Give Claude and ChatGPT better material before asking for a LinkedIn post

“Write me a LinkedIn post about customer research” gives Claude or ChatGPT a subject. It gives the model almost nothing about what you know, whom you help or why your reader should care.

The result is usually predictable: a competent introduction, familiar advice and a conclusion you could attach to somebody else's business. Editing a few phrases will not supply the missing experience.

Give the model better material before asking for prose. Harry Long's Claude writing workflow separates company context, examples of relevant content, current topics, drafting and criticism. His ChatGPT walkthrough adds a useful sequence: choose the angle and hook before drafting the whole post. Together, those ideas form an inspectable editorial process. They do not guarantee virality or require an AI agent to operate your LinkedIn account.

Start with a post worth writing

Choose a question you can answer from real material. A client objection, a decision from a project, a recurring mistake or a carefully examined public example gives the draft something to work with.

For the worked example in this guide, imagine a fictional research consultant who helps small software teams prepare customer interviews. The consultant wants to explain why an interview question about future buying intent often produces an unhelpful answer.

The initial note might say:

Buyers can say they would use a product without showing evidence that they have tried to solve the problem. I want to explain how to ask about a recent actual event instead. Use a hypothetical example. Do not invent a client conversation or a conversion result.

That note already contains an argument, an intended lesson and an evidence boundary. It is much stronger than “write something engaging about research.”

If you have no observation or reasoning to supply, use AI to help identify what you need to investigate. Do not ask it to manufacture the experience that would make the post credible.

Build a reusable author and business brief

Treat the model as a new collaborator who needs an introduction to the work. Prepare a concise reference document with the audience, the service, the problems you address, your actual methods and the claims you can support.

Include representative writing or spoken explanations. Harry suggests fifteen to twenty examples in his ChatGPT video. That can be a useful starting sample, but relevance matters more than the count. Ten recent pieces that sound like you are more useful than twenty old promotional posts written by someone else.

Ask the model to describe the patterns before it imitates anything:

Read these approved samples. Describe the author's usual vocabulary, level of formality, sentence rhythm, use of examples and way of explaining disagreement. Cite a short example for each observation. Identify patterns that vary by subject. Produce a brief I can correct; do not draft a new post yet.

Correct the brief. Perhaps the model mistakes a single sarcastic post for your normal voice, or decides that every piece should begin with a confession. Tell it which patterns are useful and which were specific to one situation.

Add a small bank of your real positions. For the fictional consultant, one might be: “An interview should uncover behavior before inviting speculation about a proposed solution.” Another might explain when a survey is sufficient. These judgments help the model write something more specific than a collection of agreeable principles.

Keep company facts separate from voice preferences. A relaxed tone is a preference. A service scope or customer result is a factual claim with a source. Update the facts when the offer changes, and keep dated material from quietly becoming current positioning.

You do not need to upload every contract, email and meeting recording. Use relevant material you have permission to use, and remove confidential details that do not help the task. A smaller, well-organized source pack is easier to inspect than a vast archive containing contradictory promises.

Study relevant posts without copying a creator

Harry recommends selecting five creators whose work reaches a similar audience, then comparing their posts. The useful question is what helps those readers understand and respond to the material.

Choose examples with enough context to examine. Record the topic, intended reader, opening, evidence, structure, format and apparent response. Include less successful examples, because looking only at viral posts makes every shared feature look like a cause of success.

A simple comparison might reveal that a worked example attracts specific questions while a broad opinion attracts mostly agreement. That is a hypothesis about the reader's needs, not proof that one format always wins.

Use a prompt that preserves the distinction:

Compare these relevant posts for topic, opening promise, supporting detail, format and reader response. Separate observations from possible explanations. Identify three techniques that could suit my own material. Do not copy wording, personal stories, distinctive metaphors or the authors' voices.

The model should help you notice mechanics. It might point out that a writer demonstrates a mistake before naming the principle, or explains a tradeoff before recommending a tool. You can use those techniques with your own research-interview example.

Do not ask for a post “exactly like” a named creator. Their authority may depend on experience and context you do not share. You cannot acquire that by reproducing paragraph length.

Keep current-topic research separate from writing skill

Harry's third layer checks what people are discussing now, with a suggested review every other day. That may suit a fast-moving subject. A consultancy explaining a stable operational problem may need a much slower rhythm.

Collect current topics when they affect the buyer's decision. Record the URL, date, what changed and why your audience should care. A product announcement, new requirement or recurring question can justify a timely post. A popular argument on an unrelated feed cannot.

Check the tool's actual access. A model without browsing cannot reliably tell you what is trending today. A feed sample does not establish search volume. A creator's engagement does not establish customer demand.

If several sources discuss AI-generated interview summaries, the fictional consultant might have a relevant question: when does a summary lose the evidence behind a research finding? The post still needs a substantive answer. “AI is changing everything” is not enough.

Keep evergreen ideas in the plan. A useful explanation of interview questions does not become obsolete because a louder subject appeared this week.

Generate angles before polishing sentences

Once the evidence and audience are clear, ask for different ways to approach the same argument. This is where the model can save useful time.

For the fictional consultant:

Give me six distinct angles on why asking “Would you buy this?” can produce weak customer research. The reader is a software founder planning early interviews. Use the supplied explanation. Each angle must identify its practical payoff and the evidence it needs. Do not add statistics or imply that all future-intent questions are worthless.

Review the options for substance. “The question every founder gets wrong” is an attention device, but it overstates the argument. “Replace a speculative question with a recent-event question” offers a more defensible lesson.

Choose the angle yourself. It should fit your actual opinion, the reader's situation and the material available. Otherwise the model may steer the post toward whichever claim sounds most dramatic.

Run a small hook lab

Harry's Claude workflow asks for eight to ten hook variations, including questions, curiosity, contrarian statements and statistics. Use the range to compare different promises, not to select the most sensational line.

For this example, possible original hooks include:

“Would you buy this?” gives an interviewee an easy way to be polite.

Before asking about a future purchase, ask about the last attempt to solve the problem.

What did the buyer do the last time this problem occurred?

Each sets up a slightly different post. The first needs an explanation of the limits of hypothetical answers. The second promises an alternative sequence. The third puts the reader directly into the research question.

A statistic-based opening belongs only when you have a relevant, verified statistic. A personal-result opening belongs only when the result is yours to describe accurately. Do not let “give me ten options” become an invitation to invent numbers.

Pick a hook that the body can fulfill. You can revise it later if drafting exposes a better explanation. The goal is alignment between opening and payoff, not preserving a clever sentence at any cost. For more writing examples, use our post and hook guide.

Draft from the approved evidence

Now give the model a bounded writing task. The following is an original prompt for the hypothetical example, not a reproduction of Harry's unpublished prompt files:

Draft a LinkedIn post for a software founder preparing customer interviews. Use the approved author brief and the selected opening about asking for the last attempt to solve a problem. Explain the limitation of “Would you buy this?” with one explicitly hypothetical example. Show an alternative question and why it produces more inspectable evidence. Keep the language direct and readable. Do not invent clients, quotations, statistics or results. Flag any claim the supplied notes do not support.

Read for meaning before tone. Does the post actually explain the problem? Has it turned a qualified point into “never ask about the future”? Has it implied that a person's past behavior guarantees a purchase?

A useful body could compare these two questions:

Would you pay for a tool that organized customer feedback?

Tell me about the last time customer feedback became difficult to organize. What happened, and what did you do?

The second question asks for an event and an attempted response. The answer can still be incomplete or inaccurate, and it does not by itself validate a product. The post should preserve that limit.

Then edit the prose. Replace vague sentences such as “Understanding your customer is essential for success” with the actual decision or example. Keep paragraph breaks where they help reading. Do not force every sentence onto a separate line if a connected paragraph explains the idea better.

Give a fresh critic a specific job

A separate review context can help surface weaknesses the drafting conversation has started accepting. Harry calls this the critic agent. It is useful, but it is not an independent fact-checker simply because it is a different chat.

Give the critic the source notes and explicit criteria:

Review this draft against the attached brief. Identify unsupported facts, invented experience, overstated conclusions, unclear audience, copied phrasing, generic filler and any mismatch between the opening and the body. Quote the sentence at issue and explain the edit needed. Preserve useful qualifications. Do not add new facts or rewrite the whole post unless a structural problem requires it.

You can ask for a score if it helps triage, but the explanation matters more. A nine-out-of-ten rating cannot establish accuracy or predict distribution.

Review the criticism yourself. A model may recommend removing a necessary qualification because the stronger claim sounds more engaging. It may add a polished conclusion that repeats the introduction. Accept changes because they improve the argument, not because a second model proposed them.

Make the final human pass factual and practical

Before publishing, verify that every example is presented honestly, every factual claim has support and the voice still reflects what you would say. Check client permissions and current links where relevant.

Read the draft aloud. Repair phrases that sound impressive but do no work. Do not introduce slang, mistakes or fake personal details to disguise assistance. The result should be clearer because someone made editorial decisions.

Decide whether a next step belongs. A post teaching an interview technique might invite the reader to try it on a question they already use. A direct service post can explain how to discuss a research review. Neither needs a stock “Agree?” ending.

Publish through a method you understand and review the exact text and schedule. Writing a post does not require granting an agent browser control of LinkedIn. If you use a connected publishing tool, confirm its actual capability, destination and timing before authorizing the action. For broader agent setup and evidence boundaries, see the Claude research workflow.

Keep the final draft, source notes and the edits that changed the argument. When a post produces a useful reader question, add that question to the next brief. Update the voice and business references only when the evidence warrants it. That gives you a repeatable writing process without asking the model to invent a business behind the words.