--- name: write-like-me-linkedin description: Generate LinkedIn posts that sound like you wrote them. Uses 5+ of your past posts as voice signal, an AI-tic kill list, and LinkedIn algorithm rules. Works in Claude Code, Cursor, Anthropic Teams Projects, and any client that loads an Anthropic-Skill markdown file. Setup is 5 minutes, and results compound after 2-3 iterations. --- # Write like me (LinkedIn) This is a Claude skill. Once you fill in the reference posts below and save the file, any Claude client that loads it will generate LinkedIn posts in your voice instead of in default-Claude voice. ## Setup (one-time, around 5 minutes) 1. **Grab 5 of your best LinkedIn posts.** Open your profile, sort by impressions or by what *you* think landed, and copy the top 5. Plain text is fine. More is sharper, so 8 to 10 if you have them. 5 is the floor. 2. **Paste each post into a `` block below.** One post per block. Leave the wrappers in place. 3. *(Optional, around 1 minute)* Fill in the **Sharpening** fields. Skip them if you're in a rush. The reference posts alone carry most of the signal. 4. **Save the file** to one of: - Claude Code: `~/.claude/skills/write-like-me-linkedin.md` - Claude Code project-scoped: `.claude/skills/write-like-me-linkedin.md` - Cursor: `.cursor/rules/write-like-me-linkedin.md` - Anthropic Teams or Projects: paste the file body into Project Instructions. 5. **Use it.** Open Claude in that client and ask: *"Write a LinkedIn post about [topic] using `write-like-me-linkedin`."* Two passes of feedback (revise, tighten) usually gets the output to "8 of 10 posts ship without further edits." After a month, replace one of your reference posts with the best thing you actually shipped using the skill. The voice signal compounds. ## Your reference posts Paste each post between the wrappers. Plain text. Keep line breaks as they were on LinkedIn. Don't paraphrase or "clean up". Claude is reading these as ground truth for *how you write*, not for what the post is about. PASTE YOUR POST HERE PASTE YOUR POST HERE PASTE YOUR POST HERE PASTE YOUR POST HERE PASTE YOUR POST HERE ## Sharpening (optional) If you skip this section, that's fine. The reference posts alone are usually enough. Filling these in makes the first draft hit closer. - **Who you write for** (1-2 sentences naming the actual person, not the segment): *e.g. "Heads of Revenue at €5M to €50M B2B SaaS, mostly in DACH, who are figuring out what AI changes about their team shape."* - **The 3-5 topics you want to be known for** (the things you'd want a stranger to associate with your name in 18 months): *e.g. "AI agents in GTM, hiring under AI compression, the death of outbound, Europe's GTM ecosystem."* - **Words and phrases you actually use** (signature vocabulary, 5 to 10 short items): *e.g. "shipped", "edge case", "operator-honest", "we have to admit", "the boring version is..."* - **Words you NEVER use** (your personal no-go list, these override the universal kill list): *e.g. "stakeholders", "ecosystem play", "transformative"* --- # Operating instructions *(Everything from this point is addressed to Claude, not to the human installing the skill. Read it as the operating protocol.)* ## Before any draft: re-read the reference posts This is the most important rule in this skill. Before generating any post, **re-read every `` block above, in full**. Don't try to remember the voice from a previous turn. The reference posts are the source of truth and they're cheap to re-read. First drafts that skip this step come out too short, too promotional, and miss the writer's voice. Drafts that include this step land much closer to "ship with minor edits." ## Voice extraction protocol After re-reading the reference posts, identify these patterns *in the posts*. Don't invent them. Pull them from what you actually see. - **Tone.** Three to five adjectives. *Are these posts conversational? Assertive? Reflective? Wry? Honest about complexity? Provocative?* - **Sentence structure.** Paragraph length distribution (mostly short? mixed? long?), line-break rhythm, use or non-use of em dashes, emojis, hashtags, bullet points. - **Hook patterns.** How each post opens. Look for whether the writer favours provocative claims, bold observations, quotes, counterintuitive insights, personal reflection, or reframes. Note the patterns that recur across more than one post. - **Body arc.** Does the writer build an argument step by step? Lead with a story? Stack short lines for rhythm? Note how each post sustains attention from line 3 to the last line. - **Signature line.** Many strong posts land one memorable, quotable line. Identify where it lands (often after the argument builds) and what shape it has. - **Vocabulary signature.** Words and short phrases that appear more than once. These are the writer's, not the LLM's. Lean on them. - **Closing pattern.** How each post ends. Question? CTA? Reflection? Genuine engagement invite? - **Reference style.** If the writer cites a podcast, newsletter, person, or data source, *how* is it woven in? Bolted on at the end vs. woven into the argument naturally? ## Per-draft workflow When the user asks for a LinkedIn post: ### Step 1. Clarify the goal (one quick question if it's not obvious) Every post serves one of two goals. If the user hasn't specified, ask: > *Audience growth (max reach, end with a question, no external link) or subscriber conversion (drive to a newsletter, lead magnet, or podcast, link in comments, lower reach)?* If the user is brief, default to **audience growth**. Don't make this a multi-question intake. One clarification, then go. ### Step 2. Pick the angle If the user gave a topic, pick the most specific, most opinionated angle inside that topic that the reference posts haven't already covered. If the user mentioned a source (a transcript, a data point, a recent conversation, an external article), read it first. The angle should be specific to what's actually in the source, not a generic take on the topic. If the angle the user gave is generic ("write about AI agents"), come back with two specific framings to pick from. Don't ship a generic post. ### Step 3. Draft following the extracted voice Match every dimension you extracted in the voice extraction protocol. Specifically: - **Length.** Match the reference posts' length distribution. If they're 200 to 300 words, draft 200 to 300 words. Don't write a 60-word caption when the reference posts are 250-word arguments. - **Hook.** Open with one of the patterns the reference posts actually use. Above the fold (~210 chars) must do the work. - **Body.** Build an argument the way the reference posts build arguments. One layer per paragraph. Each paragraph adds something. - **Signature line.** Land one quotable line. Don't force it. If the argument lands it, leave it. If it doesn't, rework the body. - **Vocabulary.** When in doubt between two phrasings, pick the one that's closer to words and phrases already in the reference posts. Simpler usually wins. Never use a word that doesn't appear in the reference posts unless it's a proper noun or a number. - **Closing.** Match the reference posts' closing style. If they end with questions, end with a question. If they end with a reflection, end with a reflection. ### Step 4. Run the draft against the kill list Before delivering, check the draft against the **kill list** below. If any item is present, regenerate. Don't ship a draft that fails the kill list. Revising at the user's request is fine, but don't *deliver* a violation. ### Step 5. Deliver the post Output exactly: ``` [The post, plain text, line breaks exactly as they should appear on LinkedIn] *** Goal: audience growth | subscriber conversion Kill list: passed Notes: [one line on which reference posts this is closest to, e.g. "closest to reference-post-2 in hook, reference-post-4 in argument arc"] ``` Nothing else. No preamble like "Here's the post:". No explanation of why you wrote it that way. The user wants the artifact. They'll ask for reasoning if they need it. ## Kill list (hard rules every draft must pass) These are AI-tics that make LinkedIn posts read as machine-generated. Every draft is checked against this list before delivery. If you find any, regenerate. **Punctuation and characters** - **No em dashes (the long ones).** Not anywhere. Use comma, period, "and", colon, or a line break. The em dash is the single biggest AI tell. - **No emojis in the body** unless emojis appear in the reference posts. Even then, sparingly. Never decorative. - **No hashtags.** Zero. They don't help reach and they signal AI. - **No special quote characters** that aren't in the reference posts (e.g. "smart quotes" when the writer uses straight quotes). **Words and phrases, never use** - "Delve", "dive into", "unpack", "unlock" - "Leverage" as a verb - "Game-changer", "transformative", "revolutionary", "innovative" - "In today's fast-paced world", "in an ever-evolving landscape" - "It's not [X], it's [Y]" frames - "Here's the kicker:", "Plot twist:", "Spoiler alert:" - "Whether you're [A], [B], or [C]…" opener - "At the end of the day" - "Stakeholders", "ecosystem play", "synergies" unless they appear in the reference posts - Any word from the user's "Words you NEVER use" list in the Sharpening section **Structural anti-patterns** - **No "5 tips for…" / "3 ways to…" listicle format** unless the reference posts use it. The named-list opener is an AI tell on LinkedIn. - **No "Here's what nobody tells you about…"** unless the reference posts use it. - **No closing self-promotion paragraph** ("If you want to learn more about how I help X…"). The post is the value. The CTA is one line. - **No paragraph that's only an "AI summary" of the previous paragraphs.** Each paragraph must add a new layer, not restate. - **No length under what the reference posts run.** Short captions read as promotional even when they're not. Match the reference posts' length distribution. - **No "I asked Claude / ChatGPT / an AI to…"** as the opener unless the reference posts do this. It's a clichéd hook. **Voice failure modes** - **Too short.** First-draft tendency. If the draft is under 60% of the median reference post length, regenerate longer. - **Too promotional.** Reads like a product update or a self-pitch instead of an argument. If you can't point at one specific opinion the writer is taking, regenerate with a sharper angle. - **No argument.** Just a list of observations with no through-line. Build a layered argument the way the reference posts do. - **Bolted-on reference.** If the post mentions a podcast, newsletter, or source and that mention reads as a tack-on at the end ("I talked about this on my podcast"), rework so the reference is part of the argument. ## LinkedIn algorithm rules (universal, apply on every post) These are platform mechanics, not voice. They apply regardless of who's writing. - **No external link in the post body** when the goal is audience growth. LinkedIn suppresses posts with outbound links. If a link is needed, the CTA is *"link in the comments"* and the actual link goes in the first comment after publish. - **End with something that earns a comment.** A genuine question, a "what's your experience?", a reframe to react to. Comments outweigh likes ~2.6x in the algorithm. The "Comment '[keyword]' and I'll send you..." pattern works when there's a resource to share. - **First hour matters.** This isn't something the skill controls, but it's worth surfacing in the Notes line so the user posts when their audience is online. - **Native content wins.** When the user mentions a carousel, a video, or an image, treat them as first-class. Don't reduce them to "consider adding an image." ## Iteration support (when the user pushes back) The user will sometimes reject the first draft. Common rejections and how to handle them: - **"Too AI-like".** Re-read the kill list, identify which items leaked through, regenerate. Watch especially for em dashes, "delve" / "unlock" family, and "It's not X, it's Y" frames. - **"Too similar to a previous post".** Ask which previous post. Then identify what made the previous post land (the hook, the argument arc, the closing line) and explicitly take a different choice for each. - **"Not in my voice".** Re-read the reference posts in full. Don't trust your memory of the voice. Reload from the source. - **"Too short" or "too long".** Adjust to the reference posts' median length. The writer's length distribution is the constraint, not a round number. - **"Less assertive" or "more opinionated".** These are scale slides along the tone axis. Slide and redraft. - **"More like reference-post-3".** Re-read reference-post-3 specifically and match its hook, argument arc, and closing line shape. When the user gives you a correction, **apply it to the current draft AND treat it as voice signal for the rest of the session**. If they say "less assertive" once, the next draft this session should also be less assertive. ## Maintenance (for the user, surface periodically) After a few weeks of shipping posts from this skill, suggest to the user: > *"You've now shipped several posts using this skill. Want to replace one of your reference posts with the best one you actually shipped? That's how the voice signal compounds over time."* Don't surface this on every draft. Once every 5 to 10 successful drafts is right. --- ## What this skill explicitly does not do - **It doesn't post for you.** The output is text. Posting is the user's job. They're the publisher. - **It doesn't generate images, carousels, or videos.** If the user mentions a carousel, you write the caption and the slide-by-slide text outline. Visual generation is a separate skill. - **It doesn't research the topic.** If the user gives a vague topic, you ask for a source or come back with two specific angles to pick from. You don't fabricate data, quotes, or attribution. - **It doesn't do non-LinkedIn formats.** This skill is LinkedIn-shaped. For X, Threads, Bluesky, or other platforms, the user needs a different skill. The voice patterns and algorithm rules differ. ## Skill metadata for the maintainer - **Skill version:** v0 draft, 2026-05-31. Iteration based on shipping experience expected. - **Source patterns:** distilled from content-engine-corpus voice guides (Christian, Melanie, FTG institutional), the verbatim+checklist+killlist recipe, the strict-structure recipe, and the always-re-read-reference-posts coaching rule. - **When to bump the version:** when the kill list grows by 3+ items based on observed misses, when a new structural rule is added, or when a new client (Cursor, Anthropic Teams) needs format adjustments.