7 Claude Prompts That Actually Sound Like You, Not Like AI

Ask Claude for a LinkedIn post about your last hiring mistake and you will get a competent post about a hiring mistake. Ask a hundred founders to do the same thing and you get a hundred versions of the same post, because you are all drawing from the same well: the model's idea of what a LinkedIn post sounds like.

The seven prompts below are built on the opposite instinct. They work today, in plain Claude, with nothing installed, no integration, no MCP server, no subscription to anything. Copy them, paste them, they run. (Written and tested 2026-07-28.)

Key takeaways

• Prompt packs flatten voice because "write a post about X" gives the model a topic and no material, so it fills the gap with the average LinkedIn post it has read a million of.

• The fix is structural, not stylistic: prompts that extract from you beat prompts that generate for you. Telling a model to "sound human" produces a different flavour of the same average.

• All seven skills below run in plain Claude today. No MCP server, no plugin, no API key.

• Every prompt carries a stop rule or a check, a condition under which Claude has to tell you it doesn't have enough to work with, instead of inventing something plausible.

• The honest limitation: prompts live in one client, forget everything between conversations, and can't see what your posts actually did on LinkedIn.

Why every prompt pack makes everyone sound the same

Here is a real-shaped brain-dump, the kind you'd talk into your phone walking back from a meeting:

"ok so we hired a second engineer in march, guy was great on paper, ex-stripe, took him like 6 weeks to ship anything and by then i realised the problem wasn't him it was that we had nothing written down. no onboarding doc, no architecture note, nothing. i'd been holding the whole codebase in my head for 2 years. we let him go in may which felt awful and honestly was my fault not his. now we write a decision doc for every non-obvious choice, takes 20 min, and the third hire shipped to prod on day 4."

Hand that to a generic "turn this into a LinkedIn post" prompt and you get something close to this:

Hiring is hard.

Last year, I made a mistake that cost me dearly.

I hired a brilliant engineer. Ex-Stripe. Impeccable résumé.

Six weeks later, he hadn't shipped a thing.

Here's the thing: it wasn't him. It was me.

The lesson? Document everything.

What's the hardest hiring lesson you've learned?

It is a perfectly serviceable post. It is also unrecognisable as yours, and the reason is visible if you compare the two side by side: almost every specific has been deleted. Two years of a codebase held in one person's head. A decision doc that takes twenty minutes. A third hire shipping to production on day four. All gone, replaced by "document everything".

That deletion isn't a bug in the model. It's the correct answer to the question that was asked. "Write a LinkedIn post about X" asks the model to produce a shape. It has read an enormous number of LinkedIn posts and holds a very confident average of them, and when you supply a topic without material, the average is the only thing it has to work with. So the average is what comes back.

This is why bigger prompt packs don't solve it. Seventeen prompts that each begin "write a post about..." are seventeen doors into the same room. You can usually spot which pack a post came out of, which should tell you everything about how much of you survived the trip.

The other common non-fix is asking the model to write "in a human tone" or "like a real founder". That swaps one house style for another. You get shorter sentences and more lowercase, and every post that used the instruction still sounds like every other post that used the instruction. Voice matching is not a tone setting. It is a question of whose material is on the page.

The principle: extract from you beats generate for you

Every prompt below does one of two things. It either pulls material out of you that the model could not possibly have invented, or it constrains the model to work only with material you already produced. Not one of them asks Claude to be interesting on your behalf.

That has a cost, and it's worth naming up front: these prompts are slower. A generation prompt hands you a finished post in eight seconds. The substance interview asks you five questions first and refuses to draft until you answer them. That is the entire trade, the eight-second post takes eight seconds precisely because nothing of yours went into it.

It's the same reason a good LinkedIn ghostwriter spends the first hour interviewing you rather than writing. The writing was never the hard part. Getting the specific, unrepeatable thing out of your head and onto the page is the hard part, and it is the part most prompts skip.

Each skill below has three parts: what it does, the prompt itself, and what good output actually looks like so you can tell whether it worked.

Skill 1 — The substance interview

This is the one to run first, and the one most people skip. It forbids Claude from drafting anything and turns it into an interviewer that goes after the vague parts of what you said. The output isn't a post, it's a substance sheet: the raw material, sharpened, that every other skill then works from.

You are interviewing me to find the material for a LinkedIn post. You are not
writing the post. Do not draft, outline, or suggest angles.

Here is my raw material:
[paste your brain-dump, voice-note transcript, or one messy paragraph]

Step 1. Ask me five questions, numbered, in one message. Target the places where
my material is vague, general, or asserted without evidence. At least one question
must be about a number, a date, or a name. At least one must ask what I believed
before I changed my mind. At least one must ask what it cost me: money, time, a
relationship, or my own credibility. Do not ask anything I have already answered.

Step 2. After my answers, name the single thinnest answer and ask two follow-ups
on that one only.

Step 3. Then produce a SUBSTANCE SHEET, in my words wherever possible:
- The claim (one sentence)
- Three specifics that prove it (each must contain a number, a name, a date, or
  a quote)
- The reversal (what I thought, then what happened)
- The cost
- The transferable rule (what someone else could do differently on Monday)
- What is still missing (what a reader would fairly push back on)

Stop rule: if the substance sheet would contain fewer than three specifics, say
"not enough material yet" and tell me exactly what to go find out. Do not fill
the gap with plausible-sounding detail.

What good output looks like: questions you find slightly annoying to answer, because they go at the part you glossed over, "what did you tell yourself the problem was in week three?" is the right kind of question. If Claude asks anything you could answer without having lived it, the prompt isn't working.

Skill 2 — The voice sampler

Paste five posts you actually wrote and make Claude derive your patterns as explicit, quotable rules rather than a vibe. The two constraints that make this work: a pattern only counts as a rule if it shows up in at least three of the five posts, and every rule has to carry a quote as evidence.

Below are five posts I wrote myself. Not posts I liked. Posts I wrote.

[post 1]
---
[post 2]
---
[post 3]
---
[post 4]
---
[post 5]

Derive my voice as explicit, checkable rules. Rules only. No compliments, and no
summary of what I write about.

Requirements:
- A pattern counts as a rule only if it appears in at least three of the five
  posts. Anything rarer is an accident, not a habit. List those separately under
  "one-offs".
- Every rule must be followed by a short quote from my posts as evidence. A rule
  with no quote is a guess: delete it.
- Cover at minimum: how I open, how I close, typical sentence length and where I
  break it, paragraph length, punctuation habits, words I reach for repeatedly,
  words and constructions I never use, whether and how I use numbers, and how I
  refer to myself and to other people.
- Include a DO NOT list: five things that would immediately read as not-me.

Output a "voice card" of no more than 20 lines that I can paste into a project
instruction.

Then calibrate: write two sentences about [any topic] that follow the card, and
two that deliberately break it. Label which is which, so I can check whether you
actually heard me.

What good output looks like: rules specific enough to be falsifiable — "opens on a concrete scene or a flat statement of fact, never a question (5/5)" beats "conversational and authentic". The calibration step is the real test: if you can't immediately tell which two sentences are the fakes, the card is too loose to use.

Skill 3 — Rewrite the writer, don't write for them

This is the load-bearing skill in the pack. It takes something you wrote badly and makes it publishable without making it Claude's — by capping how many words Claude is allowed to introduce and forcing it to mark every one of them. Deleting, reordering and fixing grammar are allowed. Upgrading your vocabulary is not.

Below is something I wrote badly. Your job is to make it publishable without
making it yours.

[paste your rough text]

Hard constraints:
- At least 80% of the words in your output must be words I already wrote. You may
  delete, reorder, split, join, and fix grammar and spelling.
- You may add new words only to make a sentence parse or to bridge a reordering.
  Wrap every added word or phrase in [+ ... ] so I can see it.
- Do not upgrade my vocabulary. If I wrote "stuff", it stays "stuff".
- Do not add a moral, a lesson line, a closing aphorism, or a question to the
  reader. If my text ends flatly, it ends flatly.
- Do not add claims, numbers, or examples I did not give you.
- Keep my order of thought, unless reordering makes it clearer why I changed my
  mind. If you reorder, say why in one line at the end.

Then report: the percentage of preserved words, and every sentence you cut, with
a one-line reason each.

What good output looks like: run on the hiring brain-dump above, the result keeps "great on paper", "which felt awful and honestly was my fault not his", and the day-4 detail, while cutting "ok so" and fixing the run-ons. Roughly nine words in ten are still yours. Nothing has been summarised into a lesson.

We hired our second engineer in March. Great on paper. Ex-Stripe.

It took him six weeks to ship anything.

By then I'd realised the problem wasn't him. It was that we had nothing written down. No onboarding doc, no architecture note, nothing. I'd been holding the whole codebase in my head for two years.

We let him go in May. That felt awful, and honestly it was my fault, not his.

Now we write a decision doc for every non-obvious choice. Takes 20 minutes.

Our third hire shipped to prod on day 4.

Compare that to the generated version further up. Same facts, same person, and only one of them could have come from anybody.

Skill 4 — The de-slop pass

A checkable list of tells, rather than the useless instruction to "make this sound less like AI". The prompt audits before it edits, quotes the offending line for each hit, and only ever fixes by deleting or restoring plainer phrasing. It also has to tell you when stripping the tells leaves nothing behind.

Audit this post for AI writing tells. Do not rewrite it yet.

[paste the post]

Check for each of these, and quote the exact line where you find it:
1. Em dashes used as the default connector: more than one or two in a short post,
   or an em dash where a full stop would do.
2. "It's not X, it's Y" and its variants: "Not X. Y." / "X isn't the problem.
   Y is."
3. A manufactured one-line closer that restates the post as an aphorism.
4. A rhetorical question as the opening line.
5. Hollow superlatives: game-changer, incredible, insane, wild, the single most
   important, this changes everything.
6. Three-item lists where the third item is a flourish rather than a fact.
7. Signposting filler: "Here's the thing", "Here's what nobody tells you", "Let
   that sink in", "The lesson?".
8. Vague quantifiers standing in for a real number: dramatically, massively, 10x,
   most people, everyone.
9. A staircase of short lines that are all roughly the same length.
10. A closing question whose only purpose is to farm comments.

For each hit, give: the tell, the quoted line, and a fix that either deletes the
line entirely or restores plainer phrasing. Never introduce a new claim in a fix.

Then answer two questions honestly:
- How many of these lines carry information? If a line can be deleted with no
  loss of meaning, say so.
- After removing the tells, is there a post left? If the answer is no, tell me
  that instead of rewriting it into something smoother.

Finally, output the cleaned version and a count of tells removed.

What good output looks like: run on the generated hiring post above, it flags seven tells — the generic opener, "Here's the thing", "The lesson?", the comment-farming closer, "cost me dearly" as a vague quantifier, and the even staircase of lines — and then says the honest thing: after removing them there are four sentences left and no post, because the tells were doing the work. De-slopping cannot restore substance that was never there. That verdict is the most useful output this skill produces.

If you want the mechanical version of this on a finished draft rather than the diagnostic conversation, our LinkedIn post humanizer runs a similar pass in one click. The prompt is better when you want to understand why a line is a tell; the tool is better when you just want it gone.

Skill 5 — Hook variants without clickbait

Six openers built from six named mechanisms, each of which has to pass a cheque test: name the exact sentence in the body that pays off the promise the hook makes. A hook that can't name one is a cheque the post can't cash, and the prompt is required to mark it as failing rather than offer it to you.

Write six opening lines for the post below. One per mechanism, labelled:

1. Concrete scene: a specific moment, place, or object.
2. Flat number: a real figure from the post, stated without dressing.
3. Admission: something that costs me to say.
4. Contradicted expectation: two facts from the post that sit badly together.
5. The specific noun: the least generic thing in the post, put first.
6. Timestamped moment: "In March", "On day 4", "Six weeks in".

[paste the post]

The cheque test, applied to every hook: name the exact sentence in the body that
pays off the promise the hook makes. If you cannot name one, the hook is a cheque
the post can't cash. Mark it FAILS and do not offer it to me.

Banned: questions addressed to the reader, "unpopular opinion", "nobody talks
about this", "most people get this wrong", and any hook that could be attached to
a different post without changing a word.

Rank the surviving hooks and say in one line why the top one wins.

What good output looks like: on the hiring post, "I'd been holding the entire codebase in my head for two years" survives and gets ranked first, while a vaguer attempt like "I found out the expensive way" is marked FAILS because no sentence in the body says what it cost. That failure mark is the point of the skill. A curiosity gap the post never closes is engagement bait, and it costs you more in credibility than it earns in impressions.

Skill 6 — Repurpose without rehashing

Most repurposing prompts produce a reworded twin: same argument, different sentences. This one inventories the source first, separates what the post actually used from what it merely mentioned, and builds the new post on something unused. The twin test at the end is a hard gate.

Step 1. Read the post below and inventory it. List every distinct claim, fact,
scene, number, and opinion in it. Mark each one USED (it carries the post) or
UNUSED (it appears in passing, or is implied but never developed).

[paste the original post]

Step 2. Show me the UNUSED list and pick the item with the most weight behind it:
the one where I clearly know more than I said.

Step 3. Build a new post whose spine is that unused item. The original's main
claim may appear only as background, in one sentence at most.

Step 4. Run the twin test before you show me anything. Write the original's
argument in one sentence, then the new post's argument in one sentence. If a
stranger would call them the same argument, throw the draft away and go back to
step 2 with the next item. Show me both sentences so I can check.

Constraint: no sentence in the new post may be a reworded sentence from the
original. If you need the same fact, state it differently, because it is doing a
different job now.

What good output looks like: from the hiring post, the UNUSED list surfaces "I'd been holding the whole codebase in my head for two years" and "what I told myself the problem was in week three" — and the second becomes a post about founder self-deception in hiring, not a second post about documentation. Run this against your archive and it doubles as a content batching engine: one experience, several genuinely different posts, which is also how evergreen content earns a second run.

Skill 7 — The post-mortem

Paste a post's real numbers alongside your own median, and extract one rule you can apply before writing rather than after. The design choices that matter: it compares you only to yourself, it normalises reach against resonance so a lucky impressions spike isn't mistaken for a good post, and it is allowed to conclude "not enough evidence".

Here are the real numbers for a post of mine, and my own baseline.

Post: [paste the full text]
Impressions: [n]  Reactions: [n]  Comments: [n]  Reposts: [n]  Profile views: [n]
My median over the last 20 posts: [impressions] / [reactions] / [comments]
Posted: [day, time]

Do this, in order:
1. Normalise. Give reactions per 1,000 impressions and comments per 1,000
   impressions, for this post and for my median. Say whether this post got more
   reach, more resonance, or both. They are different findings.
2. Compare against my own baseline only. Do not compare me to platform averages
   or to anyone else's numbers.
3. Identify the two or three concrete differences between this post and my typical
   post: opening mechanism, subject, whether I named a number, whether I admitted
   something, length, format.
4. State the evidence level plainly. One post is one data point. If the honest
   verdict is "not enough evidence", say that and stop before step 5.
5. Only if the evidence supports it: write ONE provisional rule in if-then form
   that I can apply before writing, not after. Tag it with the number of posts it
   is based on.
6. Name the next post that would test the rule, and say what result would kill it.

Append the rule to a running list. Keep the evidence count next to each rule and
raise it each time a new post supports it. A rule tested four times is worth more
than four rules tested once.

What good output looks like: mostly, "not enough evidence yet" — and then a named next test. When it does produce a rule, the rule is usable before you write ("if the post opens by naming a number from my own business, keep the first line under 12 words; evidence: 3 posts"), not a description of what already happened.

How to make these persistent

Pasting a prompt every time is friction, and friction is why packs get abandoned in week two. Three ways to make them stick, in increasing order of effort:

• A Claude project. Create one project called something like "LinkedIn", put your voice card from Skill 2 into the project instructions, and upload your last twenty posts as project knowledge. Every conversation in that project starts already knowing how you write.

• Skill files. If you use Claude with skills, each of the seven above is a reasonable skill on its own: one file, one job, invoked by name. The substance interview and the de-slop pass are the two worth setting up first, because they're the ones you'll run most often.

• A CLAUDE.md, if you write in Claude Code. Same content, checked into a folder alongside your drafts, so your voice card is version-controlled and travels with the work.

Whichever route you pick, keep the voice card as one file and re-derive it every few months. Your writing moves. A card built from posts you wrote a year ago will hold you to a version of yourself you've already outgrown.

What these prompts can't do

Three limitations, stated plainly, because a prompt pack that oversells itself is just another prompt pack.

They live in one client. The voice card sits in Claude. It isn't in your phone's voice notes, your scheduling tool, or wherever you actually publish from. Anything you work out in one place has to be manually carried to the next.

They don't remember. This is the big one. Every good detail the substance interview pulls out of you exists in one conversation and then it's gone. You mention the day-4 hire on Tuesday; by Friday the model has never heard of it, and you'll answer the same five questions again. Every conversation starts you back at zero, and the material you generated last month is scattered across chat threads you'll never search.

They can't see your results. Skill 7 only works if you go and copy the numbers by hand, and most people do that twice and stop. Claude has no view of what your posts actually did, so the loop where a real result becomes a durable rule stays manual, and manual loops decay.

That last gap is precisely the one a LinkedIn MCP server is shaped to close. A tool called capture_substance takes the specifics that surface in any Claude conversation — the day-4 hire, the objection that came up three times this month — and files them into your ContentIn substance bank, where they're still there next week and available to whatever writes your next post. Same principle as the substance interview, except the answers persist.

To be clear about what exists: the ContentIn LinkedIn MCP server is not live yet. It's in build, with a waitlist. Everything described in this section is what we're building, not something you can install today. The seven prompts above, by contrast, work right now in plain Claude and need nothing from us. If you want the setup groundwork for when the server does ship, we've written the pre-launch getting-started guide. And if you'd rather not wait for the memory part, the AI LinkedIn post generator already trains on the posts you've actually written and keeps your material between sessions.

The full pack, in one block

All seven, in order, ready to paste into a project instruction or a skills folder in one go.

=== SKILL 1: SUBSTANCE INTERVIEW ===
You are interviewing me to find the material for a LinkedIn post. You are not
writing the post. Do not draft, outline, or suggest angles.
Here is my raw material: [paste]
Step 1. Ask me five numbered questions in one message, targeting the places where
my material is vague, general, or asserted without evidence. At least one about a
number, date, or name. At least one about what I believed before I changed my
mind. At least one about what it cost me. Nothing I have already answered.
Step 2. After my answers, name the thinnest answer and ask two follow-ups on it.
Step 3. Produce a SUBSTANCE SHEET in my words: the claim (one sentence); three
specifics that prove it (each with a number, name, date, or quote); the reversal;
the cost; the transferable rule; what is still missing.
Stop rule: fewer than three specifics means "not enough material yet" plus what I
should go find out. Do not fill gaps with plausible-sounding detail.

=== SKILL 2: VOICE SAMPLER ===
Below are five posts I wrote myself. Not posts I liked. Posts I wrote.
[post 1] --- [post 2] --- [post 3] --- [post 4] --- [post 5]
Derive my voice as explicit, checkable rules. Rules only, no compliments.
- A pattern is a rule only if it appears in 3 of 5 posts. Rarer things go under
  "one-offs".
- Every rule needs a short quote as evidence. A rule with no quote is a guess:
  delete it.
- Cover: how I open, how I close, sentence length and where I break it, paragraph
  length, punctuation habits, words I repeat, words I never use, how I handle
  numbers, how I refer to myself and others.
- Include a DO NOT list of five things that would read as not-me.
Output a voice card of max 20 lines I can paste into a project instruction.
Then calibrate: two sentences that follow the card, two that break it, labelled.

=== SKILL 3: REWRITE THE WRITER ===
Below is something I wrote badly. Make it publishable without making it yours.
[paste]
- At least 80% of your output must be words I already wrote. Delete, reorder,
  split, join, fix grammar and spelling.
- Add words only to make a sentence parse or bridge a reordering, and wrap every
  addition in [+ ... ].
- Do not upgrade my vocabulary. If I wrote "stuff", it stays "stuff".
- No moral, lesson line, closing aphorism, or question to the reader. If my text
  ends flatly, it ends flatly.
- No claims, numbers, or examples I did not give you.
- Keep my order of thought unless reordering clarifies why I changed my mind. If
  you reorder, say why in one line.
Report the percentage of preserved words and every sentence you cut, with reasons.

=== SKILL 4: DE-SLOP PASS ===
Audit this post for AI writing tells. Do not rewrite it yet. [paste]
Quote the exact line for each hit:
1. Em dashes as the default connector.
2. "It's not X, it's Y" and variants.
3. A manufactured one-line closer that restates the post as an aphorism.
4. A rhetorical question as the opening line.
5. Hollow superlatives: game-changer, incredible, insane, wild, this changes
   everything.
6. Three-item lists where the third item is a flourish, not a fact.
7. Signposting filler: "Here's the thing", "Let that sink in", "The lesson?".
8. Vague quantifiers standing in for a number: dramatically, 10x, most people.
9. A staircase of short lines all roughly the same length.
10. A closing question whose only purpose is to farm comments.
For each hit: the tell, the quoted line, and a fix that deletes or restores
plainer phrasing. Never introduce a new claim in a fix.
Then, honestly: how many of these lines carry information? And after removing the
tells, is there a post left? If not, say so instead of smoothing it over.
Finally, output the cleaned version and a count of tells removed.

=== SKILL 5: HOOK VARIANTS WITHOUT CLICKBAIT ===
Write six opening lines for the post below, one per mechanism, labelled:
1. Concrete scene. 2. Flat number. 3. Admission. 4. Contradicted expectation.
5. The specific noun. 6. Timestamped moment.
[paste]
Cheque test for every hook: name the exact sentence in the body that pays off the
promise it makes. If you cannot name one, mark it FAILS and do not offer it.
Banned: questions to the reader, "unpopular opinion", "nobody talks about this",
"most people get this wrong", and any hook that would fit a different post
unchanged.
Rank the survivors and say in one line why the top one wins.

=== SKILL 6: REPURPOSE WITHOUT REHASHING ===
Step 1. Inventory the post below: every distinct claim, fact, scene, number, and
opinion, each marked USED or UNUSED. [paste]
Step 2. Show me the UNUSED list and pick the item with the most weight behind it.
Step 3. Build a new post whose spine is that item. The original's main claim may
appear only as background, in one sentence at most.
Step 4. Twin test before showing me anything: the original's argument in one
sentence, the new post's argument in one sentence. If a stranger would call them
the same argument, discard and return to step 2. Show me both sentences.
Constraint: no sentence may be a reworded sentence from the original.

=== SKILL 7: POST-MORTEM ===
Post: [paste]. Impressions / Reactions / Comments / Reposts / Profile views: [n].
My median over the last 20 posts: [i] / [r] / [c]. Posted: [day, time].
1. Normalise: reactions and comments per 1,000 impressions, for this post and my
   median. Say whether this was more reach, more resonance, or both.
2. Compare against my own baseline only, never platform averages.
3. Name the two or three concrete differences from my typical post.
4. State the evidence level. One post is one data point. If the honest verdict is
   "not enough evidence", say so and stop.
5. Only if supported: ONE provisional if-then rule I can apply before writing,
   tagged with the number of posts behind it.
6. Name the next post that would test the rule and what result would kill it.
Append to a running list with evidence counts. A rule tested four times beats
four rules tested once.

Frequently asked questions

Do these prompts need an MCP server or any integration?
No. All seven run in plain Claude today — free or paid, desktop, web, or Claude Code. Nothing to install, no API key, no connector. They also work in other assistants, though the persistence options in the section above are Claude-specific.

How many of my own posts do I need for the voice sampler?
Five is the working minimum, because the three-of-five rule needs something to count against. If you have twenty, use twenty and raise the threshold to "appears in at least half". If you have fewer than five posts you've genuinely written yourself, run Skill 1 and Skill 3 for a while first — you'll produce the sample as you go.

Won't Claude just ignore constraints like "80% of the words must be mine"?
It approximates rather than counts exactly, which is why the prompt also asks it to report the percentage and to mark every addition. You're not relying on the number being right. You're relying on the marks making violations visible, so you can see at a glance whether the output is your text tidied up or Claude's text wearing your facts.

Is this pack a replacement for a LinkedIn writing tool?
For the writing itself, it goes a long way, and that's the honest answer. Where it stops is memory and measurement: nothing you work out in one conversation survives into the next, and Claude can't see what your posts actually did. If you're posting a few times a month, the pack is probably enough. If you're posting weekly and want the rules to accumulate rather than reset, that's the gap a tool closes.

The bottom line

The seven skills above are all built from one idea: the interesting part of a post is the part only you could have written, so the job of the prompt is to get that part out of you rather than to write around its absence. Take them, change them, argue with them. They're more useful adapted to how you actually work than followed exactly.

The one thing they can't do is remember. Every specific you surface today is gone by the next conversation, and every rule you learn from a post's numbers has to be re-learned by hand. We're building the version that remembers — a LinkedIn MCP server where capture_substance keeps what comes up in any Claude conversation, so the material compounds instead of evaporating. It's in build, and the waitlist is the way in: join the LinkedIn MCP waitlist.

If you'd rather have the memory part today, in a product that already exists, start a ContentIn trial — it trains on the posts you've actually written and keeps your voice and your material between sessions.