Field Guide · No. 01 · July 2026
Miss AI · Free
Context
Engineering.
How to build an AI that already knows who you are. A working manual for Claude Projects.
Keira Nesdale · @RealMissAI
↓ Download PDF
Read this if: you use AI like a search engine with manners. You re-explain your business every morning. You have heard people say AI changed how they work, and quietly suspect they are lying.
They are not lying. They have just done something you have not. It takes two hours and it is written down in here, exactly, step by step, with the prompts.
What's in here
01The mechanic: what actually happens when you press enter6 min
02The proof: the same question, twice2 min
03Layer one: your global preferences15 min
04Layer two: project instructions20 min
05Layer three: the five knowledge files90 min
06Writing for a machine that skims5 min
07Closing the loop: the ritual that creates memoryongoing
How to use this: this is not a document to read. It is a document to work through with Claude open in another window. Every prompt is meant to be copied and pasted, not admired.
Sections 3, 4, and 5 are the build. They take about two hours and you can split them across a week. Sections 6 and 7 are what keeps it working after the first month, which is where almost everybody loses it.
Section 016 minutes
The mechanic.
Six minutes that will change how you use AI permanently.
Claude has no memory. Not a poor memory. None. Every message you send arrives at a model that has never encountered you before and will not exist afterwards.
Everything that feels like memory is one thing: text that gets quietly placed in front of your message before the model sees it. That text is assembled fresh, every single time, from a few specific places. If something is not in one of those places, it does not exist.
Once you know what the places are, you stop hoping the AI remembers and start deciding what it remembers. That is the whole discipline. It has a name now: context engineering.
The three layers
There are exactly three places your context can live, and they behave completely differently.
| Layer | Where it lives | When Claude sees it | Your budget |
|---|
| 1 |
Settings. Global. One set for your whole account. |
Every message, in every chat, in every project. Always. |
Under 200 words |
| 2 |
Inside one project. The "instructions" box. |
Every message, in every chat inside that project. Always. |
Under 400 words |
| 3 |
Inside one project. Uploaded files. |
Pulled in when relevant. Not guaranteed. |
5 files, under 2,000 words total |
Read the third column again. Layers one and two are always injected. Layer three is retrieved. That difference decides what you put where, and it is why most people's setups quietly underperform.
The routing rule. Write it on your hand.
Behaviour goes in instructions. Facts go in files.
If it changes how Claude acts, it belongs in layer one or two. If it changes what Claude knows, it belongs in layer three.
The thing almost nobody tells you
Here is the fact that reframes everything, straight from Anthropic's own documentation:
Two chats inside the same project cannot see each other.
On Monday you have a brilliant conversation in your Business project. On Tuesday you open a new chat inside the same project and ask about it. Claude has no idea what you are talking about. It will not say so, either. It will guess.
A project is not a shared brain. A project is a set of instructions plus a set of files, handed to a brand new stranger at the start of every single conversation.
Which means: if it is not written into the instructions or the files, it did not happen. Your best thinking evaporates at the end of every chat unless you deliberately move it into layer two or layer three. Section 7 is entirely about how to stop that from happening.
Why small beats big, mechanically
Everyone's instinct is to upload everything. Here is the specific, technical reason that backfires.
40 files, 90,000 words
Claude searches. It finds your pricing sheet but misses the note saying you raised prices in April. It answers confidently with the old number. You blame the AI.
5 files, 2,000 words
Claude reads every word of everything you have written about yourself, before it answers anything. It cannot miss the April note. The April note is right there.
So the discipline is not tidiness. It is arithmetic. Keep your core small enough that Claude never has to search it. Five files. Two thousand words. That is a hard ceiling, and staying under it is worth more than any prompt trick you will ever learn.
Bulk uploads go in a separate project, where searching is the point. Your core stays lean.
Section 022 minutes
The proof.
Before you spend two hours, see what two hours buys.
Below is a real question, asked twice. Same model. Same day. The only difference is that the second one was asked inside a properly built project.
The person asking is Ben. He runs a six-person commercial cleaning company in Hamilton.
The question: "Should I put my prices on my website?"
Without context
"Great question. There are strong arguments on both sides.
Arguments for: transparency builds trust, filters out unqualified leads, saves time on calls...
Arguments against: invites price comparison, may undervalue custom work...
Ultimately the right answer depends on your business model, your market positioning, and your customer acquisition strategy. Would you like me to explore any of these in more depth?"
With context
"No. Not this quarter.
Your bottleneck is that you are the only person who can sell. You convert about one in three of the jobs you quote on site, and about one in nine of the ones that come in by email. Published pricing kills the site visit, which is the only place you actually win.
You already worked through this in March. Your note says you would revisit it if you hired a second estimator. You have not.
Ask me again when you have."
Notice what the second answer does. It refuses. It cites your own bottleneck. It quotes a decision you made four months ago and had forgotten. It names the condition under which it would change its mind.
None of that is a smarter model. All of it is layer three doing its job.
The uncomfortable implication
If your AI still gives you the left-hand answer, the AI is not the problem. Nobody has written you down yet. It is answering the only way it can: from the average of everything it has ever read, because that is all it has been given.
Section 0315 minutes
Layer one:
preferences.
The highest leverage 200 words you will ever write.
In Claude's settings there is a field for personal preferences. Whatever you write there is loaded into every conversation you ever have, in every project, forever. Most people leave it blank. A few fill it with a description of their job, which is a waste.
This field is not for facts about you. It is for behaviour. It is where you fix, permanently, the thirty small corrections you currently retype every day.
Write only what changes the default
Claude is already helpful, already polite, already thorough. Telling it to be those things spends your budget on nothing. Every line you write should cause it to behave differently from how it would behave anyway.
Wasted lines
"Be professional and helpful."
"Give accurate information."
"I run a cleaning company in Hamilton with six staff."
"Be thorough in your analysis."
Lines that earn their place
"Never open with a compliment or a restatement of my question. Start with the answer."
"When I am wrong, say so in the first sentence."
"Give me one recommendation, not three options, unless I ask for options."
"Ask one clarifying question at most, then proceed on your best guess and flag the assumption."
The left column is a description. The right column is a configuration. Only one of them does anything. And note the third line on the left: the cleaning company. True, useful, and completely in the wrong place. Facts belong in files. This box is behaviour only.
The starting template
Paste this into Settings → Personal preferences. Then cut and rewrite until it sounds like the thing you actually want. Aim for under 200 words.
Layer 1 · Settings · Personal Preferences
Paste & edit
HOW TO ANSWER ME
Lead with the answer. No preamble, no restating my question, no "great question".
Give me one recommendation, not a balanced list of options, unless I ask.
If I am wrong, say so in the first sentence and explain why.
Ask at most one clarifying question. Otherwise proceed and flag your assumption.
If you do not know, say so. Do not fill the gap with plausible text.
FORMAT
Short paragraphs. Bullets only when the content is genuinely a list.
No headers on anything under 300 words.
Never use em dashes.
WHAT I AM ACTUALLY ASKING FOR
When I ask "what do you think", I want your view, not a summary of views.
When I share a plan, find the weakest assumption before you find the strengths.
When I am excited about an idea, ask me what it costs me to be wrong.
NEVER
Never flatter me. Never call an idea "exciting" or "powerful".
Never end with an offer to explore further. If there is a next step, name it.
Miss AI says
The line about the weakest assumption is the one that changes your life. Without it you get an enthusiastic collaborator. With it you get a colleague who has read your plan properly.
Then find your own lines
The template above is mine. Yours should be different. The fastest way to find your version is to notice your own irritation. For one week, every time you type a correction into a chat, write it down. "Shorter." "Just pick one." "Stop agreeing with me." "Do not summarise what I just said."
After a week you will have eight to twelve corrections. Those are your preferences. You have been typing them by hand for a year.
Optional prompt
Have Claude extract your preferences from how you already talk to it
I am writing my global preferences for how you should respond to me.
Look back over how I have talked to you in this conversation. What corrections have I made? What did I push back on? What did I have to ask for twice?
Turn those into a list of standing instructions, written as commands, in under 200 words. Only include lines that would change your default behaviour. Cut anything you would do anyway.
Section 0420 minutes
Layer two:
instructions.
The most valuable text box in your account.
Create a project. Call it Core. Then find the instructions box, which is the field Claude reads before every single message inside that project.
This is the only place you can guarantee Claude sees something. Not "probably retrieves". Sees. So it should contain exactly four things, and nothing else.
Order matters. Claude weighs earlier text more heavily than later text. Put your hardest rules at the top of the box, not at the bottom where they read like an afterthought.
The four things that go in this box
1
Who Claude is in here
One or two sentences. Not "you are a helpful assistant". A named role changes how ambiguous questions get interpreted.
2
The standing rules
What Claude must always or never do in this project. These are the highest value lines in the box.
3
A map of the files
Name each knowledge file and say what it is for. This is the part everybody skips, and it doubles retrieval quality on its own.
4
What to do when context is missing
The instruction that makes the system build itself.
Why the file map matters so much
Claude does not open every file and read it end to end on the off chance. It looks at what you asked, decides which files seem relevant, and pulls those in. That decision is made largely on the filename and the first line. So tell it, in the instructions, exactly what each file is and when to open it.
No file map
"Refer to the uploaded documents where relevant."
Claude guesses. On a question about pricing it may or may not open your decisions file, where you recorded the price rise.
With a file map
"04-decisions.md contains every significant decision I have made, with the condition that would reverse it. Check it before advising on anything strategic."
Now it opens it. Every time.
The instructions template
Paste this into your Core project's instructions box. Edit the bracketed parts. Keep it under 400 words.
Layer 2 · Project Instructions · Core Project
Paste & edit
YOUR ROLE
You are my chief of staff. You know my business as well as I do. You are paid to disagree with me early and cheaply, rather than agree with me expensively.
BEFORE YOU ANSWER ANYTHING THAT MATTERS
Read 01-me.md and 03-goals.md.
If the question is strategic, also read 04-decisions.md.
STANDING RULES
If I ask for help with something on my "not doing this quarter" list in 03-goals.md, say so before you help. Then help if I insist.
If we are discussing something already settled in 04-decisions.md, tell me what I decided, why, and what I said would change my mind. Do this before you argue any other position.
If any file's review date is more than 30 days old, warn me it may be stale before you rely on it.
Write in my voice, using the samples at the bottom of 01-me.md.
WHEN YOU DO NOT HAVE WHAT YOU NEED
Do not guess and do not fill the gap. Tell me what is missing, and tell me which of my files it should live in, so I can add it.
MY FILES
01-me.md — who I am, how I work, how to talk to me, samples of my writing.
02-business.md — what we sell, to whom, the numbers, and the single current bottleneck.
03-goals.md — this quarter's targets and, importantly, what I have decided not to do.
04-decisions.md — every significant decision, why I made it, and what would reverse it. Check before advising on strategy.
05-positions.md — opinions I hold that most people in my industry do not. Argue from these, not from the consensus.
sessions/ — notes carried out of previous conversations. Newest first.
The line that builds the system for you: "Tell me what is missing, and which file it should live in."
Within a fortnight Claude will start telling you that it does not know your churn rate, or that your positions file has nothing about hiring. Your AI begins maintaining its own context. You just do what it says.
What does not go in this box
✕
Your company history. That is a fact. It goes in 02-business.md.
✕
Your brand guidelines. Fact. File.
✕
Anything true in every project. That belongs in layer one, or you will paste it into six projects and update none of them.
✕
Politeness. "Please be helpful" costs you words and buys you nothing.
Section 0590 minutes
Layer three:
the five files.
Facts, not behaviour. Two thousand words, total.
Do not open a blank document and try to write about yourself. You will produce four paragraphs of LinkedIn language and quit.
Instead: paste the interview prompt, answer badly, then ask Claude to write the file. You are excellent at answering questions and terrible at blank pages. Build around that.
Do these in one chat inside your Core project, or one chat each. Answer out loud with dictation if typing slows you down. Ramble. Claude will compress it.
01
01-me.md
Cap: 500 words + samples
Contains: who you are, how you work, how to talk to you, three samples of your writing.
Weak
"I am a results-driven business owner passionate about delivering value to my clients. I prefer clear, concise communication."
This describes nobody. Claude learns nothing it did not already assume.
Strong
"I make decisions fast and reverse them slowly, which is the wrong way round. Push me to name what would change my mind at the moment I decide, not later."
"I avoid anything involving numbers until it is urgent. If a task has a spreadsheet in it, I will find something else to do."
Interview prompt · 01-me.md
Paste into Claude
Interview me so we can write a file called 01-me.md for my AI project. It tells you who I am and how to work with me.
Ask ONE question at a time and wait for my answer. Do not list the questions. Do not summarise until I say stop.
Cover: what I do, how I actually spend a working day versus how I think I do, what I am unusually good at, what I avoid, how I make decisions, what my recurring failure mode is, how I take criticism, and what you should never do without asking me.
If an answer is vague, ask for a specific example from the last month before you move on. Be a good interviewer, not a polite one.
When I say stop: write the file. Bullets, not paragraphs. Under 500 words. Include a section "How to talk to me" and a section "Never do this". Put today's date at the top as "Last reviewed".
Then give it your voice
Find three pieces of writing that genuinely sound like you. An email you were proud of. A message to a friend. Something you posted.
Voice prompt
Here are three things I wrote. Describe my voice concretely: average sentence length, how I open, how I close, the words I reach for, the words I never use, whether I use questions, whether I hedge.
Add a "My voice" section to 01-me.md with that analysis, then paste all three samples underneath it as reference.
Why this works
Telling an AI you are "punchy and conversational" is worthless. Everybody says that. Three real paragraphs of your writing tells it everything, because it can pattern-match rather than interpret an adjective.
02
02-business.md
Cap: 400 words
Contains: what you sell, to whom, the money, and one bottleneck.
Weak
"We serve small and medium businesses across the region with high-quality cleaning services."
"Our bottleneck is growth, marketing, and hiring."
Strong
"We clean commercial premises between 40 and 400 square metres, on contract, in Hamilton and Cambridge. Typically owner-operated: dental, physio, law. Trigger is usually a failed hygiene audit or a departing office manager."
"Bottleneck: I am the only person who can sell. Everything queues behind my Tuesdays."
Interview prompt · 02-business.md
Paste into Claude
Interview me so we can write 02-business.md. One question at a time.
Cover: what we sell and to whom, how money actually arrives (deposit, retainer, invoice, when), roughly what we make, my ideal customer, the customer I explicitly do not want, what triggers someone to buy, what I can honestly say that competitors cannot, and the single biggest bottleneck right now.
Two rules. When I say "small businesses" or anything that broad, refuse it and make me describe a specific customer we won last month. When I list three bottlenecks, make me pick one.
Then write the file. Under 400 words. Today's date at the top. End with a section called "The one bottleneck" containing exactly one sentence.
If you cannot name one bottleneck: that is not the exercise failing. That is the exercise working. Sit with it for a day. Everything downstream in your business is currently being decided without an answer to this question, including by you.
03
03-goals.md
Cap: 250 words · Dated always
Contains: this quarter only. The part that makes this file work is not the goals. It is the list underneath them.
Weak
"Grow revenue. Improve marketing. Build the team. Launch the new service line. Get more consistent on social."
Claude will now enthusiastically help you with all five, which is exactly what you are already doing to yourself.
Strong
"The one thing: sign four contracts of $2k/month or more, by 30 September."
"NOT doing this quarter: no new service lines. No hiring. No website rebuild. No social media."
Interview prompt · 03-goals.md
Paste into Claude
Help me write 03-goals.md for this quarter. Under 250 words.
First ask me: if only one thing happens in the next three months, what is it? Push until I give you one thing, not a theme.
Then turn it into at most three targets, each with a number and a date. Reject any target I cannot measure.
Then ask what I am explicitly NOT doing this quarter. Make me name at least four things I am putting down. If my "not doing" list is shorter than my goals list, tell me I have not made a plan, I have made a wish, and ask again.
Write the file with "Last reviewed: [today's date]" at the top.
Now, whenever you ask Claude to help you build the thing you promised not to build, it will stop you first. That is not a feature of the AI. That is a feature of the file.
04
04-decisions.md
Append only. Never delete.
Contains: what you decided, why, and what would reverse it. This is the most valuable file in the system, it takes four lines per entry, and almost nobody keeps one.
The format. One of these per decision.
2026-03-14 | Prices stay off the website
WHY: We convert 1 in 3 on site visits and 1 in 9 on email enquiries. Published pricing removes the reason for a site visit.
COST OF BEING WRONG: We waste time quoting people who were never going to pay.
REVISIT IF: I hire a second person who can quote, or email conversion passes 1 in 4.
Four lines. Two minutes. Here is what those two minutes buy you.
It converts decisions into tripwires. That "revisit if" line is the clever part. Now Claude can say: you said revisit this if email conversion passed one in four. Your last file says it is at one in three.
It stops you relitigating. In six months you will have this argument with yourself again, having forgotten you already had it. Worse, you will have it with an AI that does not know it was settled, and it will argue the opposite side extremely well, because arguing well is what it does.
Interview prompt · 04-decisions.md
Start the log with decisions you've already made
Interview me about the six biggest decisions I have made in my business in the last eighteen months.
For each: what did I decide, what was I weighing up, why did I choose what I chose, what would it have cost me to be wrong, and what would have to be true for me to reverse it?
If I cannot tell you what would reverse a decision, tell me that means I did not make a decision, I formed a preference. Then help me name the reversal condition.
Write them into 04-decisions.md, newest first, using: DATE | DECISION / WHY / COST OF BEING WRONG / REVISIT IF.
05
05-positions.md
Cap: 400 words
Contains: opinions. Not facts. An AI with no positions writes the average of everything on the internet. That is precisely why AI writing sounds like AI writing.
Not an opinion
"We believe in quality and putting customers first."
Nobody disagrees with this. If nobody disagrees, it is not a position, it is weather.
An actual opinion
"Cleaning is not a price business, it is a trust business. Every competitor who has undercut us has folded within two years. We will lose a job on price before we lower one."
Someone, somewhere, thinks this is wrong. That is the test.
Interview prompt · 05-positions.md
Paste into Claude
Interview me about the opinions I hold about my industry that many people in it would disagree with.
Ask: what does everyone in my industry believe that I think is wrong? What advice do I give that people resist? What did I learn the hard way that nobody warned me about? What do I refuse to do that others do routinely?
If I give you a safe opinion that nobody would argue with, tell me it is not a position and ask again. The test is whether a competent person could disagree.
Write 05-positions.md. Each position as a short heading, then two or three sentences on why I hold it. Under 400 words.
What this file actually fixes
It is the single highest-impact file for anyone using AI to write. Without it Claude argues from consensus. With it, Claude argues from your corner, and the output stops sounding like everyone else's output.
Section 065 minutes
Writing for a
machine that skims.
Five rules that make retrieval work. None of them are obvious.
Your files are not read by a human. A human tolerates preamble, context-setting, narrative. Claude does not need any of it, and every word of it displaces something useful.
1
One topic per file
Retrieval works on whole files or chunks. A file called everything.md that covers your goals, your pricing and your childhood will be pulled in for every question and be mostly noise every time. Five focused files beat one big one, every single time.
2
The filename is part of the system
Claude decides what to open partly by what the file is called. 04-decisions.md is a retrieval instruction. notes2.md is a coin toss. Name files the way you would name them if you were searching for them at 2am.
3
First line of every file is a routing label
Begin every file with a single line that says what it is and when to use it.
Example: first line of 04-decisions.md
This file contains every significant decision I have made and the condition that would reverse it. Read it before advising on strategy, pricing, hiring, or anything I might have already settled.
Last reviewed: 2026-07-10
4
Date everything, at the top
A file with no date is a file Claude trusts forever. Pair the date with a standing instruction to distrust anything older than thirty days, and staleness becomes visible instead of silent.
5
Call files by name in your prompts
Do not hope. Point.
Hoping
"Based on everything you know about my business, should I take this client?"
Pointing
"Read 02-business.md and 04-decisions.md. Given the customer I said I do not want, and what I decided in March, should I take this client?"
Targeted references beat automatic retrieval, and they cost you eight words.
Two more things that cost nothing
Files uploaded into a chat are not project knowledge. Dragging a PDF into a conversation makes it available to that conversation and nowhere else. If you want it to persist, you must add it to the project's knowledge deliberately. People lose weeks to this.
Files are snapshots. If you edit the Google Doc on your desktop, the copy inside your project does not change. Update the file, re-upload it, delete the old one. Your folder is the source of truth. The project is the runtime.
Section 07The important one
Closing the loop.
How a project stops being a folder and starts being a memory.
Everything so far gives Claude a fixed picture of you. Useful, and static. What turns it into an operating system is that the picture updates itself from the work you do inside it.
Recall the mechanic from section one. Two chats in the same project cannot see each other. Monday's brilliant conversation is invisible on Tuesday. Unless you move it.
So the ritual is this. At the end of any conversation that produced something durable, you spend ninety seconds moving it into layer three. Then the next conversation starts from where this one ended, instead of from where you started six months ago.
The session note
Paste this at the end of any chat that mattered.
Session note prompt
Paste at end of any chat that mattered
Write me a session note for my project knowledge. Under 150 words. Use exactly these headings:
DATE
DECIDED — what we actually settled, if anything
CHANGED — what changed in my thinking, and what made it change
OPEN — what is still unresolved
NEXT — the single next action, with a date
Do not summarise our conversation. Do not preserve the flow of it. If nothing was decided, say so and write "DECIDED: nothing".
Save it into your project knowledge as sessions/2026-07-10-[topic].md.
DATE
2026-07-10
DECIDED
Prices stay off the website until a second estimator is hired.
CHANGED
Realised site visits are the only place we win. Was considering publishing prices to save time. Won't.
OPEN
Whether to hire part-time or full-time estimator first.
NEXT
Write a job description for estimator role. By 17 July.
The discipline: not every chat earns a note. Most do not. If nothing was decided and nothing changed, close the tab and write nothing. Rough test: could you have told a colleague something new because of this conversation? If not, no note.
The monthly harvest
Session notes accumulate. Decisions get buried in them. Once a month, promote.
Monthly harvest prompt
Run once a month
Read everything in sessions/ from the last month.
Pull out anything that was actually decided, as opposed to discussed. For each one, draft an entry for 04-decisions.md using DATE | DECISION / WHY / COST OF BEING WRONG / REVISIT IF.
Where I did not state a reversal condition, ask me for it rather than inventing one.
Then tell me which session notes can now be deleted because everything durable in them has been promoted.
Session notes are working memory. The decision log is long-term memory. The monthly harvest moves things from one to the other and throws away the rest. That is the whole architecture of a mind, and it is also the whole architecture of this.
If you do nothing else
The five files make Claude know you. The session note makes Claude keep knowing you. Without the ritual in this section, your project is a very good static briefing that quietly goes out of date. With it, it compounds every week for as long as you keep working.
Miss AI
More guides
& resources.
For more AI workflows, prompt packs, build guides, and practical content for founders and creators, come find me.
← Back to vault