How to use Claude for LinkedIn thought leadership without sounding like AI

Claude sounds like AI on LinkedIn when the prompt is a topic and the reply is a memo. "Write thought leadership about pricing" has no buyer, no week, and no decision you made. Claude fills that gap with a balanced essay: context, both sides, three lessons, and a recap. The sentences are cleaner than a rushed ChatGPT draft. Clean is the tell. A colleague in your market can post the same paragraph after swapping the industry noun.

The fix is a note you wrote first, a short instruction that forbids the memo shape, and a read that deletes any noun you did not supply. Boykli is the place to keep that note, draft it under a length, and publish it on the personal profile you connect. Create posts does not call Claude. The post model is Grok 4.6 on both seats, with no model picker. Claude Sonnet 5 is a choice on the carousel page, for the PDF, after you already have the lesson. Boykli does not read your Claude account, your Instructions for Claude, or a skill you built there.

Logged-in visitors can browse Radar and Algorithm without a plan. Writing, images, carousels, Voice Agent, publishing, and scheduling need a paid seat. BYOK is $29 a month. Managed is $88 a month. Prices are the public ones on boykli.com.

What the search is asking

People who type this query already like Claude's ear. They have tried "make it sound more human" and received a warmer memo. They want a post a buying committee can forward, under their name, on a cadence. They do not want "stakeholders should leverage nuance."

Thought leadership, in the useful sense, is one decision a stranger can reuse: what you refused, what you sequenced, what the number counted. The mechanism is the same one in LinkedIn thought leadership automation. Automation may draft and schedule. It may not invent the opinion. A topic prompt invents the opinion. Claude does it politely.

Four places the same note can go

The second column is a discipline inside Claude. The third is a library and a queue, written by a different model. The fourth is the only Boykli surface that runs Claude, and it runs it on slides. How a topic-only ChatGPT reply fails in a louder way is why raw ChatGPT LinkedIn posts fail. How "voice" differs from "which model" is AI LinkedIn ghostwriter.

Free Creator Tool: Test your draft against LinkedIn's 2026 distribution rules using the Free LinkedIn Algorithm Checker (scores 140-char hook fold, link penalty, and dwell potential). Turn long posts into visual slides with the Free LinkedIn Carousel Generator.

Why Claude's default reads as generated

Claude is trained to be careful. Careful, with an empty brief, becomes a consultant memo. These are the lines to cut. They are patterns in the reply, not a score from a detector.

Short mode in Boykli is told to refuse a fixed opener list: "Here's the thing," "Let that sink in," "Unpopular opinion," "Hot take," "The truth is," "In a world where," "Nobody talks about this," "This is your reminder," and starting with "I." Closers such as "Agree?", "Thoughts?", and "Comment below" are on that ban list. Medium and long posts ban hashtags, emoji, and invented proof. Claude does not receive that list unless you paste it. None of this is a promise that a given draft will clear your ear.

LinkedIn's June 2026 note, Keeping conversations real, says content that looks generated and lacks a clear perspective is less likely to be widely distributed beyond the immediate network. Using AI to help write is described as acceptable when the post is still your perspective. Their initial tests identified generic content correctly 94% of the time. That 94% is LinkedIn's own figure, not an audit of your account. The September 2026 note, How LinkedIn is tackling AI slop, describes slop as low-effort, polished, and missing a point of view, and says views of content their classifier labeled as slop were down 40%. Polished and empty is the Claude failure mode. Boykli does not see that classifier and does not promise a distribution tier.

A specimen

This composite is the topic-only shape. It is not a capture of one Claude reply, and it is not a post to publish.

Pricing is one of the most nuanced decisions a founder will face. On one hand, a lower price widens the top of the funnel. On the other, it can attract the wrong stakeholders.

Three lessons stand out. First, price on value. Second, talk to customers. Third, stay willing to revisit the number.

It's worth noting that the best companies treat pricing as a conversation, not a spreadsheet. I hope this framing is useful.

Nothing in it happened. "Nuanced," "stakeholders," and "three lessons" are the model completing a school assignment. "I hope this framing is useful" is the assistant talking to you, left inside the draft.

The same founder, with the fact written first:

We moved the team plan from $49 to $79 in March after 14 expansion calls. Eleven of them asked for audit logs before they asked for a discount. The four accounts that churned in April were all on the old $49 plan and had never turned logs on. The $79 page now leads with the log, and the discount button is gone.

That note is already the post. Claude is useful if it keeps March, $49, $79, 14 calls, 11 log requests, and four churns, and if it refuses a third lesson about "pricing as a conversation."

Set Claude up so the memo has to lose

Claude's help page on personalization separates three controls, and they do different jobs.

Instructions for Claude apply to every chat. The page says to set them under Settings, on Your Claude. Older layouts call that screen Account, and the field Instructions for Claude. Use a few lines about shape: one LinkedIn post, first sentence is the decision in the message, no hashtags, no emoji, no statistic that is not in the message, no opening compliment, no closing recap, no "I hope this helps." Leave your industry out of the standing text. "I write B2B thought leadership" teaches Claude to invent a category post on every empty prompt.

Project instructions apply inside one project. The same help page says projects exist on the free plan, with a cap of five, and marks project instructions as a paid-plan control. A project called "LinkedIn" is the right drawer for the shape rules if you do not want them on every coding chat. Put source material you are willing to publish in the project. Do not paste other people's posts and ask for them "in my voice."

Skills are procedures Claude can load when a task matches. Anthropic's help page says they are available on Free, Pro, Max, Team, and Enterprise, and that they require code execution to be enabled. A custom skill can describe a format. It does not know your last quarter unless you put that quarter in the skill or in the message. A skill that says "sound like a thoughtful founder" reproduces the memo.

Standing instructions do not replace the note. They only stop the chrome. Anthropic's own prompting guidance prefers telling the model what to do, with a reason, over a long stack of "never." One reason is enough: "The reader is a buyer who will check the nouns, so keep only the facts in this message."

The workflow, in order

Do this once on a decision from the last month. Generating first and "adding your voice" second is how the three lessons get in.

1. Write the note before Claude sees it. One situation. The constraint. What you decided. If a number matters, write what it counts and over what window. If you do not have the number, leave it out. A usable note is a few sentences. Ten characters is only the minimum Create posts will accept later. It is not a useful brief.
2. Open a Claude chat that already has the shape rules, either from Instructions for Claude or from the LinkedIn project. Paste the note. Ask for one post, not three variants. Three variants push the model to decorate. Name the ceiling in the same message: short enough to read on a phone, or a body you would actually post. LinkedIn's own ceiling is 3,000 characters. Boykli's page labels, if you move across later, are short at 45–400 characters, medium at 400–1400, and long at 1400–3000.
3. Read the reply against the note with a pen. Mark every noun that was not in the note. Delete the preamble, the unused "on the other hand," the extra lessons, the recap, and any "I hope this helps." If a percentage appeared, delete it. A follow-up that says "make it more specific" often adds a new fake number. The safe follow-up names the cut: "Remove the third lesson. Keep the March price change and the 11 calls. Do not add a study."
4. If this draft should live in a library and leave on a schedule, move the original note into Boykli, not the decorated reply. Create posts on a paid seat. Short, medium, and long take the note. YouTube takes a video URL and is not allowed to write from the title when the watch fails. Article takes a link the app could read. Document takes the file. Series takes source text and returns several posts. The tone menu changes with the mode. Thought Leader asks for a strategic lesson. Personal Story, Analytical, and Teach & Explain are often the better fit for an incident. A short post can be Humor. A long post cannot. Those names are instructions for this draft. They are not a voice file. The full mode list is on the LinkedIn post generator.
5. Read the Boykli draft the same way. Each write mode allows two refinements. Managed accounts also have a monthly cap of 130 post refinements and 210 posts. BYOK has no Boykli generation cap. You still pay OpenRouter for chat and images, and Google AI Studio for Voice Agent. Manual edits in the editor are yours. They do not train the next post.
6. Check the opening on a phone width. LinkedIn does not publish a "see more" character count. Measurements people repeat put the opening near 210 characters on desktop and near 140 on a phone. Boykli does not draw that fold. Paste the finished lines into LinkedIn's composer, or use the method on LinkedIn post preview.
7. Publish now, or queue the saved text, on the one personal profile you connect with LinkedIn OAuth. The queue does not post a company Page, and it does not comment, like, or send invitations.

If the material is something you would rather say than type, Voice Agent runs a live interview and writes three drafts from that transcript. The draft model on that path is Gemini 3.5 Flash. It is not Claude, and it does not ingest a Claude export.

Where Claude Sonnet 5 actually runs

Create posts has no switch that says "use my Claude subscription" or "use Claude Sonnet 5 for this post." Posts go out as Grok 4.6. On the managed path, Grok posts use web search. On BYOK, the OpenRouter call can use the web plugin. A missing OpenRouter key fails the BYOK call. It does not fall through to Boykli's xAI, OpenAI, Anthropic, or Vertex accounts.

The carousel page is the Claude control. The picker is Grok 4.6 (the default), Gemini 3.5 Flash, Claude Sonnet 5, or GPT-5.6 Terra. The selected model is the model that runs. There is no silent swap. Pick Claude Sonnet 5 when you want that model's sentence rhythm on slide copy, and when the artifact is a PDF of 2 to 8 slides, not a text post. The source is still the lesson you supply. Claude Sonnet 5 does not receive your Instructions for Claude, your project, or your skills. Brand kit, dimensions, and the safe zone are separate choices, covered on the LinkedIn carousel generator page.

A slide deck is a second artifact. It is not a way to smuggle a Claude text post through the product. If you only needed the paragraph, stay in the chat until the nouns match the note, or use Create posts and accept Grok 4.6.

A second pass on the specimen

Medium mode, Analytical, source equal to the pricing note and nothing else, should be able to land near this:

In March we moved the team plan from $49 to $79. The reason was 14 expansion calls. Eleven of them asked for audit logs before they asked for a lower price. In April, four accounts churned. All four were still on $49, and none of them had turned logs on.

The $79 page now opens with the log. The discount button is gone. A price change that leads with the feature people already requested is a packaging fix. A price change that leads with "we raised prices" is a press release.

If a draft of that note comes back with "stakeholders," a third lesson, or "I hope this helps," the draft failed the spec. Refine once: "Remove the extra lesson and the sign-off. Keep March, both prices, 14, 11, and four." If the second pass still adds a study, edit the lines yourself.

When the Claude tab is the right tool

Use Claude when you are still finding the sentence, when the project already holds the source you are allowed to publish, or when you want a private scratchpad and you are not posting today. A brainstorm that stays in the chat cannot fail on LinkedIn. The failure starts at paste.

Stay in Claude if you want that model's cadence on the text itself. Boykli will not imitate it on Create posts. Move the finished note into Boykli when you want the length bands, Library, and a queue. The founder note on why the words have to start from something you brought is building Boykli.

FAQ

Does Boykli write the LinkedIn post with Claude?

No. Create posts uses Grok 4.6 on Managed and on BYOK. There is no Claude option on that screen. Claude Sonnet 5 is on the carousel model picker, for the PDF. Voice Agent drafts use Gemini 3.5 Flash. Boykli does not call your claude.ai account.

Will Instructions for Claude make the post sound like me?

They change the shape of Claude's replies: length, bans, whether it recaps. They do not contain this week's decision unless you put that decision in the message. A line like "sound like a founder" without the incident produces a smoother memo. Your old LinkedIn posts are not loaded unless you paste them, and pasting them as a style sample still does not supply today's fact.

Can I paste a Claude draft into Boykli and humanize it?

You can paste text into the core idea. The writer rewrites under the mode you picked. It does not run a detector, and it does not restore facts the first model invented. If the Claude draft says "research shows," that phrase is now source material. Paste the original note instead.

Does a custom skill fix the AI sound?

A skill can hold a format: one incident, one decision, no recap. Anthropic describes skills as procedures that load when the task matches, and the help page says they need code execution enabled. A skill that only says "thought leadership, warm, expert" reintroduces the memo. The skill also does not travel into Boykli.

What does a seat cost, and what is capped?

BYOK is $29 a month: your OpenRouter key for chat and images, your Google AI Studio key for Voice Agent and for YouTube when a video has no captions, and no Boykli generation cap. Managed is $88 a month: platform models, 210 posts, 150 carousels, 150 images, 25 Voice Agent sessions, and 130 post refinements. Each draft still allows only two refinements on either seat. A cancelled seat works until the current period ends. An on-hold seat does not.

Is Claude safer than ChatGPT for thought leadership?

Claude's default is more polished and more balanced. On an empty topic prompt, that polish is the slop LinkedIn described: generic, finished, and missing a point of view. ChatGPT more often adds a bait closer and a fake percentage. Both are fine once the note holds the decision and you delete every noun you did not write. The model is not the safeguard. The note is.