AI Playbook Undeniable Speaking Trainings About Blog Subscribe Work With Me
← All articles
AI Enablement

How to Make AI Learn From Its Mistakes (The Weekly Dream Report)

The Short Version
  • If you want to know how to make AI learn from its mistakes, give it a scheduled review: replay recent sessions, notice the repeat mistakes, and rewrite them into rules.
  • Rules only stick when they live somewhere your AI reads before every chat, like memory or project instructions, with a visible copy in your second brain.
  • Use one approval, two saves: every approved rule goes into memory and into a Notion page at the same time, and Notion wins if they ever disagree.
  • Treat the review like any new agent. Put it on an on-ramp, watch the cost, and make it earn its spot.

Your AI makes the same mistake on Tuesday that you corrected on Monday. You explain your preferences again. You fix the same formatting problem again. If you have been searching for how to make AI learn from its mistakes, the answer is a simple weekly habit: have your AI look back at its own work, name the patterns, and turn them into rules it reads every time it starts.

I tested this live on a recent group coaching call. We all ran the same prompt at the same time, set a four minute timer, and compared what came back. Some of the results were gold. Some came back almost empty. Both taught us something.

Can AI Learn From Its Mistakes?

Out of the box, every session starts at zero. Great work today, brand new tomorrow.

That is why the same mistakes show up twice and your corrections fade. I told my group about a writing problem that has been driving me nuts. AI does not understand paragraph spacing. I asked it to fix it, and it overcorrected by putting every sentence on its own line.

"You fix it Monday, but if you don't put it as a rule in your second brain or commit this to memory, it'll happen again."

So yes, AI can learn from its mistakes. It needs two things to do it. It needs a moment where it reviews what happened, and it needs a place to store the lesson that it reads automatically next time.

My rule for this is short. See it once, commit it as a rule.


What Is AI Dreaming?

Anthropic calls the concept dreaming. Between sessions, the AI reviews what it has done, sees the patterns, and rewrites its notes for next time.

I love this idea because it mirrors something I do every night. The last thing I do before I fall asleep is visualize my goals. Your mind keeps working while you sleep, so you might as well point it at what you want to create.

AI can do the same thing. The dream has three steps:

  1. Replay. Read back through the recent sessions.
  2. Notice. Find the repeat mistakes, the workflows that worked, and the preferences you showed through your corrections.
  3. Rewrite. Update the rules so tomorrow starts smarter.

How to Make AI Learn From Its Mistakes in Claude or ChatGPT?

This is the exact prompt my group ran together. Paste it into an existing project that already holds real work.

It's dream time. Let's look back at our conversations from the past seven days.
1. Replay: list what we worked on.
2. Notice: find three repeat mistakes, three workflows that worked, and three preferences I showed through my corrections and approvals.
3. Rewrite: turn it into three new rules for working with me and flag anything you know about me that's outdated or duplicated.
Give me a dream report. Lead with your top recommendations. Change nothing until I approve.

Where you run it matters more than the prompt. One member pasted it into a plain chat with no project behind it. The answer came back thin. It knew their name, their email, their company, and their time zone, and that was it.

"Your AI is mirroring back to you how good or a lack thereof your AI is."

Another member ran it inside a real project and got a lot back. The report showed what was working, where things were stuck, and that a few tasks it said it had done were still parked. It said it would correct that and set a new rule. It also caught that when they give a command, they want action right away.

That second result is the whole point. The mistakes you never noticed get surfaced, named, and turned into a rule.

If you are still building that foundation, start by learning how to set up Claude Projects for your business so the dream has real context to work with.


How Do I Make Claude Remember My Rules in Every Chat?

A sharp question came up right after the exercise. If Claude forgets things, how does it remember the dream report a week from now?

I asked Claude live on the call, and the answer was clear. Claude stays current through the layers it loads at the start of every chat:

  • Memory loads in every chat. Best for rules, preferences, and corrections.
  • Project instructions load in every chat inside that project. This is your standing playbook.
  • Project files get searched when they are relevant.
  • Your Notion second brain only gets read when you send Claude there.

Notion is a library Claude visits when you send it. Memory and project instructions are the notes Claude reads before every chat. This is also why it helps to understand whether Claude projects share memory, since each project carries its own context.

Memory has one weakness. It feels invisible. You have no easy way to see what is in there, whether it is old, or whether it is right.

So we landed on a simple system: one approval, two saves. When you approve a rule from the dream report, it goes into memory and into a Notion page called Claude Rules in the same step. You never update them separately. If the two ever disagree, Notion is the truth and memory gets corrected to match. Once a month, the dream compares the two. Each dream report can also open with a short snapshot of what memory holds right now, so you see it every week without asking.

This is exactly how I run my AI second brain in Notion. One of my standing rules tells my agents to lead with their recommendation and give me a decision card whenever they need me. Every new agent gets onboarded with that same rule. When the rule changes, it changes everywhere.


How to Make AI Learn From Its Mistakes Without Adding Too Many Rules?

My own first dream report taught me something fast. I did not read it. It was too wordy.

So without reading it, I asked for a visual report that was easy to read. The second version picked up little nuanced things. It flagged the facts about my business it needs to refresh before every chat, like how many podcast episodes I have and what my tagline is.

That one hit home. At one point AI started adding a tagline to all of my work that I never actually said. It picked it up somewhere, and I could not figure out how to reverse it. A conflicts report helps catch that kind of drift. Ask your AI where it sees one thing in one place and something different in another.

That is the next level of this practice. Newest truth wins. Your self-improvement loop has to prioritize new over old, because a second brain you have used for months or years is full of things you have moved past.

I think about it like pruning a rose. The rose blooms by losing the dead petals that are sapping its energy. What do you need to prune from your AI system?

Pruning works in both directions. Claude's own verdict on the dream report named the risks: more decisions for you, too many rules, and reports that run unread. Adding rule after rule can backfire, which is why I wrote about why your AI gets worse when you add more rules. The dream should clean up your rules as often as it adds new ones.

When the dream suggests a cleanup, validate and verify before you approve it. One member's report recommended retiring an old, duplicated version of their knowledge base. They hesitated, and they were right to. My advice was to speak that hesitation right back to the AI. The newest version is the one we want. Make sure nothing breaks. Then approve it.


How Often Should You Review Your AI's Mistakes?

Start weekly. Claude's recommendation for my setup was a small weekly cleanup with a four week test to prove it works.

Then hold it to the same standard as any new agent. Every new agent or routine I build goes through a seven day on-ramp. It has to get a green light every day for seven days before it graduates to the roster. I learned that the hard way when I tried to make my first agent do a million things at once.

The dream report gets the same treatment. "It can fit, but does it earn its spot?"

Watch the cost too. You are paying for the tokens. If a weekly dream report costs you $37, is it generating $37 worth of improvements?

My honest verdict after one live test is that this needs more reps. Dream more work. I see real value in it, and we are not romantic about it either. It earns its place by getting better results, or it goes.


How Do You Know If an AI Self-Improvement Loop Is Working?

One member has already built a loop of their own. Their email agent spots issues that keep coming up and suggests new agents to handle them. Their chief of staff agent reads their checklist at the end of each day and tells the other agents what they can take on.

That is a self-improvement loop. The next step is a watchdog, another agent watching the loop to make sure the improvement actually happens. An agent can find the pattern, tell you it is spinning up a fix, and you might realize later that it never actually did the thing.

Ask yourself two questions. What did I start that did not get completed? Is that my fault or the AI's fault? Both can be true. My own stale report exists because I realized I was the roadblock for everything.

Your Next Step: Start an AI Dream Report This Week

Run the dream prompt this week inside the project where you do your most important work.

  1. Paste the prompt and add "make it visual and easy to read."
  2. Read the three repeat mistakes first.
  3. Approve only the rules you would want every future chat to follow.
  4. Save each approved rule to memory and to your Claude Rules page in Notion in the same step.
  5. Run it again next week and compare.

That is how to make AI learn from its mistakes, one weekly dream at a time.

Your AI gets better the same way you do. It looks back honestly, keeps what works, and turns every lesson into a standard. See it once, commit it as a rule, and let your AI start every tomorrow smarter than today.

Undeniable Studio

Ready to build at the speed of your imagination?

Grow your business with AI, made simple and fun. Like having a Chief AI Officer by your side.

Weekly Live Building · Human-First AI