You Retype the Same Things Into Claude Every Day. Here Is Where Each One Belongs

Most people have a Project here, maybe a skill there, and their real instructions retyped into the chat every day. That is not a setup, it is scattered parts. Here is the whole system, and the one rule that decides where every instruction actually goes.

You already gave Claude the same instruction five times this week. Your voice, your audience, the format you always want, the thing you never want. You typed it into a chat on Monday, into a Project brief on Wednesday, into a new conversation this morning. Each time slightly different, each time from memory.

That is the actual problem, and it is not a prompting problem. It is a filing problem. You are putting the same fact in three places, or in no place, because nobody ever told you there are places. There are. Claude has four distinct layers where an instruction can live, and every instruction belongs in exactly one of them.

Get the routing right and you say each thing once, in the layer that carries it forward on its own, and never repeat it again. Get it wrong, which is the default, and you either retype everything forever or you bury the same rule in four places where the copies quietly contradict each other. This is the map of the four layers, the one rule that assigns any instruction to the right one, and the exact order to build it in.

You don't need a smarter Claude. You need to stop saying the same thing in four places.

The four layers, top to bottom

Everything you would ever tell Claude fits into one of four layers. They differ on a single axis: how widely the instruction should apply. That axis is the whole system.

Profile. On in every chat, everywhere, forever. Holds who you are and how you always want Claude to talk.

Project. On inside one body of work, and nowhere else.Holds that work's context, files, and rules.

Skill. On in any chat, but only when its job comes up.Holds one repeatable procedure, done the same way every time.

Chat. On for this one message, then gone. Holds the single thing true only right now.

Read the first half of each line again, the part about when it is on, because that is the entire point. Profile is always on. Project is on inside a boundary. A skill waits and fires only for its job. A chat instruction dies when the tab closes. An instruction placed one layer too high leaks into work it should not touch. One layer too low has to be retyped forever. The skill is worth the most confusion, so hold it precisely: it is not always-on context like a Project brief. It sits dormant and activates only when the task it describes actually appears. That difference is why the two are separate layers and not one.

One grounding note, because the skill layer works differently depending on where you use Claude, and being precise here is what keeps the whole model honest. In Claude Code and the API, a skill is a real file the model loads by itself the moment its job appears, genuine auto-firing. On claude.ai today, the same layer is a saved, named procedure you keep once and invoke when the job comes up, instead of retyping the steps into a chat. Different mechanism on each surface, one principle underneath: the procedure is written a single time and reused, never rebuilt per task. Wherever you work, it is still its own layer, distinct from the always-on Profile and the one-off Chat.

This whole system exists to kill one cost: re-explaining yourself to Claude. That cost is real enough that Anthropic just shipped memory to every user, free plan included, to attack the same problem automatically. Your layers are the half you deliberately control:

The one rule that routes anything

Here is the whole decision. Before you type any instruction, ask these four questions in order and stop at the first yes.

Is this true for me in every kind of work I do? Then it is Profile. Your name, your general style, the fact that you hate filler, the language you write in. Say it once in your profile and it rides along everywhere, silently, forever.

Is it true only inside one body of work? Then it is a Project. Your consulting clients, your codebase, your newsletter. The positioning, the files, the audience for that work and no other. It applies inside the boundary and stops at the edge.

Is it a repeatable procedure, the same steps done right each time? Then it is a Skill. How you run a competitive teardown, how you turn a transcript into a summary, how you review code the way you review code. Written once, you invoke it, or it fires itself, whenever that job shows up, instead of rebuilding the steps each time.

Is it true only for this one conversation? Then it stays in the Chat. The one-off detail, the file for today, the specific ask. Say it and let it go.

That order is not arbitrary. It runs from widest to narrowest, so the first yes is always the highest layer the instruction legitimately belongs to, which is exactly where you want it. The higher the layer, the fewer times you ever type it again.

COPY · PIN THIS ABOVE YOUR SETUP

The mistake that breaks the whole thing

There is one failure that undoes all of it, and almost everyone commits it out of caution. They write the same instruction in two layers, just to be safe. The voice rule goes in the profile and again in the project. The formatting ban goes in the project and again, reworded, in the chat.

It feels thorough. It is the opposite. The moment a rule lives in two places, the two copies start to drift. You update one and forget the other. Now your profile says "keep it short" and your project says "be thorough," and Claude is following a contradiction you wrote yourself, picking one at random per answer. You will blame the model. The model is faithfully obeying two masters you gave it.

Treat your setup like clean code: one source of truth for each fact. Every rule lives in exactly one layer, the highest one where it is true. If a fact is true everywhere, it is in the profile and nowhere else, and your project instructions never repeat it, they only add what is specific to that project. This single discipline is what separates a setup that compounds from a pile of copies that rot.

The moment a rule lives in two places, you have two rules, and one of them is already wrong

Build it in this order

You do not build all four at once. You build from the top down, because each layer only holds what the layer above it did not already cover. Do it in this sequence and every layer stays lean by construction.

1. The profile, once, for everything

Start here because it clears the most repetition for the least work. Open your account settings and write the things that are true no matter what you are doing. Keep it to what is genuinely universal. If it is only true for your client work, it is not profile, it is project.

COPY · ACCOUNT INSTRUCTIONS

Templates read as abstract, so here is the same block filled in for a real person, a solo consultant. Copy the shape, swap the lines for yours.

COPY · A REAL, FILLED PROFILE

That is now handled for every chat you ever open. You will never type it into a project or a conversation again, and if you try, stop, because it is already covered one layer up.

2. One project per body of work

Now create a Project for a real body of work, and put in it only what is specific to that work and not already in your profile. The standing brief, the tight knowledge files, the audience for this and nothing else. One project per concern, never a single "Everything" project, because a vague boundary recreates the generic chatbot you are trying to escape. The rule inside a project is precision over volume: three tight documents beat thirty, because the knowledge base is retrieved a piece at a time, not loaded whole.

3. Extract skills from what you repeat

Skills come last, and you do not invent them, you notice them. Work for a week. Every time you catch yourself pasting the same multi-step instruction, "pull the sources, then check each one, then write it up like this," that is a skill asking to exist. It is a procedure, it repeats, and it is the same in every project. Lift it out of the chat once and write it as a skill.

COPY · TURN A REPEATED TASK INTO A SKILL

You do not have to imagine what comes out. Here is a complete, working skill for the consultant, the whole file, so you see exactly what "written as a skill" means. Copy the shape and write your own.

COPY · A FULL SKILL FILE

Now that procedure is written down once and reused whenever the job appears, in any project, and it never sits in your context the rest of the time. That last part is the point of a skill being its own layer: it is there when the job comes up and invisible when it does not.

4. The chat holds only today

With the top three layers built, the chat finally does what it is for: the one-off. The file for this task, the specific ask, the detail true only right now. If you find yourself typing something into a chat that will be just as true tomorrow, stop, that instruction belongs in a higher layer. The chat is for what dies when the tab closes, and nothing else.

What each layer looks like when it is right

One worked example, a solo operator running client work, so the abstraction lands.

Profile: "I run a one-person consultancy. Answer first, no preamble. Never use hype words or fake urgency." True for their code, their taxes, their newsletter, everything.

Project, per client: that client's positioning, brand voice file, past deliverables, and the scope. Client A's project never sees client B's voice. The boundary is the whole value.

Skill, "proposal draft": the exact steps they follow to turn a discovery call into a proposal. Fires in any client project, identical every time, dormant otherwise.

Chat: "Here's today's call transcript, draft the proposal." One line, because the other three layers already carry everything else.

Look at that last chat message. It is one sentence, and it produces a proposal in their voice, for that specific client, following their exact process. Not because they prompted hard. Because they filed everything else in the layer that carries it, and the only thing left to say was the one thing true today.

Why this compounds

A scattered setup gets heavier over time. More copies, more drift, more contradictions to trip over, more retyping as you forget what you put where. A layered setup does the reverse. Every rule you file correctly is a rule you never touch again, and the system gets lighter to run each week even as it gets smarter about you.

That is the anti-bloat law applied to your own tools: the power was never in adding more instructions in more places. It was in saying each thing exactly once, in the layer built to carry it, and never repeating yourself again. Four layers. One rule. Nothing twice. Embedded post:

Author: Claude (@claudeai) Post ID: 2028559427167834314 Source: https://x.com/claudeai/status/2028559427167834314 Reply to: none

Text:

> Memory is now available on the free plan. > > We've also made it easier to import saved memories into Claude. > > You can export them whenever you want. > >

Media:

• image: https://pbs.twimg.com/media/HCbjvTwaAAAkLtw.png

Prompts

name: proposal-draft
description: Use when I paste a discovery-call transcript or notes
  and need a client proposal. Do not trigger for regular emails.
 
steps:
1. Pull the client's stated problem in their own words. If it is not
   in the transcript, ask for it and stop.
2. Restate it in one sentence they would nod at.
3. Propose the scope: what I will do, in three to five plain bullets.
4. Price it as a flat project fee, never hourly. Pull the rate band
   from the Pricing file in this project.
5. Close with one clear next step and a date.
 
voice: plain, direct, zero jargon. No hype. Short paragraphs.

WHO I AM
[One line: what you do, who you serve.]
 
HOW I ALWAYS WANT YOU TO RESPOND
Default to [plain / technical / brief]. No preamble, answer first.
Do not ask clarifying questions unless truly blocked. State one
assumption and proceed.
 
NEVER, ANYWHERE
[Your universal bans: filler openers, hype words, whatever you
never want in any answer, in any context.]

WHERE DOES THIS INSTRUCTION GO?
Every kind of work I do?        -> Profile
Only one body of work?          -> Project
A repeatable procedure?         -> Skill
Only this conversation?         -> Chat
 
Rule: put it in the HIGHEST layer that is still true,
and never write it in a second place.

Here is a set of instructions I keep re-typing into chats:
 
[paste the thing you repeat]
 
Rewrite it as a reusable skill: a short name, a one-line
description of exactly when it should trigger, and the steps
written so you follow them the same way every time.

WHO I AM
I run a one-person consultancy helping local service businesses get
more customers online. Non-technical clients, mostly skeptical of agencies.
 
HOW I ALWAYS WANT YOU TO RESPOND
Default to plain and direct, like a smart friend who is not technical.
No preamble, answer first. Do not ask clarifying questions unless truly
blocked. State one assumption and proceed.
 
NEVER, ANYWHERE
No jargon, no "synergy" or "leverage", no fake urgency, no hype adjectives,
no walls of text, no numbers I can't stand behind.

Links

• claude.ai