17 AI skills for finding customers and making content

linkedin-post-editor

make anything you write sound like a person, not a bot

How the two work together

Claude thinks it through. Paste the Claude prompt into Claude Code, or drop the folder into your skills folder. Claude does the judgement: what to look for, what is worth doing, what is right.

Codex gets it done. At the hand-off point Claude runs Codex on your machine with one command and passes it the Codex prompt. Codex does the mechanical part and hands the result back. Claude checks it before you see it.

No API key to set up: Claude calls the Codex you already have installed. If Codex is not installed, Claude does that half itself and tells you.

Prompt for Claude

---
name: linkedin-post-editor
description: Edit a draft LinkedIn post so it reads as the author wrote it, with the AI tells marked then removed, every claim kept and the post under 150 words; use on any draft before it is published.
---

# Get your LinkedIn draft read to the last line

Your post, in your voice, the AI tells gone and every claim intact.

## Claude does

1. Get {DRAFT}, one or two of the author's past posts as {VOICE_SAMPLE} (optional), and whether long-form is wanted (default no). Read the sample before the draft. Record register, first person, humour, emoji and em dash habits as {VOICE_NOTES}. No sample: the draft's plainest sentences stand in.
2. Write {AUTHOR_POINT}: one sentence, one committed stance. If the draft argues both sides, ask the author which. Never choose for them.
3. Set {WORD_TARGET}: 150, or 300 only if long-form was explicitly asked for.
4. Mark pass, no edits. Search the draft for each pattern and list every hit with its pattern name as {TELL_LIST}: cut-on-sight words, "not X but Y", forced triplets, false "from X to Y" ranges, "serves as" for "is", brochure adjectives, "experts say", a sentence restating the last, chatbot residue (preamble, both-sides balance, closing recap), bold, emoji, markdown. Act on clusters, not one hit. Skip quotations, titles and names.
5. List every "many", "significant", "a lot", "various" and each unanchored claim as {SPECIFICS}. Ask the author for the number, name or moment behind each. None given: mark "delete". Never invent one.
6. Name the one dominant rhythm pattern (same-length sentences, one-line paragraphs, announce-say-recap) as {RHYTHM_NOTE}.
7. Judge the first line alone, as the feed preview shows it. If it fails, write three from the draft's own material and pick one: {FIRST_LINE}.
8. Pick one close as {LAST_LINE}: a question only this post could ask, a plain statement of the point, or the ask. Never a recap.

## Then Codex does

Codex restates each marked point in the author's words instead of patching phrases, applies the specifics, breaks the one rhythm pattern, sets paragraphs, counts words with wc and saves the post and the change list. Fixed instructions on bulk text: Codex's job.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Fill in first: {DRAFT}, {VOICE_NOTES}, {AUTHOR_POINT}, {WORD_TARGET}, {TELL_LIST}, {SPECIFICS}, {RHYTHM_NOTE}, {FIRST_LINE}, {LAST_LINE}.

## Claude checks

Read edited-post.md and changes.md before the user sees anything.

1. Claims, sentence by sentence against {DRAFT}: nothing added or lost; no "may" to "will", "some" to "most", "associated with" to "caused"; qualifiers and attributions intact.
2. Search, never recall: every {TELL_LIST} phrase gone, cut-on-sight list clear, em dashes within the rule, no bold or emoji the author did not use. wc -w under {WORD_TARGET}.
3. Opens with {FIRST_LINE}, ends with {LAST_LINE}, no recap before it.
4. Each {SPECIFICS} sentence carries the supplied detail or is gone. No new specific anywhere.
5. Over-correction: no fake slang, staged typos, invented anecdotes or feelings, opinions the author did not hold. Sentence spread: longest minus shortest 20 words or more, under half in the 10 to 20 word band, variation from the content, not scattered.
6. Read it aloud. "What still sounds AI-generated?" A remaining cluster fails.
7. No preamble, changelog, [Name], TK or markdown in edited-post.md; changes.md has both counts and five bullets at most.

Any failure: send Codex the exact line and the fix, rerun, recheck until it passes. Never patch it silently.

## Rules

- Editing, not rewriting. Keep the author's point and voice; the least invasive edit wins.
- Cut on sight: "In today's fast-paced world", "game-changer", "unlock", "leverage", "delve", "navigate the landscape", "it's not just X, it's Y", "let that sink in", "stands as a testament"; emoji bullets, and emoji anywhere unless the author already uses them; any sentence that restates the previous one; rhetorical triplets unless there is exactly one and it lands; "Thoughts?" or "Agree?" as a closing line.
- Em dash: at most one, none if the author's sample has none, their rate if they use them. One "however", a formal word or curly quotes alone is never a tell.
- Under 150 words unless long-form was explicitly wanted; then under 300 and every paragraph earns its place.
- Never add a claim the author did not make. Never add a statistic. A confident wrong detail is itself a tell.
- Structure before words; no synonym cycling.
- No fabricated humanity, no random sentence-length wobble; fix the dominant pattern, not the count.
- Personality only where the author's voice has it; never strip it from a personal post.
- Never mark the draft clean from memory. Count.
- Not detector evasion; never promise an AI score.

## If Codex is not installed

Claude does the Codex section itself and says so.

## Built from

- blader/humanizer, https://github.com/blader/humanizer, 42,729 stars on api.github.com: 35 named patterns, keep every claim, sample beats rules.
- Wikipedia: Signs of AI writing, https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing, the WikiProject AI Cleanup field guide behind every skill above.
- harshaneel/humanize, https://github.com/harshaneel/humanize, 416 stars: countable rules, em dash rate, sentence spread.
- Anbeeld/WRITING.md, https://github.com/Anbeeld/WRITING.md, 363 stars: least invasive edit, over-correction check, meaning kept.
- Louis-Francois Bouchard, https://www.louisbouchard.ai/ai-editing/, Towards AI editor: structure before language, silent self-check.

Prompt for Codex

# Get your LinkedIn draft read to the last line - Codex task

## You are given

- {DRAFT}: the author's post, verbatim.
- {VOICE_NOTES}: register, first person, humour, emoji and em dash habits. Match them.
- {AUTHOR_POINT}: the one point and the one stance. Both stay.
- {WORD_TARGET}: 150 or 300. Come in under it.
- {TELL_LIST}: each flagged phrase with its pattern name. Every entry goes.
- {SPECIFICS}: each vague word or unanchored claim with the number, name or moment that replaces it, or "delete" for that sentence.
- {RHYTHM_NOTE}: the one dominant rhythm pattern to break.
- {FIRST_LINE} and {LAST_LINE}: use verbatim.

## Produce

Two files in the working folder.

1. `edited-post.md`: the post only, plain text, ready to paste.
   - Opens with {FIRST_LINE}. Ends with {LAST_LINE}, nothing after it.
   - Work through {TELL_LIST} top to bottom. For each hit, restate the sentence's point in the author's plain words. Where the phrase carries no information, delete it and join the neighbours. Never swap a synonym in place.
   - "not X but Y" becomes the Y. A forced triplet becomes the one or two items that matter. "serves as" becomes "is" or "has". Bold mini-heading bullets become sentences. A sentence restating the previous one goes.
   - Apply {SPECIFICS} exactly: the supplied detail, or delete the sentence.
   - Break {RHYTHM_NOTE} only. Let length follow content: a short sentence where the point is short, a long one where it carries a condition. Target: longest minus shortest sentence at least 20 words, fewer than half the sentences 10 to 20 words. No programmed wobble.
   - Paragraphs of one to three lines, blank line between.
   - Em dashes: at most one, and zero unless {VOICE_NOTES} says the author uses them; then match their rate.
   - Plain text: no markdown, no bold, no emoji unless {VOICE_NOTES} allows it, no placeholders.
   - Before saving, silently check the result against every rule below and read it back once.
2. `changes.md`:
   - Line one: "Before: N words. After: N words." Both from `wc -w`, never estimated.
   - Then at most five bullets, "what changed: why", biggest structural change first.

## Rules

- Keep every claim. Every fact, name, number, date, quote and ranking in the output is already in {DRAFT} or {SPECIFICS}. Never add one to sound specific.
- Never change meaning: "may" stays "may", "some" stays "some", "associated with" stays; keep every qualifier and who said what.
- Do not rewrite anything inside quotations, titles or names.
- No fabricated humanity: no slang, typos, anecdotes, feelings, "actually" turns or opinions the author did not write. Match {VOICE_NOTES}, not a casual default.
- Least invasive edit: a sentence not on {TELL_LIST} stays as written unless {RHYTHM_NOTE} or {SPECIFICS} touches it.
- Introduce no new tell: no "leverage", "unlock", "game-changer", "delve", no new "not X but Y", no new triplet, no closing recap.
- No preamble, changelog or commentary inside edited-post.md.
- Create or change nothing outside the two files.

## Return

Print, in this order:

1. The full contents of edited-post.md.
2. Before and after word counts.
3. Counts of em dashes, bold markers and emoji in edited-post.md, from grep, not memory.
4. Longest and shortest sentence lengths in words.
5. Full path of each file written.
6. Any {TELL_LIST} or {SPECIFICS} entry you could not apply, with the reason, or "none".

Built from the best public work on this

linkedin-post-editor: sources

Built from

  • blader/humanizer, https://github.com/blader/humanizer, 42,729 stars on api.github.com (v2.11.2, MIT): 35 named AI-writing patterns with before/after pairs, keep every claim, invent nothing, the writer's sample beats the rules.
  • Wikipedia: Signs of AI writing, https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing, the WikiProject AI Cleanup field guide compiled from thousands of AI-generated submissions; the declared source of every skill above.
  • harshaneel/humanize, https://github.com/harshaneel/humanize, 416 stars on api.github.com: seven countable hard rules (em dash rate, sentence-length spread, banned vocabulary), strip chatbot voice, count occurrences rather than trusting memory.
  • Anbeeld/WRITING.md, https://github.com/Anbeeld/WRITING.md, 363 stars on api.github.com (MIT): least invasive edit, supported specificity, meaning preserved during edits, the over-correction check against fabricated humanity.
  • Louis-Francois Bouchard, https://www.louisbouchard.ai/ai-editing/, Towards AI editor of thousands of AI-assisted submissions: fix structure before language, silently check your own output before presenting it.

Best public prompt we found for this job

Source: https://github.com/blader/humanizer/blob/main/SKILL.md (sections "What to do" and "Rewrite process", v2.11.2, MIT; verbatim except one sentence about return modes and the fiction exemption removed for length).

```text

When given text to humanize:

1. Find AI patterns. Check the text against the patterns below.

2. Keep every claim. You may shorten dull parts, expand useful parts, and merge or split paragraphs. Keep the information even when you change the structure.

3. Do not invent facts. Do not add a fact, name, number, date, quote, or citation unless it comes from the source or the user. If a sentence needs a missing detail, ask for it or use a simpler sentence. You may add an opinion or reaction when the writer's voice calls for one, but you may not add a factual claim.

4. Match the voice. Use the right tone for the text, such as formal, casual, or technical. Add personality only when the text and the writer call for it.

[...]

Rewrite process

1. Read the source and mark each AI pattern.

2. Write a draft. Read it aloud. Check the rhythm, details, simple verbs such as is and has, and the right level of formality.

3. Ask two questions:

  • "What still sounds AI-generated?"
  • "Did the rewrite add or remove any fact, name, number, date, quote, citation, ranking, or other claim?"

Treat any unsupported addition or lost claim as an error.

4. Write the final version. State each point naturally instead of patching one flagged phrase at a time. If a sentence stays awkward, rewrite the paragraph around its main point. Apply the dash rule in §14.

```

Want this running in your business?

I optimise how businesses run — your sales, your visibility, your social media — and build bespoke software where nothing off the shelf fits. The first conversation is free. Work starts from £150 a day.