Linkedin Humanizer

LinkedIn Humanizer

A draft that sounds generated fails twice. Readers who recognise the cadence stop at the second line, and the author's name is now attached to a paragraph that could have been posted by anyone. The usual fix makes it worse: someone runs a find-and-replace on a list of forbidden words, chops the long sentences into fragments, sprinkles in "honestly", and produces text that sounds like a machine imitating a person who is trying not to sound like a machine.

This skill treats the problem as editing, with a catalogue. Thirty-three rules are sorted into three tiers by what should happen to the text: remove (tool residue no person writes), reduce (habits that are fine once and a signature in bulk), and review (choices a careful writer makes on purpose). Each rule carries the reason it fires, the fix, and the case for keeping the text as it is. A separate scorer handles emoji placement, and a fingerprint built from the author's own past posts decides which habits are theirs and must survive the edit.

Scope boundary. This skill edits and audits a draft that already exists. It does not choose an angle or write a post from a blank page; that is linkedin-post-writer. It does not classify the opening lines of other people's posts for reuse; that is linkedin-hook-analyzer. Comments and replies have their own skills (linkedin-comment-writer, linkedin-reply-manager), though the catalogue here applies to any short text pasted in. Interviewing the author for the missing story or figure is linkedin-story-interviewer; this skill only reports that the figure is missing. It makes no claim about how any automated classifier will score the text and does not try to defeat one: the target is a human reader. Everything runs offline on text the user supplies. Nothing is posted, scheduled, fetched or sent anywhere; the user pastes the finished draft into LinkedIn themselves.

When to use this skill

• A draft was produced with a writing assistant and is about to go out under a person's name
• A post "feels off" and nobody can say which sentence is responsible
• An editor or ghostwriter needs to justify a change to an author who likes the sentence
• A team wants one pre-publish gate that every draft passes, with a documented exit code
• An earlier clean-up pass removed every dash and contraction and the result reads stiff
• Several past posts are available and drafts keep drifting away from how the author writes

Inputs the skill expects

• The draft as plain text (a .txt file or stdin). Formatting should be what will be pasted
• Optional: three to ten of the author's own past posts, separated by --- lines, for the fingerprint
• Optional: the facts the draft is missing (a figure with what it counts, a name, a date). The skill asks for them; it never invents them
• Who will read the post, when the audience is unusually strict about style

Clarify First

Before editing, confirm these inputs. If any is unknown or vague, ASK — do not assume:

• Audit only, or audit and rewrite — an audit returns findings and leaves every word alone; a rewrite changes the author's text and needs their sign-off on meaning
• Whether past posts are available — with a fingerprint, the author's real habits (their dashes, their fragments) are waived; without one, every reduce-tier rule applies at face value and the edit may strip their voice
• The facts behind vague sentences — when the audit reports "nothing only the author could know", the only honest fix is a real figure or name from the author; the agent must not supply one
• How strict the audience is — a general feed needs the remove and reduce tiers; an audience that hunts for generated text may justify acting on the review tier too

Stop rule: ask only the 2-3 that most change the output. If the user says "just clean it up," run the remove and reduce tiers, change nothing in the review tier, and list every assumption and every unfilled fact at the top of the reply.

Workflows

Quick start: run Workflow 1 on the draft, fix what it reports, and re-run until the exit code is 0.

Workflow 1 — Audit a draft before it is published

1. Save the draft exactly as it will be pasted, including line breaks.
2. Run the tell audit. Read the remove tier first: those hits block the gate on their own and usually mean a paste accident, not a style problem.
3. Run the emoji audit. It scores placement, not presence.
4. For each reduce-tier finding, read the seen: lines and decide per instance. The allowance exists because one contrast or one triple is ordinary writing.
5. Validate the result: the gate must pass (exit 0) before the draft is handed back. If it fails only on rule RD-16, stop and get the missing detail from the author.

python3 tools/linkedin/linkedin-humanizer/scripts/tell_audit.py \
  --input tools/linkedin/linkedin-humanizer/assets/sample_draft.txt --why

python3 tools/linkedin/linkedin-humanizer/scripts/emoji_audit.py \
  --input tools/linkedin/linkedin-humanizer/assets/sample_draft.txt

The sample draft fails with four blockers and a habit load of 29 against a limit of 6. assets/sample_draft_edited.txt is the same post after Workflow 2 and passes both tools.

Workflow 2 — Rewrite without leaving edit marks

1. Clear the remove tier completely, then re-run. Removing a pasted preamble changes which line is the opening, and the opener rule only inspects line one.
2. Work paragraph by paragraph, worst first. Where RD-19 fired, rewrite the paragraph from its underlying fact; replacing words one at a time yields the same empty paragraph in different vocabulary.
3. Fix each remaining reduce-tier hit with the move in references/rewrite-playbook.md. Excess dashes become commas, colons or brackets, never full stops: splitting at a dash manufactures fragments.
4. Ask the author for any missing figure, name or date. Leave a visible gap in the reply if they have none.
5. Run the over-edit checklist (playbook §6): no new fragments, no added candor phrases, contractions and at least one long sentence still present.
6. Re-run the audit with --tier review and confirm the gate passes.

python3 tools/linkedin/linkedin-humanizer/scripts/tell_rules.py --explain RD-03

python3 tools/linkedin/linkedin-humanizer/scripts/tell_audit.py \
  --input tools/linkedin/linkedin-humanizer/assets/sample_draft_edited.txt --tier review

Workflow 3 — Build a voice fingerprint and edit against it

1. Collect at least five posts the author wrote themselves, without assistance, into one file separated by --- lines. Reshares and announcements written by a comms team do not count.
2. Build the fingerprint and save the JSON. Read the Protect lines: these are reduce-tier rules the author's own writing already exceeds.
3. Audit the draft with --voice. Protected rules are still displayed, marked as the author's habit, and add nothing to the load.
4. Compare the draft to the fingerprint. Drift on three or more measures means the draft is in someone else's voice even if no rule fired.
5. Record the result in assets/voice_profile_template.md so the next draft starts from it.

python3 tools/linkedin/linkedin-humanizer/scripts/voice_fingerprint.py \
  --input tools/linkedin/linkedin-humanizer/assets/sample_past_posts.txt \
  --compare tools/linkedin/linkedin-humanizer/assets/sample_draft.txt

python3 tools/linkedin/linkedin-humanizer/scripts/tell_audit.py \
  --input tools/linkedin/linkedin-humanizer/assets/sample_draft.txt \
  --voice tools/linkedin/linkedin-humanizer/assets/sample_voice.json

Decision frameworks

The three tiers

The tiers are a statement about evidence. A remove-tier hit is proof of a paste accident. A reduce-tier hit is a pattern that only means something in quantity. A review-tier hit is a taste question, and treating taste as proof is how good sentences get deleted. All allowances and weights are editorial heuristics.

What to do with a finding