Craft Library
LLM tics & tells — the de-LLM catalog
Companion to the Craft Glossary — the named sentence-level habits that make prose read as machine-written. Each tell below links to a full explainer; this page is the quick catalog with BAD → GOOD examples you can self-audit against. The studio counts these deterministically (
./run.sh tics) and hunts them in the de-LLM loop — not to ban rhetoric, but to restore modulation: plain where plain belongs, sharp where sharp earns it.
0. What a "tell" is
A tic is a small recurring habit — yours or the model's. An LLM tell is a tic so common in AI prose that readers feel the machinery even when they cannot name it.
The deepest tells are not individual words. They are statistical habits:
- mathematically even cadence (every sentence the same weight)
- uniform gravitas (nothing allowed to be throwaway)
- prose from nowhere (no situated speaker with a stake)
- compulsive over-significance (everything "underscores" something)
The word-lists below are symptoms. Fix the rhythm and stance; the words often fall away on their own.
→ Full theory: Tic / machine-tell / LLM tell · Evenness of register · The root cause: no situated speaker
1. The headline tells (hunt these first)
"Not X. Y." / "Not X, but Y." — the reframe
The model's default way of sounding deep. Powerful once; a tell in bulk (~6% of ChatGPT output uses a "not just X, but Y" variant).
BAD: It wasn't fear. It was recognition.
GOOD: Her hands stopped shaking. She knew that face.
BAD: Not just a tunnel — a machine.
GOOD: The tunnel was a machine.
BAD: No panic, no flinch — just focus.
GOOD: She focused.
→ Negative parallelism / the reframe · Reframe device
Em-dash addiction
The dash doing a comma's or full stop's job — every sentence interrupted, nothing allowed to end cleanly.
BAD: She opened the door — the hinge screamed — and stepped inside — the air cold — older than the stone.
GOOD: She opened the door. The hinge screamed. The air inside was cold and old.
BAD: He wasn't angry — not exactly — more tired than anything — though tired wasn't the word.
GOOD: He looked tired. That wasn't the word, but it would do for now.
Rule: Reserve the em dash for genuine interruption or apposition. Track density per book.
"The way…" — the simile template
Often the single most over-used construction in AI prose: the way a [noun] [verbs], repeated until every comparison sounds stamped from the same die.
BAD: The silence settled the way dust settles on unused furniture.
GOOD: The silence settled.
BAD: He moved the way a man moves when he already knows the answer.
GOOD: He moved like he already knew.
BAD: It hit her the way bad news always hits — slow, then all at once.
GOOD: Bad news. She felt it in her teeth first.
Rule: Cut roughly half. Keep the ~40% where the comparison only that comparison could carry the meaning. Else state the thing directly.
Even cadence / gravitas inflation
Every sentence the same length and weight; every moment made profound; nothing plain.
BAD: The chamber was ancient. The stone was patient. The air was heavy with meaning.
GOOD: The chamber was old. Dust. A drip somewhere she couldn't see.
BAD: She understood, in that moment, the true cost of knowing.
GOOD: She wished she didn't know.
→ Cadence · Gravitas / evenness of register · Modulation
2. Vocabulary & construction tells
- AI vocabulary cluster — delve, tapestry, underscore, pivotal, nuanced, landscape, foster, nestled… → Rewrite for specificity; not a banned-word purge. → The "AI vocabulary" cluster
- Rule of three — three adjectives, three parallel clauses, everywhere → Break the rhythm. → Rule-of-three overuse
- Verb-of-significance — serves as, stands as, represents → Plain is/was; dramatize instead. → "Serves as / stands as"
- Caption tell — show, then explain what you showed → Trust the reader. → Over-explanation
- The hedge — ultimately, in many ways… → End on the image, not a summary. → The hedge
- Generic over specific — smooth abstraction replaces the odd true detail → Restore the precise noun. → Generic over specific
- Frictionless transitions — every paragraph topic-sentence → support → tidy close → Allow hard cuts. → Frictionless transitions
3. The self-audit (one pass)
Read a chapter aloud. Flag any paragraph where:
- More than one "Not X…" reframe appears
- More than two em dashes in five lines
- "The way" appears twice on the same page
- Every sentence is roughly the same length
- A line states what the previous line already showed
Cut or rewrite the worst three instances. Re-read. Stop when only earned rhetoric remains.
→ The one revision rule · Anti-Patterns (structural habits above the sentence)
4. What the workshop does with tells
The engine does not silently rewrite your voice. It:
- counts each tell band against per-book targets
- flags clusters in finish reports and cold-reads
- blocks only when craft gates fail (continuity, mythos — separate from tics)
- leaves surgical cuts to the author (or author-approved edit passes)
That is why a Rothfuss-tier manuscript can run through the pipeline and still sound like Rothfuss — the tells are measured and thinned, not replaced with generic "humanizer" slop.
→ The workshop — for authors & editors · The technology (de-LLM loop)
Back to the Craft Glossary · Craft Library overview
Overview · Glossary · Doctrine · Anti-patterns · Triptych form
Arjuna Badger Press