train-voice — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited train-voice (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
The most useful thing you can do for someone before they write a LinkedIn post is figure out how they actually sound. Every other AI tool skips this step and produces the same templated voice for everyone, which is exactly why LinkedIn is full of identical-sounding posts. This skill exists to fix that.
LinkedIn's 2026 algorithm rewards content that holds attention. Generic-sounding content gets scrolled past. The fastest way to sound generic is to write in the average LinkedIn AI voice. The fastest way to sound like a real person is to write in the voice the user actually writes in.
Real voice is not "professional but approachable" or any other tone-deck label. It is concrete: sentence length, word choice, how the user opens and closes, whether they use semicolons, what they refuse to say, what they get excited about. The job here is to capture that from real samples and write it down so other skills can use it.
Trigger when:
Do NOT trigger when:
Samples of their actual writing. In order of preference:
Ask once: "Do you have past posts I can read, or a LinkedIn data export? Pasting 10-15 of your best past posts works too. If you have nothing yet, I can interview you instead." Do not loop on this. Pick the best source available and go.
Read the samples carefully. Look for these specific things, in this order:
Sentence rhythm. Average sentence length. Range from shortest to longest. Does the user mix short and long, or stay even? Do they use one-line sentences for emphasis?
Opening patterns. How do their posts start? Common types: a number, a question, a quote, a confession, a contrarian claim, a story setup, a definition. Note which the user actually uses.
Closing patterns. Some users close with a question. Some with a one-line summary. Some just stop. Some have a signature phrase. Note what they do.
Word choice register. Casual ("yeah," "weird," "stuff"), neutral, formal, technical. Most people mix. Note where they land and where they vary.
Punctuation habits. Em dashes, semicolons, parentheses, ellipses, all-caps, em-spaces, line breaks. Note what they use and what they avoid.
Recurring phrases or verbal tics. Things the user says often. "the thing is," "look," specific industry jargon, particular metaphors. These are voice fingerprints.
What they refuse to do. Look at what is NOT in the samples. No emojis? No hashtags? No "thoughts?" closer? No bold text? These are also voice signals.
Topics and angles. What do they write about, and at what angle? A founder writing about product, lessons learned, the team, hiring, mistakes. Note the recurring beats.
Tells of their actual job/expertise. The specific details that prove they know what they are talking about. Numbers, named tools, dates, customer types.
A focused profile, not a dissertation. Save to voice-profile.md in the workspace root (or ~/.linkedin-maxxing/voice-profile.md if a shared location is configured).
The file should contain:
# Voice profile for [user's name if shared, otherwise "user"]
Generated [date] from [number of samples] samples of [source: LinkedIn posts / blog / interview / mixed].
## Sentence rhythm
[1-2 lines on length, mix, use of short sentences for emphasis]
## Opens
[2-4 specific opening patterns the user actually uses, with one real example each]
## Closes
[1-2 closing patterns, with example]
## Word choice
[register and notable habits, with examples of words the user uses and words they avoid]
## Punctuation
[what they use, what they avoid]
## Verbal tics and recurring phrases
[3-6 actual phrases from their writing]
## What they refuse to do
[bullet list of what is absent: emojis, hashtags, certain closers, etc.]
## Topics and angles
[3-5 recurring topic-and-angle pairs, with example post titles or one-liners]
## Tells of expertise
[2-4 things they reference that prove their domain knowledge]
## Things to NEVER do when writing in this voice
[hard nos: anything the user has explicitly said they hate, plus inferred patterns they avoid]
## Things to ALWAYS do
[2-3 things that are core to the voice, like "open with a number" or "end without a CTA"]Length target: 200-400 words. Tight enough that downstream skills can read it quickly, specific enough to actually change output.
If the user has no samples, run a short interview (4 questions, asked once):
From these answers, build a partial voice profile, marked at the top with "PROVISIONAL: built from preferences, not samples. Re-run train-voice with real samples when available."
Save the voice-profile.md to the workspace. Tell the user where it is and what it contains in 2-3 sentences. Show them the file path. Do not paste the whole file inline unless they ask.
End with: "Future writing skills will read this automatically. To update it, run train-voice again with new samples."
Every writing skill in this repo (write-post, write-comment, write-longform, write-recommendation, write-hook, write-dm) is instructed to read voice-profile.md before drafting. If the file is missing, they fall back to asking one question about tone. If it is present, they match the voice patterns you wrote down.
This means the voice profile has real consequences. Be specific, not flowery. "Mixes short punchy sentences with longer ones, often opens with a number, never uses bullet points, signature closer is a question that sounds rhetorical but isn't" is useful. "Conversational and engaging" is not.
User uploads Shares.csv from LinkedIn data export with 47 past posts.
Skill reads them, then produces voice-profile.md that includes (truncated):
## Sentence rhythm
Mostly medium sentences (15-25 words), with frequent one-line sentences for emphasis, usually after a longer setup. Average ~18 words. Rarely uses sentences longer than 35 words.
## Opens
- A specific number: "47% of our pipeline closes in the last 5 days of the quarter."
- A confession: "I shipped a feature this week I'm not sure about."
- A contrarian claim: "Most B2B PMs are wrong about onboarding."
## Closes
- Often ends with a question that is not rhetorical, like "What is your team doing differently?"
- Sometimes just stops after the strongest line.
## Verbal tics
- "Honestly" used twice per post on average, never as an opener (mid-sentence)
- "Worth knowing" as a transition into a point
- "Here is what I noticed" as a setup for a list
- Almost never uses "I think". Instead just states the claimThis file is now what every write-* skill reads before drafting a post.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.