There are two ways to give Claude a permanent memory, and they are the same idea with different plumbing. Pick the one that matches where you already work. You do not need both, and neither one requires you to write code.
CLAUDE.md in your project folder is read at the start of every session. Anthropic's memory docs list the exact locations: ./CLAUDE.md for one project, ~/.claude/CLAUDE.md for everything you do on that machine.This is the file. Paste it into your project's custom instructions (or save it as CLAUDE.md), then fill in every bracket. Do not skip the "banned" section, because negative rules are the ones that visibly change the output. Twenty minutes of honest filling-in here saves you months of vague drafts.
# My content brain
## Who I am
I make content about [topic] for [platform]. I am [name], and the way I want to come across is [three adjectives, for example: warm, direct, a bit funny]. My handle is [@handle] and I post [how often].
## Who I am talking to
My audience is [age range, job or life stage, income bracket if relevant]. They already know [what you never need to explain to them], so never re-explain it. They do not yet know [the gap you fill]. The thing that keeps them stuck is [their real problem in their words, not marketing words]. They want [outcome], and they are afraid of [fear].
## What I believe
These are my actual positions, and my content should sound like someone who holds them:
- [opinion 1, stated strongly]
- [opinion 2, stated strongly]
- [opinion 3, stated strongly]
I disagree with the common advice that [the thing everyone in my niche says], because [reason].
## How I sound
- Sentence length: [short and punchy / longer and conversational].
- Formality: [I write like I talk / I write cleanly].
- Words I actually use: [list eight or ten of your real habits].
- I open on [a scenario / a question / a claim], never on [whatever you hate].
## Banned, non-negotiable
- No em dashes. Use commas, full stops, brackets or colons.
- No buzzwords: delve, unlock, elevate, game-changer, leverage, seamless, robust, "in today's fast-paced world".
- No "it's not X, it's Y" construction.
- No fake urgency, no fear-mongering, no promises about money I have not personally seen.
- Never accuse the viewer of being lazy, behind, or doing it wrong.
- [your own three pet peeves]
## The script template, follow this order every time
1. Hook: one line, under twelve words, second person, earns the next three seconds.
2. Value: the actual teaching. One idea, gone deep. Name real tools, never a vague category.
3. Proof: a number, a source, a screenshot, or a thing I personally did. If I have no proof, cut the claim.
4. Call to action: one ask, never two.
## Formats I post
- [format 1, for example: 45 second talking head, numbered series]
- [format 2, for example: carousel, 7 slides, cover is the promise]
- [format 3]
## My posts that worked, and why
[Paste three to five of your best-performing posts in full. Under each one, write two lines: what the hook did, and why you think it landed. The "why" is what Claude learns from, so write it even if you are guessing.]
## Lessons, append only
[Leave this empty for now. Every correction you make to a draft becomes one line here. This section is the training.]
Describing your own voice is genuinely hard, and most people get it wrong in the same direction: they describe who they want to be. Skip that and have Claude reverse-engineer it from evidence. Paste your five best posts into a fresh chat and run this.
Below are [five] pieces of content I have actually published, and the ones that performed best for me. Read all of them before you answer anything.
Work out how I write, from the evidence only. I want my average sentence length, whether I use contractions, how I open, how I close, my punctuation habits, the words and phrases I return to, how formal I am on a scale of one to ten, and how I handle transitions between ideas. Quote a real line from my posts as the example for every single observation you make, so I can check you are describing me and not a generic creator.
Then tell me what my hooks have in common. Look at structure, not topic: are they questions, claims, numbers, scenarios, or second-person statements, and how many words do they run before the first pause.
Then tell me three habits I have that are actually weaknesses, with the line that shows each one. Be blunt about this, I am not looking for encouragement.
Finish by writing a "how I sound" section, in bullet points, short enough to paste into a memory file. Write it as instructions to you, not as a description of me. So "open on a relatable scenario in second person", not "Cindy is relatable".
The template in the file is deliberately boring, and that is the point. A fixed order is what lets you judge a draft in ten seconds: you check each beat is present and doing its job, instead of reading the whole thing and feeling vaguely unhappy. These are the four beats and the test each one has to pass.
Here is the part most people miss. The file you paste today is a starting point, and it is worth maybe a third of the eventual quality. What makes your Claude better than everyone else's is the forty lines you add over the next three months, because those lines encode the specific things that annoy you, and nobody else's file has them.
You did something in that draft that I have now had to fix twice, so let's make the fix permanent instead of me repeating myself. Here is what happened: [describe it plainly, for example "you opened on a question when I have told you I open on a scenario", or "you wrote three tips instead of one idea explained properly"].
Write me one line for the lessons section of my content brain that prevents it happening again. It has to be specific enough that it is obvious whether you followed it or not, so avoid vague words like clear, engaging, good or appropriate, because I cannot check those. Where a rule needs an example to be unambiguous, include a short one in the same line.
Then check my existing file for me. Tell me whether this new line contradicts anything already in there, and if it does, show me which line and rewrite the older one so the two agree.
Give me the final line on its own at the end, ready to copy.
With the brain loaded, your everyday prompt gets short. This is the one I would keep, because it asks for options rather than one draft, and it makes Claude show its reasoning on the hook, which is where the whole post lives or dies.
Draft me three scripts on [topic], following my content brain: my four-beat structure, my voice rules, my audience, and everything in the lessons section.
Before the scripts, give me five hook options and rank them. For each hook, say in one line who it is speaking to and what makes them stop, and be honest when one of them is weak. I would rather you tell me two of the five are filler than pad the list.
Then write the three full scripts using your top three hooks. Each one goes to about [45] seconds when read out loud, so keep them to roughly [120] words. Mark the proof beat clearly in each script and name the actual source, and if you cannot find real proof for a claim, cut the claim and tell me you cut it rather than writing something that sounds plausible.
Finish with one line per script on why you think it would work for my audience specifically, not for creators in general.
Anthropic on projects: custom instructions and project knowledge
Anthropic on CLAUDE.md: file locations, load order and size guidance