Kartini Cooper
← Back to portfolio
// Playbook 02 · 2026
Playbook · Memory and context

Stop re-explaining yourself.

Most people use Claude as if every chat were the first. They paste the same role, repeat the same style rules, re-upload the same reference docs. Four layers exist for a reason — chat, Style, Project, Memory — each with a different scope of what Claude knows. Set them up once, get the leverage every chat.

!
// A note on currency Feature details below are current as of May 2026. Claude is moving fast — Memory shipped to all plans (free included) in March 2026, Projects has had multiple capability bumps in the past year. The principles in this playbook don't change. The settings paths and free-tier limits might. Verify the specifics in your own Claude.ai before you set up.
// Who it's for
Anyone using Claude for ongoing work
// Time investment
30 minutes to set up · saves hours/week
// You'll need
A Claude.ai account · free is fine
// You'll walk away with
A working configured Claude
01 —
// The mental model

Four layers. Inside out.

Claude doesn't know one thing about you. It knows four things, at four different scopes. The innermost is what you typed this turn. The outermost is what it remembers about you across every conversation. Each layer answers a different question — and belongs to a different kind of information.

// LAYER 01
The chat
SCOPE — this conversation
"What did I just say?"
// LAYER 02
The Style
SCOPE — every chat using it
"How should you behave?"
// LAYER 03
The Project
SCOPE — every chat in it
"What do you know about my work?"
// LAYER 04
The Memory
SCOPE — across everything
"What do you know about me?"
// Smallest scope · this turn
// Broadest scope · across all chats
1
Layer 01 · The smallest scope

The chat.

Scope · this conversation

The chat is what you've said in this conversation, plus what Claude has said back. It resets when the chat ends. This is the layer everyone uses by default — and the layer everyone over-uses, because they don't know the other three exist.

// What belongs here vs. doesn't
// Goes in the chat
  • The task — what you want done right now
  • The input — the document, data, draft you're working with
  • Task-specific length and format — "in 200 words", "as a table"
  • One-off context — the specific audience for this output
// Doesn't belong here
  • Your role and standing context — that's Layer 2 or 4
  • Your writing style preferences — that's Layer 2
  • Reference documents you reuse — that's Layer 3
  • Facts about you Claude should remember — that's Layer 4
// Working example · the unconfigured prompt~200 words
You are a senior advisor preparing a board member for a meeting. I work at a financial services firm in Melbourne, Australia. We use Australian English, our house style is crisp, declarative sentences, no jargon, no "leverage" or "synergy". Read the attached document. Summarise it in a way the board member can absorb in two minutes. Lead with the decision being asked for, if any. Ignore detailed methodology — focus on findings and implications. 500 words maximum. Format as three short sections — Decision, Findings, Risks. If a field isn't clearly stated, write "Not stated" — don't guess. The board member is chair of the risk committee, so flag anything risk-related explicitly. Document attached.
i
The pattern to notice

About 80% of that prompt is the same thing you'd type next time. The role, the firm context, the style rules, the spelling, the "don't guess" rule — they're all standing context. Only the task, the document, and the specific audience are genuinely one-off. That's the case for moving the rest into Layers 2, 3, and 4.

2
Layer 02 · How Claude behaves

The Style.

Scope · every chat using it

A Style is Claude's standing instructions about how to behave — voice, tone, spelling, what to always include, what to never include. In Claude.ai, you'll find Styles in the chat composer; you can pick a built-in one or create your own. Set once, applies to every conversation using it. This is where standing rules belong, not standing facts.

// What belongs here vs. doesn't
// Goes in the Style
  • Voice rules — short sentences, fragments where they land, active voice
  • Spelling conventions — Australian English, no Americanisms
  • What to always do — flag uncertainty, lead with the headline
  • What to never do — no "leverage", no closing summaries, no hedging
  • How to handle missing information — say so, don't guess
// Doesn't belong here
  • Facts about specific projects — that's Layer 3
  • Reference documents — that's Layer 3
  • Personal facts about you — that's Layer 4
  • Task-specific instructions — that's Layer 1
// How to set it up
01
In the chat composer, open the Style picker.
Below the message input, there's a Style selector (defaults to "Normal"). Click it.
02
Create a custom Style.
Name it something specific (Kartini · House style). Paste your voice and behaviour rules. Aim for under 300 words — long Styles get skimmed.
03
Set it as your default.
In Style settings, mark it as the default so every new chat starts with it loaded. Switch back to Normal only when you specifically need a different voice.
!
// The anti-pattern

Stuffing project context into the Style. "I work at Acme Corp on the supply chain transformation" doesn't belong in a Style — it belongs in a Project (Layer 3). If you put it here, every chat thinks it's about Acme, including the ones that aren't. Keep Styles about behaviour, not subject matter.

3
Layer 03 · What Claude knows about your work

The Project.

Scope · every chat in it

A Project is a workspace with two persistent layers of its own: instructions (a text block applied to every chat in the Project) and knowledge (files uploaded once and referenced by every chat in the Project). Free plans get up to 5 Projects; paid plans get unlimited plus a larger knowledge base. This is where the highest leverage usually lives.

// What belongs here vs. doesn't
// Goes in a Project
  • Instructions specific to this work — "this Project is for board paper summaries"
  • Style guides and voice docs — your firm's, your team's, your own
  • Reference templates — board paper formats, briefing structures
  • Examples of good output — three or four real ones you're proud of
  • Policy or process docs you cite repeatedly
// Doesn't belong here
  • Confidential client files — those stay in a single chat
  • One-off documents — upload to the chat itself, not the Project
  • Personal preferences about voice — that's Layer 2
  • Anything you wouldn't want every chat in this Project to see
// How to set it up
01
Create a Project, named for the actual work.
Projects → + New Project. "Board paper summaries" beats "Work 1". The more specific, the easier to manage when you have a few.
02
Write the Project instructions.
200–500 words. Describe what this Project is for, who you are when working in it, and how Claude should approach the work. Different from your Style — this is about subject, not voice.
03
Upload reference documents to the knowledge base.
Templates, style guides, examples of past output you'd want Claude to match. Two or three good documents beat a 60-page manual. Don't upload anything you wouldn't want every chat in this Project to see.
04
Maintain it weekly.
If you find yourself re-prompting Claude for the same correction twice, add it to the instructions. Five minutes a week. Compounds fast.
i
Insight from the field

The Project knowledge base is where most individuals get exponential returns. A Style fixes voice — a Project fixes understanding. Three real examples of board papers you've written, uploaded once, do more for output quality than any amount of instruction writing. Show, don't tell, scales here.

4
Layer 04 · What Claude knows about you

The Memory.

Scope · across everything

Memory is the newest layer. When enabled, Claude extracts and stores facts about you across conversations — your role, your business, your working preferences, ongoing context — and uses them to inform future chats. Available to free and paid plans since March 2026. Toggle in Settings → Capabilities → Memory. You can view, edit, or delete stored memories anytime.

// What belongs here vs. doesn't
// Belongs in Memory
  • Who you are — name, role, profession, city
  • Ongoing context — current job, current programme of work
  • Working preferences — how you like to work, when you're available
  • Standing facts — your team, your stakeholders, your tools
// Doesn't belong here
  • Confidential information — client identifiers, financial details, HR data
  • Anything covered by a privilege or compliance obligation
  • Project-specific detail — that's Layer 3
  • Style preferences about voice — that's Layer 2
  • Anything you'd be uncomfortable seeing in a future chat
// How to set it up
01
Turn it on.
Settings → Capabilities → Memory. Toggle to on. Free and paid plans both have access as of May 2026.
02
Seed it explicitly, don't wait for inference.
Open a new chat and tell Claude the standing facts directly: "For future chats, remember: I'm Kartini, a transformation lead in Melbourne. I work in financial services. I write in Australian English." Explicit beats inferred — every time.
03
Audit it monthly.
Settings → Capabilities → Memory shows everything Claude has stored. Read through. Delete anything stale, wrong, or no longer relevant. Memories compound — pruning is how you keep them useful.
04
Use temporary chat for confidential work.
Anything you don't want remembered — client matters, sensitive conversations, exploratory work — open as a temporary chat. It stays out of Memory and out of search.
!
// The anti-pattern

Treating Memory as set-and-forget. Memory drifts. The job you had six months ago is still in there. The "current project" you finished last quarter is still in there. The preference you've since changed your mind on is still in there. Without the monthly audit, Memory becomes a museum of past you — and Claude responds to past you, not current you.

i
A note on confidentiality

Memory is convenient. It's also persistent. For anyone with professional confidentiality obligations — legal, medical, financial, HR — be deliberate about what you let Memory hold. The default isn't malicious, but the default also isn't aware of your obligations. Temporary chat exists for a reason; use it.

02 —
// The payoff

From 200 words to 20. Same output.

Watch the same task — summarising a board paper — across the four layers. The chat shrinks every time. Everything in strikethrough moved to a deeper layer. Everything in orange is the new minimum needed.

// Stage 1 · Nothing configured~200 words
You are a senior advisor preparing a board member for a meeting. I work at a financial services firm in Melbourne, Australia. Use Australian English, crisp declarative sentences, no jargon, no "leverage" or "synergy". Read the attached document. Summarise it in a way the board member can absorb in two minutes. Lead with the decision being asked for. Ignore methodology. 500 words max. Format: three sections — Decision, Findings, Risks. If a field isn't clearly stated, write "Not stated" — don't guess. The board member is the chair of the risk committee. Flag anything risk-related explicitly. Document attached.
// Stage 2 · Style configured (voice + spelling rules moved out)~140 words
You are a senior advisor preparing a board member for a meeting. I work at a financial services firm in Melbourne, Australia. Use Australian English, crisp declarative sentences, no jargon, no "leverage" or "synergy". Read the attached document. Summarise it in a way the board member can absorb in two minutes. Lead with the decision being asked for. Ignore methodology. 500 words max. Format: three sections — Decision, Findings, Risks. If a field isn't clearly stated, write "Not stated" — don't guess. The board member is the chair of the risk committee. Flag anything risk-related explicitly. Document attached.
// Stage 3 · Project configured (work context + format rules moved out)~60 words
You are a senior advisor preparing a board member for a meeting. I work at a financial services firm in Melbourne, Australia. Use Australian English, crisp declarative sentences, no jargon, no "leverage" or "synergy". Read the attached document. Summarise it in a way the board member can absorb in two minutes. Lead with the decision being asked for. Ignore methodology. 500 words max. Format: three sections — Decision, Findings, Risks. If a field isn't clearly stated, write "Not stated" — don't guess. The board member is the chair of the risk committee. Flag anything risk-related explicitly. Document attached.
// Stage 4 · Memory configured (role + city moved out)~20 words
Summarise this for the chair of the risk committee. Flag anything risk-related explicitly. Document attached.
The work is in the setup. Then every chat is twenty words.
03 —
// Common pitfalls

Four ways the layers collapse.

The four-layer model only works if each layer does its own job. The most common failures are layer-confusion — putting the right information in the wrong place — and layer-neglect, leaving a layer empty when it would do real work. These four are the ones I see most.

// Pitfall 01

The kitchen-sink Style.

You discover Styles, get excited, and pour everything into one — voice rules, project context, role, working preferences, the lot. The Style hits 800 words. Claude starts ignoring half of it.

The fix Styles are about behaviour, not facts. Keep them under 300 words. Move project context to Layer 3 and personal facts to Layer 4. A short Style that's followed beats a long Style that's skimmed.
// Pitfall 02

The everything Project.

One Project named "Work" with 40 files in the knowledge base. Half the files contradict each other. Every chat inherits all of them and Claude can't tell which is canonical.

The fix One Project per kind of work. Three or four focused Projects beat one Project that does everything. The knowledge base is for reference, not archive — fewer, better documents.
// Pitfall 03

The museum Memory.

Memory is on, but never reviewed. It still thinks you work where you worked a year ago. It still references projects that wrapped six months back. Every chat starts with subtly wrong context.

The fix Audit Memory monthly. Settings → Capabilities → Memory. Delete stale entries, correct wrong ones. Fifteen minutes. The cost of not doing it is Claude responding to a version of you that no longer exists.
// Pitfall 04

The confidentiality blur.

You upload a sensitive client file to a Project. You let Memory store details from a confidential conversation. The information persists across chats that shouldn't have access to it.

The fix Confidential work belongs in a single chat — ideally a temporary chat — not in any of the persistent layers. The persistent layers are for standing context, not specific cases. When in doubt, keep it ephemeral.
04 —
// Your decision guide

What to put where.

A quick question-to-layer mapping. When you catch yourself typing the same thing twice, ask which layer it belongs in. The answer is almost always not Layer 1.

// Question
Is this about how Claude should write?
// Layer 2 — Style
Voice, tone, spelling, what to always do, what to never do. Behaviour, not subject.
// Question
Is this about what Claude is working on?
// Layer 3 — Project
Subject context, templates, examples, reference documents. One Project per kind of work.
// Question
Is this about who I am?
// Layer 4 — Memory
Role, city, business, working preferences. Standing facts about you, not your work.
// Question
Is this just for this task?
// Layer 1 — Chat
The specific task, the input document, the one-off audience. If you'll never repeat it, it stays here.
// Question
Is this confidential?
// Layer 1 — Temporary chat
Open a temporary chat. Out of Memory. Out of search. Ephemeral on purpose.

Using this in practice?

Most people who say "Claude is useful but not as useful as I'd hoped" are running it on Layer 1 alone. The other three layers are where the leverage lives. Half an hour setting them up properly pays back in the first week — and keeps paying after that.

This playbook pairs naturally with Playbook 01 (Prompting) — Pattern 05 there pointed at the system prompt, this is the proper unpack. It also feeds into Playbook 03 (Skills): once you've found instructions worth keeping in a Style or Project, the next step is often turning them into a Skill that triggers automatically.

A reminder on the note up top: the features evolve, the principles don't. Claude.ai's settings menus and free-tier limits will keep shifting. The four layers — what you're saying now, how Claude should behave, what Claude knows about your work, what Claude knows about you — will keep being four layers. Build for the structure, adapt the specifics.