Removing AI Tells

Iterative passes over a draft, removing the patterns that make a fatigued reader think "a machine wrote this" before they consciously work out why.

Announce at start: "Removing AI tells from $FILE. Working on a copy, three passes maximum."

Arguments:

• $1 — path to the file to clean

Example:

/slushpile:removing-ai-tells applications/Acme/Engineering/Staff-SRE/cover_letter.txt

Why This Is Separate From the Voice Agent

The voice agent writes in a person's voice. That is a different job from hunting a specific list of patterns, and combining them produces worse results at both.

A voice agent asked to "write well and also avoid these forty patterns" will trade one against the other silently. Running the passes separately means each one has a single objective, and the gatekeeper between them can see exactly what changed and why.

The gatekeeper is the point. About a third of what a tell-hunting pass flags is doing real work in the document. Applied without review, this process reliably makes prose worse — flatter, more uniform, and stripped of exactly the moves that made it sound like a person. That is the failure mode this skill is built around, not an edge case.

Prerequisites

• The file to clean
• The voice agent named in preferences.yaml under voice.agent (default: aaddrick-voice)
• The user's voice agent definition. A habit documented there outranks this entire checklist.

The Process

1. Work on a copy

Never edit the original.

cp cover_letter.txt cover_letter.v2.txt

Keeping the previous version is what lets you revert a rejected change precisely rather than re-editing from memory.

2. First pass

Send the copy to a fresh instance of the voice agent with the checklist below. Ask it to edit the file directly and return a summary of every change with its rationale.

The rationale is not optional. A change you cannot evaluate is a change you have to accept or revert blind.

3. Gatekeeper review

Read the updated file. For each change: accept or reject.

Accept when the change removes a real tell without losing meaning or voice.

Reject when:

• The flagged pattern was doing structural work — a transition between sections, a precision qualifier that changes meaning, a section opener that motivates what follows
• The removal changes the meaning or weakens the argument
• The fix introduces a different problem
• The pattern is documented in the user's voice agent as one of their own habits

Revert rejected changes directly. Write down why you rejected each one — the next pass gets that list and will otherwise make the same change again.

4. Feedback pass

Send remaining issues to a new voice agent instance. Always fresh, never resumed: an agent that has already defended a change will defend it again rather than re-evaluating it.

Include what needs fixing, and what the previous pass got wrong.

5. Stop at three passes

Iterate steps 3 and 4. Where you and the agent disagree, the gatekeeper wins.

Three passes is the ceiling. Past that the changes are churn — synonym swaps and sentence reorderings that trade one phrasing for an equivalent one. If the draft still reads as machine-written after three passes, the problem is the content, not the phrasing, and no amount of editing at this level reaches it.

6. Punctuation pass, if the voice calls for it

Check the user's voice agent first. If it documents an absence — no em dashes across the whole corpus, for instance — run a dedicated pass replacing every instance with a contextual rewrite: commas, periods, semicolons, colons, parentheses, or a restructured sentence.

If the voice agent shows the user does use em dashes, skip this step entirely. Removing a punctuation mark the person actually uses moves the prose away from their voice, which is the opposite of the objective.

7. Check the replacements

Verify each replacement preserves the grammatical relationship. Flag weak ones:

• A comma creating a restrictive appositive where the original was nonrestrictive
• A colon overloading a sentence that then pivots with "but"
• A period splitting a clause that depended on the one before it

Send flagged replacements back with an explanation of the relationship that needs preserving.

The Checklist

Send this to the voice agent on each pass.

Vocabulary

Watchlist: landscape, paradigm, ecosystem, leverage, robust, comprehensive, crucial, facilitate, utilize, streamline, underscore, delve, harness, illuminate, bolster, tapestry, realm, beacon, cacophony.

Also: formal register that does not match the rest of the piece, and generic language where a specific example would land harder.

Sentence and paragraph patterns

• Uniform sentence length. The strongest signal there is. Human writing varies; generated writing converges on a comfortable middle.
• Repetitive structures, especially subject-verb-object in sequence
• Overly neat parallelism in consecutive sentence pairs
• Triplet structures used repeatedly as a rhetorical device. One or two in a piece is natural.
• Participial phrase endings: "The update shipped, revealing a deeper issue."
• Passive voice where active is more natural
• Lists where every item follows the same syntactic template
• Even, unvaried paragraph length

Rhetorical devices

• The "it's not X, it's Y" reframe, and its compressed variants: "Not X. Y." as a two-sentence punch, "X, not Y" as a comma correction, "X isn't just Y" as a setup
• "Not X" as emphasis: "Not a benchmark number." State the positive claim.
• Staccato fragment pairs as punch. One fragment is fine; two or more back to back is a rhythm device.
• Paired beats: "That's X. It's also Y."
• Announcing frames: "Here's the mechanism that makes X powerful:"
• Dramatic openers: "The takeaway:", "The bottom line:"

Filler and hedging

• Balanced or hedged language where the author would commit to a position
• Transition filler: Moreover, Furthermore, Additionally, It's worth noting
• The whole "it's worth [verb]-ing" family: worth noting, worth pausing on, worth sitting with
• "It's important to note that", "generally speaking", "to some extent", "from a broader perspective"
• Authority claims: "Let's be clear", "To be sure", "The reality is", "Make no mistake"

Structural redundancy

• Bookend summaries that restate the intro instead of advancing
• Sentences that summarize what was just said
• Summary openers: Overall, In summary, In conclusion
• "From X to Y" scene-setting: "From simple scripts to full pipelines..."
• Cross-section repetition

Significance labeling

• "That's the gap between policy and reality." If the fact is strong, it lands without a label.
• "That changes the framing considerably." Let the fact do it.
• Vague gestural conclusions: "says a lot about where they are", "is the most telling thing about". Either say what it tells you or let it stand alone.

Post-edit verification

Two errors that only appear after editing, and are worth a dedicated read:

• Orphaned pronouns. "It builds..." where the referent was in a sentence that got cut.
• Broken ordering. Text claiming A happens before B when the surrounding edits reversed them.

Gatekeeper Principles

Not every pattern on the checklist is always wrong. Telling the two apart is the whole job.

Real tells — remove:

• Section openers that restate the intro
• Announcing frames before content that speaks for itself
• Significance labels after a paragraph that already landed
• Filler qualifiers
• Watchlist vocabulary where a plainer word works

Structural work disguised as a tell — keep:

• Transitions that connect sections. "That's the gap these files close." looks like a significance label and is carrying the reader between two ideas.
• Precision qualifiers that change meaning. "structurally" before an analogy is not filler.
• Section openers that motivate what follows.
• Colons that set up a mechanism.
• Deliberate scope-escalation fragments. "$30M CAPEX. Four facilities. Seven clients." This is a human writing move and frequently the strongest line in a cover letter. A tell-hunting pass will flag it every single time. Reject it every single time.
• Punchy inventory lists that read clipped on purpose. Adding the missing subject to each item kills the delivery.
• Anything the user's voice agent documents as their habit. The voice agent outranks this checklist without exception. Someone whose corpus is full of triplets keeps their triplets — the pattern is a tell in general and a fingerprint in their case.

Prompt Template

Read [file path].

Review the entire draft for AI tells:

[paste the checklist]

Edit [file path] directly. Be surgical. Change only what needs changing, and do
not rewrite sections that already sound human. Preserve the argument structure
and every factual claim.

Do not flag or change:
[the rejections from the previous pass, with reasons]

After editing, summarize every change and why you made it.

Anti-Patterns

• Do not run this without a gatekeeper. An unreviewed pass makes prose flatter and more uniform, which is the same direction as the problem it is solving.
• Do not run more than three passes. After three the changes are churn.
• Do not resume a voice agent instance between passes. It will defend its previous changes rather than re-evaluate them.
• Do not run the punctuation pass without checking the voice agent first. Removing punctuation the person genuinely uses moves the prose away from their voice.
• Do not apply this to a resume the same way as prose. A resume bullet is a compressed form and legitimately looks patterned. Use /slushpile:application-builder's humanization step for resumes; use this for letters and prose.
• Do not let this run on quoted material. Anything in quotation marks, any job description text, and any product name stays verbatim.