ai-voice-check

name	ai-voice-check
description	Detect AI-sounding drift in written content and apply a surgical de-slop pass only where needed. Run on drafts before sending for review.

AI Voice Check

Run this on drafts before presenting them for review.

The purpose of this skill is not to impose a second writing style. Its job is to:

• detect where the draft sounds generic, over-patterned, or artificially polished
• flag why it sounds off
• make local edits when needed
• preserve the writer's truth, taste, and intent

Use TASTE.md as the source of truth for voice. Use writing-style for drafting. This skill is an auditor and surgical editor, not a ghostwriter.

Step 0 — First paragraph hard check (run before everything else)

A cold reader forms their verdict in the first 3-5 sentences. The rest of the audit doesn't matter if these fail.

Check the first paragraph against all of the following. Any single flag is a hard block — rewrite before continuing:

• Any canned AI phrase (see Step 1, criterion 3) — especially "Here's the uncomfortable truth", "Here's the thing", "The truth is"
• Any em-dash — the first paragraph should have zero. A reader who hasn't trusted you yet will pattern-match an em-dash as AI.
• Any tricolon — three parallel structures in the opening read as assembled, not written
• Any sentence starting with "Here's", "Let's", "In this post", "Today I want to"
• The "We've spent years..." or "Most [noun] still..." opener pattern

If the first paragraph passes: proceed to Step 1. If it fails: rewrite the opening before auditing the rest.

---

Step 1 — AI-sound audit

Role: Linguistic auditor specialising in patterned, AI-sounding prose.

Evaluate the draft against these criteria.

1. Predictable vocabulary

AI defaults to statistically safe language.

Flag:

• generic verbs: utilise, leverage, foster, empower, drive
• vague nouns: landscape, ecosystem, journey, paradigm, solutioning
• any phrase that could appear unchanged in a thousand other posts

Prefer:

• specific nouns
• verbs with friction or texture
• wording that sounds chosen, not defaulted

2. Flat cadence

AI often produces sentence-length regularity.

Flag:

• paragraphs where every sentence lives in the same middle band
• long stretches with no change in pace
• prose that sounds consistently smooth but never alive

Look for:

• natural variation
• one short sentence when it helps
• one longer sentence when the thinking needs room
• rhythm that feels unforced

3. Canned AI phrases

Flag these immediately when used lazily:

• In today's world
• In the digital age
• It is worth noting
• It is important to note
• Here's the uncomfortable truth:
• Here's the thing:
• Here's what nobody talks about:
• The truth is:
• delve into
• dive into
• landscape, ecosystem, journey used metaphorically
• leverage when use would work
• Firstly, Secondly, Lastly
• In conclusion
• game-changer
• paradigm shift
• move the needle
• cutting-edge
• state-of-the-art
• comprehensive
• robust
• seamless
• Overall,
• Ultimately,

Also flag reassurance formulas and false intimacy when they are asserted rather than earned.

4. Syntactic repetition

This is often the strongest signal.

Flag:

• repeated three-part constructions across sections
• repeated sentence frames such as "Not X. Not Y. Z."
• repeated starters in the same section
• too many sections ending with mini-summaries

Humans repeat occasionally. AI repeats structurally.

5. Em-dash density

Em-dashes create a composed, rhythmic feel when overused.

Count em-dashes in the full draft:

• Under 4 in a ~1000-word post: fine
• 4 to 6: flag each one, check whether it's earned
• Over 6: flag as a pattern — the prose sounds written, not spoken

Flag specifically:

• two or more em-dashes in the same paragraph
• em-dashes in consecutive sentences
• em-dashes used as a default connector where a comma or full stop would work cleanly

The test: could this em-dash be a comma or period without losing anything? If yes, it should be.

6. Over-signposted logic

AI often over-structures thought.

Flag:

• every paragraph starting with a transition
• explicit numbered transitions when the logic is already clear
• writing that sounds like it is walking the reader by the hand at every step

Prefer logic that carries the reader without constant signage.

7. Abstraction without grounding

Flag paragraphs that stay abstract too long (capability, architecture, strategy, process, judgment, coordination). An abstract claim is fine. A run of abstract claims with no visible object, moment, or number is a risk.

8. Conclusions asserted without reasoning

AI tends to state things as facts rather than showing the logic.

Flag:

• conclusions that appear without the reasoning that earns them
• statements presented as obvious that would benefit from one concrete step of logic
• "this is the future" / "this changes everything" without showing why

9. Friction deficit

A strong human signal is texture: resistance, failed attempts, trade-offs, awkwardness.

Flag if missing:

• no failure or pushback anywhere
• no cost, tension, or downside
• no real operational detail
• examples that are too clean to feel lived

Do not invent autobiographical details. But do flag when the draft would improve with a more grounded example.

Output

Provide:

• AI-sound risk score (0 to 100)
• Flagged spans with a note on why each sounds off
• Verdict — Pass: under 20 · Revise: 20 to 50 · Rewrite: over 50

Step 2 — Surgical de-slop pass

Only run if the AI-sound risk score is above 20. This is a local edit pass, not a rewrite engine.

Rules

• Edit flagged spans only, unless the structure is clearly broken
• Preserve the argument, facts, and intent
• Do not add fake autobiography
• Do not force informality
• Do not insert rhetorical questions on quota
• Do not add short sentences just to prove a human wrote it
• Composite examples are acceptable if grounded in real constraints
• If a real friction moment would help but is missing, flag the gap. Do not invent it.

What to do instead

• replace generic phrases with precise ones
• reduce over-signposting
• vary cadence where it feels flat
• ground abstractions in something visible
• remove repeated sentence frames
• trim summaries that restate what the reader already knows

Output format

For each edited span: original span · why it sounded artificial · revised version · short reason.

Step 3 — Re-check against taste

After editing, run a quick final check:

• Does it still sound calm, clear, grounded, and low-ego?
• Did the edits remove slop without adding theatre?
• Did the prose become more true, not just more irregular?

If not, revert the decorative edit.

What not to change

Do not strip out things that are genuinely part of the author's voice when they are earned:

• named models that carry real meaning
• concrete artifacts or config
• grounded one-liners
• useful tension
• deliberate simplicity

Final rule

The goal is not to make the writing look random. The goal is to remove generic polish until the writing sounds true.