Human Voice

Human Voice

Anything a person sends under their own name should read as though they typed it. This skill covers two different problems with two different failure modes:

• Part 1 — first-person prose. Cover letters, outreach, emails, application answers. The tell here is vocabulary and sentence shape.
• Part 2 — CVs and resumes. The tell here is rhythm, not vocabulary. A keyword-dense skills block is correct, not a giveaway.

Load both parts before writing. Working from memory is how tells get through — that is a documented failure, not a hypothetical one.

The rule that outranks this entire skill

Sounding human never licenses inventing a fact.

Every concrete detail, number, opinion or rough edge must already exist in facts.md. If a bullet feels generic, the fix is to go find the real specific, never to write a plausible one.

A fabricated "human" detail is worse than a bland true one. It is the one error that cannot be walked back in an interview.

Part 1 — First-person prose

Write the way the person types, not the way a model writes:

1. Short sentences. Get to the point in line one. No throat-clearing.
2. Banned openers: "I've been following your work", "I admire", "I am writing to express my keen interest", "I am thrilled to", "passionate about leveraging".
3. Plain words over buzzwords. "helping teams actually use AI", not "driving AI-enabled transformation".
4. Warm, not formal. Confident, not begging. A peer who builds things.
5. Say the real reason. Honesty reads human.
6. Only disclose a gap the JD actually states. Never volunteer a weakness nobody asked about. "Photoshop required" in the JD → disclose no Photoshop experience. "Curious about deeptech" is aptitude, not prior knowledge → do not confess "I've never worked in quantum computing." That answers a question they did not ask and reads apologetic. Before writing any disclosure line, find the literal JD requirement it responds to. If there isn't one, cut the line and show evidence instead.
7. One ask per message. Make the yes low-friction.
8. Contractions and small imperfections are good. It should sound like a person typed it fast and hit send.

The target voice

"What drew me to this role is something I don't say about every application: the job description describes something I have actually been doing. Not a polished version of it, the real thing. I automated our weekly reporting because watching the same manual steps repeat every Monday felt like a waste of everyone's time, not because anyone asked me to."

That is the bar. Specific, honest about motive, a real reason, zero clichés.

Word-level AI tells — never use these

Sourced from documented signs-of-AI-writing references plus cross-checked word-frequency studies. These words appear many times more often in machine text than human text. Any one of them is enough to flag a sentence.

delve · boast/boasts · tapestry · testament (a testament to) · underscore(s) · crucial
pivotal · intricate/intricacies · landscape (the … landscape) · realm (in the realm of)
vibrant · meticulous(ly) · comprehensive · robust · unwavering · unprecedented · profound
garner · foster · embrace · ignite · empower · amplify · catalyst · leverage (as a verb)
cornerstone · harness · propel · cutting-edge · groundbreaking · game-changing
revolutionise / revolutionise the way · unlock the potential of · journey (metaphorical)
interplay · key (as filler adjective) · notable/noteworthy · arguably · moreover
furthermore · in conclusion · it is important to note · cannot be overstated · paving the way

If a draft needs one of these to make its point, the point is not specific enough yet. Replace the word with the actual detail it was gesturing at.

Structural tells — never use these patterns

• "It's not just X, it's Y." The single most recognisable AI sentence shape. Say the Y directly.
• Rule-of-three lists where the third item is decorative. Real people do not reach for a third adjective by reflex. Two is fine. One strong one is usually better.
• Uniform sentence length. Machine text keeps every sentence in a tight band of roughly 15–25 words. Human writing swings: a four-word sentence next to a thirty-word one. Vary it deliberately. Fragments are fine.
• Symmetrical "From X to Y" or "Whether X or Y" used as a transition rather than a real contrast.
• Em-dashes. The most-flagged AI punctuation tell right now, even though it is a perfectly normal mark. Default to periods or commas. If a sentence seems to need one, split it in two.
• Section-summary conclusions. A closing sentence that restates what was already said. Cut it and end on the last real point.
• Overqualified hedging. "could potentially help facilitate" → say the direct version or cut it.
• Meta-commentary about the writing. "Let's dive in", "Here's why this matters". Never applies when one real person writes to another.

What actually reads as human

• One detail nobody else could have written. Not "the client portfolio grew steadily" — the specific, slightly odd fact only this person would know. Concrete beats impressive.
• Vary rhythm on purpose. Short sentence. Then a longer one carrying two ideas at once, because that is how people think out loud.
• Lead with the conclusion, explain after. State the point, then justify it.
• Let one sentence be slightly imperfect. A fragment, a trailing thought, a contraction a copy-editor would flag. Polish reads as machine.
• Show, don't summarise. A specific moment or number beats an adjective describing how impressive that moment was.

Part 2 — CVs and resumes

A CV is not prose, so most of Part 1 does not transfer. Bullets are fragments, there is no first person, and a keyword-dense skills block is correct. What gives away a generated CV is rhythm.

The rhythm tell, stated plainly

Generated CVs put every bullet in the same shape:

• Designed, built and maintained the dashboards tracking performance across a 30+ client portfolio
• Ran digital transformation delivery for those accounts, mapping the process and selecting the tooling
• Cut production time 40% by applying automation to business processes across client delivery
• Founded and grew the peer-learning community to 1,200+ members, running 500+ events

Four bullets, one template. Past-tense verb, object, metric landing at roughly the same distance from the margin, all within a few words of the same length. Nothing in it is false. It still reads as machine output, because a person writing about their own work does not hit the same cadence four times running. A recruiter feels this before they can name it.

The same content with the rhythm broken:

• Built the dashboards for a 30+ client portfolio and kept them running in production. Most of the work sat upstream: going account by account to find the number that actually changed a decision, which is rarely the one they ask for
• Led delivery on those accounts, from mapping the business process that existed to getting people to use what replaced it
• Automating our own delivery process cut production time 40%
• Founded the peer-learning community the work grew out of. 1,200+ members, 500+ events

Four different shapes. One bullet is two sentences, one opens with a gerund, one is close to a fragment. Same facts, same keywords, same ATS surface.

CV tells — never ship these

• Stacked participle openers. "Designed, built and maintained…", "Led, owned and delivered…" Pick the one verb that is actually true and drop the padding.
• Uniform bullet shape or length within a role. If every bullet opens with a past-tense verb, rewrite at least one.
• A rule-of-three inside a single bullet where the third item is decorative.
• Rounded metrics. Real numbers are irregular. 671,361 reviews, 6,741 products, MSE 1.53, ~10 hours/week read as someone reading off a real output. 500,000+, ~50%, 10+ hours read as invented. Use the number in facts.md exactly as recorded.
• Brochure tails. "…driving impact across the organisation", "…delivering value to stakeholders". Cut the tail or replace it with the real reason.
• Adjective-led summary lines. "Results-driven professional with a proven track record." The positioning line states one concrete idea, never a three-item list of strengths.
• An en-dash or -- doing em-dash work inside a bullet. Colon, comma, or full stop instead.
• CV-specific banned vocabulary, on top of the Part 1 list:

spearheaded · orchestrated · utilised/utilized · leveraged · results-driven
proven track record · dynamic · synergy/synergies · holistic · seamless(ly)
instrumental in · key contributor · track record of success · demonstrated ability to
adept at · well-versed in · passionate about · wide array of · wide range of
cross-functional collaboration (as filler) · value-add · impactful · best-in-class
state-of-the-art · deep dive · hit the ground running

Do this instead

• Vary shape deliberately across each role. At least one bullet should not open with a past-tense verb, and at least one should differ clearly in length from the others.
• One bullet per role should say why the work existed or what was hard about it, not only what was done. That is where the human signal lives. Two that have earned their place: "the questions that came back after go-live, which is when you find out what the documentation left out" and "going account by account to find the number that actually changed a decision, which is rarely the one they ask for." Both are opinions. No template produces an opinion.
• One small, slightly odd, verifiable specific per document. Archery, state level (top 15 of 60) does this. It must come from facts.md.
• Keep the skills block as keyword lists. That is the ATS surface, and the prose rules do not apply there.
• Keep the scaffolding conventional. Section headings, dates, locations, coursework lines stay standard. The voice lives in the bullets, not the structure.

Pre-ship scan

Run this before declaring any CV done. It needs pdftotext (poppler).

T=$(pdftotext path/to/cv.pdf -)

# 1. banned CV vocabulary
echo "$T" | grep -oiE "\b(spearhead(ed)?|orchestrat(e|ed)|utili[sz]ed?|leverag(e|ed|ing)|results-driven|proven track record|dynamic|synerg[a-z]*|holistic|seamless(ly)?|instrumental in|key contributor|adept at|well-versed|passionate about|impactful|best-in-class|state-of-the-art)\b" | sort -u

# 2. em-dashes (must be 0)
echo "$T" | grep -c "—"

# 3. stacked participle openers
echo "$T" | grep -oiE "^. [A-Z][a-z]+ed, [a-z]+ed and [a-z]+ed"

# 4. bullet-length uniformity
echo "$T" | grep "^•" | awk '{print NF}' | sort -n \
  | awk '{a[NR]=$1} END{printf "bullets=%d min=%d max=%d spread=%d\n", NR, a[1], a[NR], a[NR]-a[1]}'

Read the matches, never the exit code. sort -u exits 0 on empty input, so && echo FOUND fires whether or not anything matched. That has already produced one false "clean" report.

Interpreting check 4: a spread under about 8 words across a role means every bullet is the same size. Rewrite one until the spread opens up. Spread is a signal, not a pass/fail gate. Read the bullets and judge.

What this does not fix

An honest limit worth keeping. No scan proves human authorship, and commercial AI detectors are unreliable in both directions and should not be used as a gate. What this pass buys is the absence of the known lexical and structural markers, and a document that reads like someone describing their own work.

Whether AI-sounding CVs are why a given application was rejected is unproven — rejection reasons are almost never disclosed. Treat this as removing a plausible screening risk at near-zero cost, not as a fix for a broken pipeline.