hancrafted
Work with me (opens in a new tab)

Why AI has a weird attention span you must learn to manage

A long chat behaves like a long call: its attention is a limited window, so you manage what goes into it.

Hi. For the next 30 minutes, you're the AI. Pay attention!

Your attention span

Thought experiment

How attention works: you versus an AI.

Let's start light. You're only human.

So, how's your attention?

  1. 1Fine
  2. 2Sweating
  3. 3Fried
Your attention span

Thought experiment Page 1 of 2

The long phone call

Called to settle one thing. An hour later you remember the start, the end and one standout, and the one thing you called for is still open.

SisterCall ended · 1:21:40
  1. Let's sort out the holidays!
  2. Yes! But first, did you hear about Mum’s new phone?
The start stays sharp

Primacy effect (opens in a new tab): the first line sets the frame.

  1. No, what happened?
  2. She sent me forty photos of her thumb.
  3. Ha! Did you see the neighbours’ renovation?
  4. Don’t get me started. And my back hurts again.
The middle blurs

Serial-position effect (opens in a new tab): mum’s phone, the renovation, the back: said, not kept.

  1. And Pepper ate a whole birthday cake, candles included!
The standout survives

Von Restorff effect (opens in a new tab): the cake is unlike the rest of the call.

  1. Did you try that new recipe?
  1. Not yet. Oh, dinner is ready, I have to run.
  2. Talk soon, bye!
The end stays sharp

Recency effect (opens in a new tab): the last minutes are still fresh.

…wait, we never fixed the date.

The pattern is measured on word lists in the lab; that it holds for an hour-long call is an inference 

Thought experiment Page 2 of 2

AI forgets the middle too

A long AI chat has the same shape as the call: the very first instructions and the newest messages get used, and what you decided after them, your first message included, can get lost.

SisterCall · 1:21:40
Summer partyChat with AI · one hour
  1. Let's sort out the holidays!
  2. Yes! But first, did you hear about Mum’s new phone?
The start
  1. Plan a summer party for our team: 40 people, end of June.
  2. Happy to help! Indoor or outdoor? Any dietary needs?
  1. Let’s cap the holidays at 2,000 euro.
  2. Deal! She sent me forty photos of her thumb.
  3. Ha! And my back hurts again.
The middle
  1. Let’s keep the budget at 40,000 euro.
  2. Noted! Three venues in Lisbon: Quinta terrace, LX Factory rooftop, Estrela pavilion.
  3. Fetched 10 pages · 15k of extracts
  1. Right, the holidays! What’s our budget?
  2. We settled that an hour ago!
The end
  1. Great, put the full plan together.
  2. Here is the full plan: venue 28,000 €, catering 16,000 €, band 8,000 €. Total 52,000 €.

With your sister, a lost detail is funny. With a deadline, it’s expensive.

Measured by planting a fact in a long input, not in today’s chat apps  · how deep the middle dips varies 

The AI's attention span

Now the cause: how the AI holds a chat.

Dialling it up. I hope you are ready?

Still with me? How's your attention?

  1. 1Fine
  2. 2Sweating
  3. 3Fried
Your attention span

The AI's attention span Page 1 of 4

The context window

The whole chat is handed over with every message, and the space is limited: the context window.

Conversation
Hover or drag the timeline to replay
Plan a summer party for our team: 40 people, end of June, budget 40,000 euro.Faded: the AI may fill this gap itself
Happy to help! Indoor or outdoor? Any dietary needs I should know about?
Outdoor if we can. 12 guests are vegetarian, please keep that in mind.
Noted: 12 vegetarian guests. Shall I look for venues?
Find three venues in Lisbon.
Three options: Quinta terrace, LX Factory rooftop, Estrela garden pavilion.

Step 9 of 21: The AI answers in three lines, +18 tokens, 26% used.

  • With every message

    The AI is handed the whole chat again , its own replies included .

  • The limit

    The space for it is limited. That limit is the context window.

  • In chat apps

    Apps may manage the window their own way. 

    illustrative numbers

The AI's attention span Page 2 of 4

What happens when the context fills up

As the chat fills the window, only the very beginning and the newest stay sharp, the rest blurs, your own first message included, one repeated detail survives, and answers get worse: context rot. 

Conversation
Hover or drag the timeline to replay
Plan a summer party for our team: 40 people, end of June, budget 40,000 euro.Faded: the AI may fill this gap itself
Happy to help! Indoor or outdoor? Any dietary needs I should know about?
Outdoor if we can. 12 guests are vegetarian, please keep that in mind.
Noted: 12 vegetarian guests. Shall I look for venues?
Find three venues in Lisbon.
Three options: Quinta terrace, LX Factory rooftop, Estrela garden pavilion.
The catering offers are in the project folder. Which one fits?
Sabor Verde fits best: it includes a vegetarian menu for 12.
Check hotels near the venue for 15 guests from abroad.
Hotel Lumen and Casa do Rio both have 15 rooms free on 27 June.
The guest list is in there too.
Got it: 40 guests, 15 travelling, 12 vegetarian.
Great, put the full plan together.
Full plan: venue 14,800, catering with the vegetarian menu 21,600, hotels 9,400, music 6,500. Total 52,300 euro.
That's 12,300 over our budget. Also check the venue contract.
Conversation compacted
Context window (22%)
44.1K of 200K
Limit 200k
Attention
illustrative
Step 24 of 28 · 100 units of attention

Only the very first lines and the newest messages stay sharp. Your first message and the rest spread thin.

Step 24 of 28: Conversation compacted, −141,045 tokens, 22% used.

Your first message

… budget 40,000 euro.

the AI may fill this gap itself 
The AI's plan

… Total 52,300 euro.

the budget got forgotten

Your first message already falls off, and it is one short number among many similar ones 

Studies use planted facts, not chat apps 

The AI's attention span Page 3 of 4

Attention is a budget

A hundred units of attention go mostly to the very first lines and the newest messages; everything between, your own first message included, gets what is left, until it is too thin to trust.

Context window (64%)
128K of
Smart zone by Matt Pocock 80K
Limit 200k
Attention
illustrative
Step 21 of 21 · 100 units of attention

Only the very first lines and the newest messages stay sharp. Your first message and the rest spread thin.

Past Matt Pocock's smart zone

Step 21 of 21: The AI's plan: 52,300 euro, +30 tokens, 64% used, past Matt Pocock's smart zone.

Known as

“Lost in the middle” 

The very first lines and the newest messages get the attention ; what sits in between, your first message too, fades .

Where the middle gets too thin differs per model and task. illustrative numbers

The AI's attention span Page 4 of 4

What we learned so far

Long AI chats show the same effects as long calls, because the AI's attention span is its context window: a limited resource you manage actively.

Human attention

  1. Let's sort out the holidays!
  2. Yes! But first, did you hear about Mum's new phone?
  3. No, what happened?
  4. She sent me forty photos of her thumb.
  5. Ha! Did you see the neighbours' renovation?
  6. Don't get me started. And my back hurts again.
  7. And Pepper ate a whole birthday cake, candles included!
  8. Did you try that new recipe?
  9. Not yet. Oh, dinner is ready, I have to run.
  10. Talk soon, bye!

Our memory follows the primacy and recency effects: we recall the start, the end and the standout.

the same shapean analogy, not a proof 

AI attention (context window)

Context window (64%)
128K of 200K
Context category
Messages91.9k (46.0%)
Free space55.5k (27.7%)Autocompact buffer16.5k (8.3%)

Step 21 of 21: The AI's plan: 52,300 euro, +30 tokens, 64% used, past Matt Pocock's smart zone.

  • Important information and decisions get lost in the middle.
  • When the window is full, auto-compaction can drop details without telling you.

Learn to manage attention: optional for you, mandatory for AI.

What fills the window

Your messages are only a part of it.

How's the brain? Anything fried yet? I eat 200K for breakfast.

Quick check: how full are you?

  1. 1Fine
  2. 2Sweating
  3. 3Fried
Your attention span

What fills the window Page 1 of 3

What's really in the context window

The chat is only the part you see. Before your first word, the context window already holds instructions and notes about you.

Token counts illustrative
Context window (64%)
128K of 200K
Smart zone by Matt Pocock 80K
Limit 200k
Context category
Free space55.5k · 28%Kept free for summaries16.5k · 8%
ExplainSimple
What fills the window

Starting instructions

34.1ktokens
17.1%of the window
1step

Before you type anything, the app hands the AI hidden instructions: who it is, how to answer, which tools it may use. They sit at the very top and travel with every message.

Step 21 of 21: The AI's plan: 52,300 euro, +30 tokens, 64% used, past Matt Pocock's smart zone.

Starting instructions

Text placed before your first message, before you type anything. 

The longer they get, the more room they take and the less closely they are followed. 

Memory

Short notes about you, saved earlier and put back in front of the AI, so a new chat already knows you. 

Claude and Gemini keep memory across chats too. 

Named today, not explained

  • Files the AI read
  • Web pages
  • Results from tools and connectors
  • System reminders

Past chats are fetched only when asked, not loaded up front .

What fills the window Page 2 of 3

One innocent question

One line in the chat. Twenty pages in the window.

Token counts illustrative 
Context window (54%)
107.8K of 200K
Context category
Messages71.7k (35.8%)
Free space75.7k (37.9%)Autocompact buffer16.5k (8.3%)
ExplainTechnical
Step 11 · pinned

The AI read catering-offers.pdf

200Tool calls
55kTool results

Glob finds the file; Read takes PDFs over 10 pages in ranges of at most 20, as page text plus page images, about 2,700 tokens a page:

Read("offers/catering-offers.pdf", pages="1-20") → 55,000 tokens

Step 12 of 21: The AI picks a caterer, +15 tokens, 54% used, past Matt Pocock's smart zone.

One line in the chat, the whole file behind it

To answer, the AI read all 20 pages of the catering PDF, not shown in the chat .

The two web searches before it brought back extracts, a modest bite . Both stay in the window for every later question .

Same scale, tokens

  • Your question15
  • Two web searches15k
  • The catering PDF55k
  • The answer15

What fills the window Page 3 of 3

The chat looks the same. The window doesn't.

Left, the full window; right, the same chat after auto-compaction. On screen it is still one ongoing chat.

A moment before · step 23

Context window
Context category
Free space0 · 0%Kept free for summaries14.9k · 7%

Step 23 of 28: The AI read the contract, +57,120 tokens, 93% used.

The fill reaches the space kept free for summaries.

The chat, before and after: the same 

Conversation
Plan a summer party for our team: 40 people, end of June, budget 40,000 euro.Faded: the AI may fill this gap itself
Happy to help! Indoor or outdoor? Any dietary needs I should know about?
Outdoor if we can. 12 guests are vegetarian, please keep that in mind.
Noted: 12 vegetarian guests. Shall I look for venues?
Find three venues in Lisbon.
Three options: Quinta terrace, LX Factory rooftop, Estrela garden pavilion.
The catering offers are in the project folder. Which one fits?
Sabor Verde fits best: it includes a vegetarian menu for 12.
Check hotels near the venue for 15 guests from abroad.
Hotel Lumen and Casa do Rio both have 15 rooms free on 27 June.
The guest list is in there too.
Got it: 40 guests, 15 travelling, 12 vegetarian.
Great, put the full plan together.
Full plan: venue 14,800, catering with the vegetarian menu 21,600, hotels 9,400, music 6,500. Total 52,300 euro.
That's 12,300 over our budget. Also check the venue contract.
Conversation compacted

Step 24 of 28: Conversation compacted, −141,045 tokens, 22% used.

Right after · step 24

Context window
Context category
Free space139.4k · 70%Kept free for summaries16.5k · 8%

Step 24 of 28: Conversation compacted, −141,045 tokens, 22% used.

One block in place of every message ; details can drop out 

No notice in the chat; at most a short "organizing its thoughts" Documented for Claude; ChatGPT, Gemini and Copilot don't say 

Most chat apps never show you how full the window is.

Claude summarises older messages in the background ; the others document nothing about it .

Managing the attention span

You can't read a gauge, so you build habits.

Almost there. One more thing, then I compact you.

Last check: how's your attention?

  1. 1Fine
  2. 2Sweating
  3. 3Fried
Your attention span

Managing the attention span Page 1 of 4

Everyday habits

Split the project into jobs, keep each chat to one job, and carry only a short handoff into the next one, so the new chat starts small and sharp.

Old chat: summer party

Old chat128K of 200K · 64%

Handoff card

Party for 40, 27 June, budget 40,000 €. Venue: Estrela pavilion. Caterer: Sabor Verde (12 vegetarian). Next: invitations.

first message of the new chat

Fresh chat: invitations

Fresh chat38K of 200K · 19%
  1. Break it down

    Split the project into jobs before the first chat.

  2. New job, fresh chat

    When the topic changes, open a new chat.

  3. Ask for a handoff

    After real progress or before a break: the goal, the decisions, the next step.

  4. Read it, correct it

    Fix what is wrong before you carry it on.

  5. Start the next chat with it

    Paste it in as the first message.

  6. Check the context if you can

    In agentic tools, run /context regularly.

Token counts are illustrative · 200K window

Managing the attention span Page 2 of 4

The workflow for bigger projects

Break the project down first, keep one chat as the planner, and give each step a fresh chat that hands back one small result.

One chat for everything

Summer party128K of 200K · 64%

Past the smart zone; every step carries every other step's reading.

A workflow: one planner, a fresh chat per step

Planner45.3K of 200K · 23%
Worker: venues51.8K of 200K · 26% · thrown away
Worker: catering90.9K of 200K · 45%
Worker: invites36.1K of 200K · 18%

One small result back per step. Short tasks beat one long one ; you coordinate, like a project manager.

Token counts are illustrative · 200K window

Managing the attention span Page 3 of 4

Human in the loop workflow

You are the captain: a planner chat prepares each handoff, worker chats do the jobs, and you check every artifact in between.

  1. Plan a summer party

    Break it down
    1. Travel
    2. Catering
    3. Venue
    4. Invitations
    5. Budget
  2. Interview

    Interview yourself with AI
    1. AI asks you3 questions,
      1 you forgot
  3. Requirements

    Requirements first
    1. Gate 1

      Requirements card

      • 200 guests
      • €150 per person
      • Step-free
      • Vegetarian menu
      • Quiet to talk

      Signed off

  4. Research

    Fresh start for each piece
    1. Venue 1Checked against the card
    2. Venue 2Checked against the card
    3. Venue 3Checked against the card
  5. Compare

    Evaluate each step
    1. Rank 3 venuesagainst the cardGate 2
  6. Recommendation

    for your boss
    1. Venue plan

      by Friday

A fresh chat for each job  keeps every window small; you coordinate, like a project manager.

Managing the attention span Page 4 of 4

Congratulations

Now you know how it feels. One more time? Or again in a minute?

Your attention span