← zachy blog

LLM Cheatsheet: signal words that actually matter

When you build software with LLMs, the words in your prompts are your primary steering mechanism. A single well-chosen operator word can swing accuracy by 10+ points. Formatting and structure shifts can swing output quality by up to 76 points (ZIP paper, arXiv 2502.03418).

Three categories do most of the work:

  1. Behavioral triggers. Verbs that tell the model how to think.
  2. Constraint markers. Words that rule out unwanted directions.
  3. Domain anchors. Single terms that pull the model into expert register.

Combine one from each. Usually 8 to 12 words. That works better than paragraphs of vague instruction.

Precision verbs (behavioral triggers)

These activate a processing mode rather than just a topic. Swapping one operator word moves output considerably. The ZIP paper found that nouns dominate importance rankings across every model tested.

Term What it triggers Stronger than
Audit Check against a standard; flag what fails review (model defaults to descriptive)
Diagnose Find the non-obvious root cause explain (descriptive, not causal)
Steelman Build the strongest opposing position counter-argument (produces strawmen)
Critique Find specifically what is weak give feedback (too open)
Enumerate Exhaustive list, every item list (implies a few is fine)
Extract Pull specific data, no invention find (no structure implied)
Identify Name the specific thing, skip the explanation analyze (produces paragraphs)
Trace Map a causal or ownership chain end to end follow (no chain notion)
Consolidate Merge N sources, deduping combine (produces concatenation)
Synthesize Combine threads into a novel framework summarize (reduces, does not create)
Reconcile Resolve contradictions between two accounts compare (does not demand resolution)

Paired operators chain two steps. "Steelman the case for doing nothing. Then identify the weakest point." Or "Diagnose the root cause, assuming the obvious answer is ruled out. Then trace the failure to its origin." Generate exhaustively, then filter with structure.

Constraint markers

The sinc-LLM framework measured constraints as the single highest-importance component, roughly 43% weight, above persona, context, or task. Constraints compress the output space by ruling out directions. High work per token.

Prefer noun-based constraints over adjectives. Adjectives are interpreted relative to the model baseline. Nouns are unambiguous.

Noun-based Adjective-based (relative)
"Without the words leverage, synergy" "Keep it simple" (whose version of simple?)
"Never start a sentence with..." "Do not be too formal"
"Maximum 80 words" or "Exactly 3 bullets" "Make it brief" (how brief?)
"Focus solely on data-loss risks, not style" "Avoid being too technical"
"Return only a JSON array" "Improve this" (improve how?)

The five-constraint ceiling. Constraints listed after the 5th item show measurable attention decay. Accuracy dropped 12 points (78% to 66%) when a 7th was added. The model does not refuse; it quietly stops complying. Max 5 constraints per prompt. Harden that ceiling. Split the prompt if you need more.

The negative constraint. Model training defaults optimize for comprehensiveness, politeness, and safety. If you need something different, the not [default] pattern tells the model what to sacrifice:

The goal is to get a reply, not to explain everything.

"Return the top 3, not an exhaustive list" "Be blunt, not diplomatic" "Quote exact language, not paraphrases"

Without the not clause, the model resolves tradeoffs against its training default. It lands on comprehensive, polite, easy.

Domain anchors

A single domain term pulls the model into an entire expert register: vocabulary, reasoning patterns, formatting conventions. No paragraph of context needed. The ZIP paper confirms nouns carry the most signal in any prompt.

Term Register it activates
fiduciary Legal/financial. Duty, risk qualification, precedent.
latency Systems. Measurement, bottlenecks, quantified tradeoffs.
idempotent Systems design. Side-effect awareness, retry safety.
monotonic Math. Formal guarantees, bounds.
proximal cause Causal reasoning. Chain-of-causation.
amortization Financial schedule. Time-value, allocation.
thermal runaway Engineering safety. Cascade failure, compounding effects.

Use 1 or 2 per prompt. More than that produces jargon without substance.

Context framing

Four framing elements that each carry the weight of paragraphs. They change what a good answer looks like. Without them the model invents its own assumptions:

The audience is CFOs at mid-market SaaS companies skeptical of AI hype.
The goal is to get them to try again, not to explain how AI works.
This is a legal notice. Any ambiguity creates liability. Precision over readability.
This will appear in a LinkedIn post. Mobile-first, no images.

The audience, goal (with its not clause), and stakes lines do more than two paragraphs of context.

The SCOPE framework

For production prompts you reuse, run these five before writing:

  1. Signal. What processing mode? (audit, critique, extract, derive)
  2. Constrain. Format plus length plus boundaries.
  3. Orient. Domain anchor for expert register.
  4. Persona. Who is the model? Who is the audience?
  5. Exclude. What to avoid explicitly.

Covering just Signal, Constrain, and Exclude puts you in the top 20% of prompts. All five is top 5% (sinc-LLM framework methodology).

Typically: embed Orient and Persona in the system prompt (they are stable). Apply Signal, Constrain, and Exclude per query.

Structure and ordering

Instruction last. The Instruction Placement Effect puts the task instruction at the end of the prompt. That improves accuracy by roughly 5.6% (recency bias). Order: context, data, constraints, format, then task.

Semantic density. Ultra-dense prompts (SDE over 0.80, every token carrying signal) outperform diluted ones by roughly 8.4 points with zero extra tokens. Cut polite preambles ("can you please"), modal softeners ("maybe", "sort of"), meta-commentary ("I want you to tell me about"), restated context. Replace "briefly" with "in 80 words".

Task before format. Specifying the container before the task lets the container shape the reasoning. Task first, then format.

The failure-mode probe

Append "Where is this analysis most likely to be wrong?" to any analytical prompt. The model surfaces its own assumptions and edge cases. That answer can be worth more than the analysis itself.

Two-pass for accuracy

When accuracy matters more than token cost (contracts, compliance, medical):

  1. Pass 1. Extract or generate.
  2. Pass 2. A separate prompt reviews Pass 1 against the source, flagging unsupported claims.

Two-pass can cut hallucination by roughly 31% at 2x token cost. Worth using when a hallucinated answer costs more than the extra tokens.

What to avoid (the noise)

Word Why it is noise
"thorough" or "comprehensive" No definition. The model fills in its own version.
"good" Zero information. Good how? Fast? Accurate? Concise?
"help me with" Activates tutor mode, kills task specificity
"discuss" or "talk about" Weakest possible triggers. Shallow descriptive output.
"maybe" or "could you" Modal softeners that increase model uncertainty
"this is critical" Urgency language degrades accuracy. Longer, vaguer, less committed output.
"you are an expert" Every role is expert. No signal. Be specific about the actual role.

Other traps:

Model quirks

Agent-system vocabulary

If you are building agents, these terms define the state machine. A single word becomes a behavioral contract in system prompts:

Architecture patterns to know: ReAct (Thought-Action-Observation loop, the default scaffold), compaction (summarize context near the token limit), sub-agents (bounded workers with clean context windows returning condensed summaries), and planner-executor (separate strategic reasoning from tactical execution).

Further reading

The models are capable enough. The limiting factor in output quality is almost always the specification, not the generation.