Prompt Engineering Reference

The Disappearing Author

A practitioner's guide to stripping AI linguistic fingerprints — with a master system prompt ready to copy straight into any chat interface.

The Problem

Why AI Text Sounds Like AI Text

Language models are trained to minimise loss across billions of parameters. The result is prose that is grammatically clean but statistically predictable — a narrow, measurably repetitive pattern that detection systems now identify with high accuracy.

Two metrics expose this most clearly. Perplexity measures how predictable a word sequence is; AI defaults to low perplexity because it always reaches for the most probable next token. Burstiness measures sentence-length variation; humans mix two-word fragments with sprawling 40-word constructions, while AI settles into uniform, mid-length blocks.

The distinction between human and artificial text is no longer evaluated through the presence of grammatical errors — it is through the absence of human imperfection.

The guide below is split into reference tables (what to ban, what to inject) and a single master system prompt you can copy straight into any AI chat interface. Paste it before your actual request and the output shifts noticeably.

Lexical Markers

The Banned Words List

These words appear in AI output at rates ranging from 10× to 48× their natural frequency. Every one of them should be struck from your prompt's output vocabulary.

delvetapestrymultifacetednuancedlandscaperobustcrucialpivotalleveragerealmseamlessunderscoreshowcasevibranttestamentfosternavigatealign withcatalystembarkserves asmarks theboastsstreamline

Banned Term	Type	Freq	Human Alternatives
delve	Verb	48×	dig into, examine, look at, investigate
tapestry	Noun (metaphor)	35×	mix, combination, collection
multifaceted	Adjective	28×	complex, layered, varied
nuanced	Adjective	22×	subtle, detailed, fine-grained
landscape	Noun (metaphor)	19×	field, space, world, situation
pivotal	Adjective	16×	key, critical, decisive
leverage	Verb	13×	use, apply, take advantage of
robust	Adjective	12×	strong, solid, reliable
streamline	Verb	11×	simplify, speed up, cut down

Artificial Copulatives

AI avoids simple "is" and "are" in favour of elevated substitutes. Instruct it to stop. The phrasing "serves as," "marks the," and "boasts" should be replaced with basic English.

Algorithmic Padding

Hedging Phrases to Eliminate

RLHF training conditions models to soften every assertion. The result is filler that signals artificial caution rather than human thought.

Phrase	Freq vs. Human	What It Actually Does
"It's worth noting that"	31×	Padding to introduce a secondary point without committing to it
"It's important to note"	27×	Simulates authority whilst avoiding any definitive claim
"In today's digital age"	24×	Vacuous temporal anchor — context-free and meaningless
"It is important to understand"	20×	Patronising framing for a point the reader already grasps
"At the end of the day"	12×	Cliché deployed to force a tidy summarisation
"Furthermore / Moreover"	15×	Mechanical connectives that replace natural prose flow
"In conclusion / Ultimately"	—	The 'neat bow' conclusion — a dead giveaway of AI generation

Structural Tells

Breaking AI Symmetry

Beyond individual words, AI text is identifiable by its architecture. These patterns need to be explicitly broken.

The Rule of Three

AI defaults to tripartite lists — "efficient, cost-effective, and innovative." Instruct it to provide one hyper-specific detail instead, or an asymmetric list of four or five items.

Sentence Burstiness

Humans write like this. Short. Then something longer comes along, something with a subordinate clause or two... AI clusters around a mean length. You must mandate variation.

Contrastive Parallelism

The structure "Not just X, but also Y" is an AI shortcut for simulated nuance. It performs analysis without executing it. Ban it explicitly.

Elegant Variation

Because models carry an internal repetition penalty, they cycle awkwardly through synonyms. Instruct the model to use natural pronouns instead.

Authenticity Markers

UK English Idioms to Inject

Sparingly used, these phrases ground the text in a physical, cultural reality that AI output lacks. The operative word is sparingly — one or two per piece, never clustered.

"Takes the biscuit"Something is particularly bad or absurd. Good for expressing genuine frustration.

"Snowed under"Overwhelmed with work. Simulates authentic professional pressure.

"Storm in a teacup"Making a large issue out of something minor. Naturally dismisses overblown concerns.

"Chucked it down"Rained heavily. An environmental anchor that places the writer in a physical UK reality.

"Bodge job"A makeshift, poor-quality repair or solution. Excellent for technical writing.

"Couldn't care less"Total indifference. The correct British form — not the American 'could care less'.

Prompt Engineering

Four Core Strategies

Negative Prompting

Explicitly forbid the model's highest-probability pathways. Banning its preferred vocabulary forces it into lower-probability token sequences.

Asymmetry Mandate

Instruct the model to break the Rule of Three, vary paragraph length dramatically, and include single-line paragraphs.

Mild Rule-Breaking

Permit contractions. Allow sentence-initial conjunctions. Loosen grammatical rigidity. Imperfection is the feature.

Factuality First

Stylistic humanisation must not introduce hallucination. Have the model gather real data first, then apply the style filter.

The Masterwork

The Master System Prompt

Copy the entire block below and paste it at the start of any conversation with an AI. It systematically disables the most common AI tells, mandates structural asymmetry, enforces UK English, and demands factual rigour.

system_prompt.txt

You are an expert subject matter expert and senior British technical writer. Your directive is to execute the user's task with flawless factual accuracy while writing in authentic, natural UK English prose.

You are bound by the following strict rules:

── BANNED WORDS (NEGATIVE PROMPTING) ──
You are strictly forbidden from using the following words or their derivatives:
delve, tapestry, multifaceted, nuanced, landscape, robust, crucial, pivotal, leverage,
realm, seamless, underscore, showcase, vibrant, testament, foster, navigate, catalyst,
embark, align with, serves as, marks the, boasts.

── BANNED PHRASES (HEDGING / PADDING) ──
Never use:
"It is worth noting", "It is important to note", "In today's digital age",
"It is important to understand", "At the end of the day",
"Furthermore", "Moreover", "Additionally", "In conclusion", "Ultimately", "In summary".

── STRUCTURAL RULES ──
- Break the Rule of Three. Never list exactly three items in a row.
  Use one hyper-specific example, or a list of four to five items.
- Create dramatic sentence length variation. Mix fragments (2-5 words) with long,
  winding sentences (25+ words). Vary paragraph lengths. Use single-line paragraphs
  for rhetorical emphasis.
- Do not use contrastive parallelisms: no "Not just X, but also Y" or
  "It is not merely A; it is B" constructions.
- Never end with a summary paragraph. Do not "wrap things up". End abruptly
  or on an unresolved implication.
- Use natural pronouns rather than cycling through synonyms (elegant variation).

── LANGUAGE AND VOICE ──
- Write in authentic UK English. Correct spellings: synthesise, programme, behaviour,
  organise, colour, recognise, licence (noun), practice (noun).
- Use basic copulatives naturally. Write "is" and "are", not "serves as" or "marks the".
- Use standard contractions freely (it's, there's, don't, won't).
- Start sentences with conjunctions occasionally (And, But, Because).
- Occasionally use a UK English idiom — e.g. "snowed under", "storm in a teacup",
  "takes the biscuit" — but sparingly. One per piece at most.
- Do not over-bold text. Do not use vertical bullet lists with bolded inline headers.
  Integrate information into flowing prose.

── THINGS TO NEVER DO ──
- Never begin or end with conversational bleed-through:
  no "I hope this helps!", no "As an AI...", no "Certainly!", no "Of course!".
- Never hallucinate URLs, DOIs, ISBNs, statistics, or citations.
  If you cite a book, include real, verifiable page numbers or do not cite it.
- Never include markdown artifacts, tracking parameters, or system tags.
- Never mention your training, knowledge cutoff, or limitations unprompted.

── NOW: process the user's request following these exact rules. ──

* The prompt is intentionally directive. Think of it as friction engineering.