discipline from the start of the subchapter: save the named distinction rather than the raw session, and let the loop add only what you reach for again. A small library you open and load beats a large one you avoid.

This subchapter teaches saving as the save stage of the loop: the principle, the minimum library, and the one habit that keeps each run from resetting to zero. The reusable AI assets chapter goes deeper into the mechanics, including how to organize a full library by the job each asset does and how to bundle your standards, briefs, and examples with context and instruction files into a that starts every session already briefed. Start small here. Build it out there.

Pull the keepers from a session into your libraryPaid book · Claude reviews one recent AI session and proposes the standards, briefs, and examples worth saving, phrased so a future session can load them. You approve what goes into your library.