AI prompt cheat sheet
An AI prompt cheat sheet is a short list of phrases that change what a model gives back. These are the ones that do the most work in ChatGPT, Claude and Gemini: a role, the audience, one task, a length, hard rules, permission to say UNKNOWN, a format, examples and a self-check. Each line below is wording you can paste. It is not a jailbreak list.
They map onto the five parts of a structured prompt — Role, Context, Task, Constraints, Output — which is the format every prompt in the library uses. Below them is the full cheat sheet: 648 terms in 12 groups, each explained in one line, with a search box.
Paste this → it changes that
| Write this | What it changes | Part |
|---|---|---|
You are a [job title] who [standard you hold]. | Vocabulary, depth and what “good” means. | Role |
Audience: [who will read it and what they know]. | Register, jargon and what gets explained. | Context |
What I already have: [draft / notes / data]. | Stops it inventing a starting point. | Context |
Wrap material: <doc>…</doc> | Separates your instructions from the text it works on. | Context |
Your task: [verb] [one deliverable]. | One clear job instead of a topic. | Task |
Think step by step, then give the final answer. | Better results on multi-step problems. | Task |
Max [N] words. / Exactly [N] bullets. | Length you can predict. | Rules |
Never [X]. Always [Y]. | Hard rules it can check itself against. | Rules |
If you don’t know, write UNKNOWN. | Permission not to guess — fewer confident inventions. | Rules |
Mark estimates as EST. | Separates guesses from facts. | Rules |
Cite the source for every figure, or leave it out. | Numbers you can verify. | Rules |
If something essential is missing, ask me one question first. | A question instead of a wrong assumption. | Rules |
Output as a table with columns: [A, B, C]. | A shape you can paste into a doc or sheet. | Output |
Return only valid JSON matching: { … } | Machine-readable output for tools and code. | Output |
Example of good output: [one example] | Shows the pattern — often better than describing it. | Example |
Give three options, then recommend one and say why. | Breadth first, then a decision. | Output |
Check your draft against every rule above, then give the corrected version. | A self-check pass before you see it. | Output |
Put the parts in this order: Role, Context, Task, Rules, Output — then any example. Every prompt in the library uses that order.
The full cheat sheet — 648 terms
648 terms in 12 groups
Prompt structure & techniques 92
The parts of a brief and the named techniques people mean when they talk about prompting.
role- Who the model should be for this task — sets vocabulary, depth and what “good” means.
context- The situation: who it is for, what already exists, what the output will be used for.
task- The one job, as a verb and a deliverable: “Draft a two-paragraph apology email.”
constraints- Hard rules the answer must obey — length, tone, never/always lines, what to do when unsure.
output- The exact shape of the answer: headings, a table with named columns, word limits, JSON.
RCTCO- Role, Context, Task, Constraints, Output — the five-part structure every library prompt uses.
brief- A structured prompt written like an instruction to a professional, not a one-line question.
brief-writer- Someone (or a skill) that turns a rough request into a complete brief.
system-prompt- Standing instructions that apply to a whole conversation or app, set before the user speaks.
developer-message- Instructions from the app builder that sit between the system prompt and the user.
user-message- What the person types in the chat; lowest in the instruction order.
few-shot- Giving two to five worked examples so the model copies the pattern.
zero-shot- Asking with no examples — instructions only.
one-shot- Exactly one example to show the pattern.
multi-shot- Several examples, usually chosen to cover different cases.
exemplar- One example in a few-shot prompt.
exemplar-set- The chosen group of examples; variety matters more than number.
negative-exemplar- An example of what NOT to produce, labelled as such.
chain-of-thought · CoT- Asking the model to reason step by step before the final answer.
zero-shot-CoT- Adding “think step by step” with no worked examples.
self-consistency- Sampling several independent answers and keeping the one most of them agree on.
majority-vote- Choosing the answer that appears most often across samples.
tree-of-thoughts · ToT- Exploring several lines of reasoning, scoring them, and continuing the best.
graph-of-thoughts · GoT- Like tree-of-thoughts, but ideas can merge and feed back into each other.
least-to-most- Solve the easiest sub-problem first, then use it to solve the next.
analogical-prompting- Ask the model to recall a similar solved problem before tackling this one.
generated-knowledge- Have the model list relevant facts first, then answer using them.
directional-stimulus- Give a small hint or keyword list that steers the answer.
program-of-thoughts · PoT- Have the model write code to compute an answer instead of doing maths in prose.
ReAct- Reason, act (call a tool), observe the result, repeat — the basic agent loop.
reflexion- The model reviews its failed attempt, writes a lesson, and tries again.
critique- A pass that looks for faults in a draft against stated criteria.
self-critique- The model checks its own draft against your rules, then gives a corrected version.
self-ask- The model asks itself follow-up questions and answers them before the final answer.
decomposed-prompting- Split a big job into sub-prompts, each handled separately.
meta-prompting- Using the model to write or improve a prompt against criteria you set.
recursive-prompting- Feeding an output back in as the input for the next round.
scratchpad- A space where the model works things out before giving a clean answer.
plan-and-solve- First write a plan, then carry it out step by step.
step-back · take-a-step-back- Ask the general principle first (“take a step back”), then apply it to the specific question.
emotion-prompting- Adding stakes (“this matters for my job”); effects are small and inconsistent — don’t rely on it.
role-persona- A detailed role: job, experience, and the standard they hold.
delimiter- Markers that separate your instructions from pasted material, e.g. ### or """.
XML-tags- Wrapping material in tags like <doc>…</doc> so the model knows what is what.
markdown-fence- Triple backticks around code or text to keep it separate.
JSON-enforcer- An instruction (or tool setting) that forces valid JSON output.
schema-first- Define the output schema before writing the rest of the prompt.
output-contract- A precise promise of what the answer will contain, so code or people can rely on it.
stop-sequence- Text that makes the model stop generating when it appears.
temperature- Randomness of the output: low for facts and formats, higher for ideas.
top-p- Sampling only from the most likely words that together make up p of the probability.
top-k- Sampling only from the k most likely next words.
presence-penalty- Discourages repeating topics that have already appeared.
frequency-penalty- Discourages repeating the same words often.
seed- A number that makes sampling repeatable where the API supports it.
logit-bias- Raising or lowering the chance of specific tokens appearing.
max-tokens- The cap on how long the answer can be.
context-window- How much text (in tokens) the model can take in at once, prompt plus answer.
working-memory- What the model can actively juggle within one answer — keep asks focused.
cognitive-load- How much effort a reader (or a prompt) demands; lower it with structure.
intrinsic-load- Difficulty that comes from the topic itself.
extraneous-load- Difficulty added by poor presentation — clutter, jargon, bad order.
germane-load- Effort that goes into actually learning or building understanding.
schema-theory- People learn by fitting new facts into existing mental structures.
dual-process- The idea of fast intuitive thinking vs slow deliberate thinking.
system-1- Fast, automatic thinking.
system-2- Slow, careful, step-by-step thinking.
attention-sink- Early tokens a model over-attends to; one reason the start of a prompt matters.
lost-in-the-middle- Models use the start and end of long inputs better than the middle — put key facts at the edges.
recency-bias- Over-weighting what came last in the prompt.
primacy-bias- Over-weighting what came first.
instruction-hierarchy- Which instructions win on conflict: system, then developer, then user, then pasted content.
prompt-injection- Hidden instructions inside pasted text or web pages that try to take over the model.
jailbreak- An attempt to make a model break its rules — not what this library is for.
leakage- The model revealing instructions or data it should keep private.
over-refusal- Declining a safe request out of excess caution.
under-refusal- Going along with a request it should have declined.
sycophancy- Agreeing with the user or flattering them instead of being accurate.
hallucination · confabulation- A confident statement that is false or made up — filling a gap with plausible invented detail.
groundedness- How well the answer sticks to the material you supplied.
faithfulness- Whether a summary or answer says only what the source says.
attribution- Saying which source each claim came from.
citation-needed- A marker for a claim that must be sourced or removed.
UNKNOWN- The word you tell the model to write instead of guessing: “If you don’t know, write UNKNOWN.”
EST- A label for estimates, so guesses are never mistaken for facts.
hedge-token- Words like “may” or “likely” that mark uncertainty honestly.
compression- Cutting a prompt to the words that carry the instruction.
distillation- Training a smaller model to copy a larger one; also used for condensing prompts.
prompt-folding- Packing a long instruction set into a shorter, reusable form.
latent-prompt- An instruction implied by examples or format rather than stated.
soft-prompt- A learned vector prompt used in fine-tuning, not written words.
prefix-tuning- A fine-tuning method that learns a prefix instead of changing the model.
Words that sharpen a brief 65
Single words that do a lot of work when you put them in a prompt. Name them and the model stops guessing.
specificity- Replace adjectives with facts: “under 150 words for busy parents” beats “short and friendly”.
audience- Who reads it and what they already know — the single biggest lever on tone.
format- The shape of the answer; name it (table, checklist, email) rather than hoping.
tone- How it should sound: warm, formal, plain, firm. Give an example line if it matters.
length- A number: words, bullets, minutes of reading. “Brief” means nothing.
examples- Two or three samples of what good looks like.
directive- A clear command verb at the start of the task.
verb- The action you want — draft, compare, rank, rewrite — not “help with”.
noun- The deliverable — an email, a plan, a table — so the output has an edge.
number- Counts make answers checkable: “five options”, “three risks”.
unit- State the unit — €, hours, kg — so numbers don’t come back ambiguous.
deadline- When it is due or what date it refers to; changes urgency and content.
owner- Who does each action; ask for an owner column in plans.
source- Where facts must come from: “only the pasted text”, “cite each figure”.
confidence- Ask the model to rate how sure it is, and why.
assumption- Ask it to list what it assumed so you can correct it.
estimate- Allowed guesses, clearly marked (EST) with the reasoning.
range- A low–high band instead of a single false-precise number.
floor- The minimum acceptable result or value.
ceiling- The maximum — budget, length, risk.
default- What to do when nothing else is specified.
exception- When the default does not apply — name the cases.
edge- Unusual cases the answer must handle (edge cases).
hazard- Something that could cause harm or a costly mistake — ask it to flag them.
likelihood- How probable a risk is; pair with impact.
impact- How bad it would be if it happened.
effort- How much work an option takes — useful for ranking choices.
reversible- Whether a decision can be undone cheaply; reversible ones can move faster.
blast-radius- How much breaks or who is affected if something goes wrong.
rollback- How to undo a change; ask for a rollback step in any plan.
flag- Ask the model to flag, not fix, things you must decide yourself.
variant- One version in a comparison or test.
control- The unchanged version you compare against.
treatment- The changed version in a test.
hypothesis- What you expect to happen and why — write it before testing.
metric- The number that tells you whether it worked.
denominator- What a rate is out of — always ask “out of what?”
numerator- The count on top of a rate.
baseline- Where things stand before the change.
lift- Improvement over the baseline.
delta- The difference between two values.
slope- The rate of change over time.
lag- A measure that shows results after the fact (lagging indicator).
lead- A measure that moves early and predicts results (leading indicator).
cohort- A group that started at the same time, tracked together.
segment- A slice of an audience with something in common.
persona- A described typical reader or customer to write for.
seat- The reader’s position — buyer, approver, end user; name it so the pitch fits.
room- Headroom left — in budget, time or word count — before a limit is hit.
layer- One level of instructions or content, e.g. system layer, user layer.
report- A structured write-up with findings and recommendations.
orphan- Something with no link or support — an unsourced claim, an unlinked page.
duplicate- Content that repeats something that already exists; ask the model to merge or remove.
canonical- The one authoritative version everything else points to.
alias- Another name for the same thing; list aliases so search catches them.
slug- The short URL-friendly name of a page, e.g. salon-loyalty-plan.
body- The main text, as opposed to title and metadata.
footer- The closing block — sign-off, source line, disclaimer.
heading- A section title; ask for headings to make long answers scannable.
table- Rows and columns; name the columns you want.
bullet- One point per line; set a maximum.
sentence- Limit by sentences for tight summaries: “three sentences”.
paragraph- Limit by paragraphs for emails and posts.
clause- One part of a contract or rule; ask for clause-by-clause review.
token- The unit models read and bill in — roughly ¾ of an English word.
Verbs for instructions 82
Start the task with one of these. A precise verb tells the model what kind of work to do.
specify- Spell out exactly what is required.
constrain- Add limits the answer must stay inside.
refuse- Say when the model should decline — e.g. requests for medical dosing.
compress- Shorten while keeping every instruction or fact that matters.
decompose- Break a big task into ordered sub-tasks.
truncate- Cut to a length, keeping the most important part.
cite- Give the source for each claim.
attribute- Say who said or produced what.
disclose- State openly — e.g. that AI helped write something.
meter- Count and limit usage.
gate- Only let something through when a condition is met.
budget- Set a limit of money, time, tokens or words.
invoke- Call a tool, skill or function.
install- Set up a skill, tool or file where it will be used.
uninstall- Remove it cleanly.
wrap- Put material inside delimiters or tags.
reset- Start fresh, dropping earlier context.
freeze- Lock a version so it no longer changes.
version- Save a numbered copy so changes can be compared.
score- Rate against criteria, with a number and a reason.
label- Tag each item with a category.
file- Put something where it belongs (a folder, a category).
merge- Combine duplicates or versions into one.
split- Separate one thing into several.
own- Assign responsibility to a named person.
kill- Stop or cancel a project or process on purpose.
ship- Release or send it.
preview- Show what it will look like before committing.
export- Produce it in a format for another tool.
import- Bring material in from elsewhere.
redact- Remove personal or confidential details.
hash- Turn data into a fixed fingerprint (used for checks and privacy).
sign- Mark as approved or authored.
pause- Stop temporarily, keeping state.
resume- Continue from where it paused.
verify- Check against a source or test.
falsify- Look for evidence that would prove the claim wrong.
pre-register- Write the plan and success criteria before you look at results.
calibrate- Adjust so stated confidence matches how often it is right.
normalize- Put data in one consistent format.
tokenize- Split text into tokens.
chunk- Split a long document into pieces for retrieval.
embed- Turn text into a vector for similarity search.
retrieve- Fetch the most relevant pieces for a question.
generate- Produce new text, images or code.
revise- Improve a draft against feedback.
accept- Approve as good enough.
reject- Decline, with the reason.
queue- Line up tasks to handle in order.
approve- Give the go-ahead (often a human step).
deny- Refuse permission.
log- Record what happened and when.
measure- Put a number on it.
compare- Set options side by side against the same criteria.
rank- Order from best to worst by a stated rule.
vote- Pick by majority across several answers or judges.
ensemble- Combine several models or answers.
distill- Boil down to the essentials.
quantize- Shrink a model by storing numbers with less precision.
serve- Run a model so apps can call it.
stream- Send output as it is generated.
batch- Process many requests together.
cache- Store a result so it can be reused quickly.
evict- Remove from a cache.
pin- Keep fixed in place (a version, a message, a file).
unpin- Release it.
mount- Attach a folder, drive or resource so a tool can use it.
unmount- Detach it.
isolate- Keep separate so one failure can’t spread.
sandbox- Run in a safe, restricted environment.
allowlist- Only these are permitted.
denylist- These are blocked.
throttle- Limit how often something can happen.
retry- Try again after a failure.
backoff- Wait longer between each retry.
circuit-break- Stop calling a failing service for a while.
degrade- Keep working with reduced features when something fails.
fallback- The plan B used when plan A fails.
handoff- Pass the work to another person or agent, with context.
escalate- Send it up to someone with more authority.
page- Alert the on-call person urgently.
ack- Acknowledge receipt of an alert or message.
Context engineering 47
Everything that goes into the model’s window besides your instruction — and how to budget it.
context-engineering- Deciding what goes into the model’s window — instructions, files, tools, memory — and in what order.
context-budget- How much of the window you allow each kind of content.
token-budget- The token limit you plan around for a task.
reserve-tokens- Room kept free for the answer so it isn’t cut off.
load-order- The order content is placed in the window; key facts go first or last.
progressive-disclosure- Load a short summary first and the details only when needed.
skill-card- The short name and description of a skill the model sees before loading it.
tool-schema- The definition of a tool’s inputs and outputs.
tool-result- What a tool returned, fed back to the model.
retrieval-chunk- One piece of a document fetched for a question.
rerank- Re-order retrieved chunks by true relevance before using them.
hybrid-search- Combining keyword search and meaning-based (vector) search.
dense-retrieval- Search by meaning using embeddings.
sparse-retrieval- Search by keywords (like BM25).
embedding- A list of numbers that represents the meaning of a text.
embedding-model- The model that turns text into embeddings.
chunk-size- How long each document piece is.
overlap- How much neighbouring chunks share, so ideas aren’t cut in half.
parent-document- Retrieve small chunks, but give the model the larger section they came from.
late-chunking- Embed the whole document first, then split — keeps context in each piece.
contextual-retrieval- Adding a short description of where a chunk sits before indexing it.
HyDE- Write a hypothetical answer first, then search with it.
MMR- Maximal marginal relevance — pick results that are relevant but not repetitive.
memory- Information kept across turns or sessions.
short-term-memory- What the model holds within the current conversation.
long-term-memory- Facts saved and brought back in later sessions.
episodic-memory- Memory of specific past events or conversations.
graph-memory- Memory stored as connected facts (entities and relations).
write-policy- The rules for what gets saved to memory — and what never does.
working-trace- The running log of steps in a long task.
transcript-reset- Starting a fresh window with a summary to shed clutter.
summarization-gate- A point where the conversation is summarised before continuing.
compaction- Shrinking earlier context into a summary to free space.
system-layer- The instructions at the system level.
developer-layer- Instructions from the app builder.
user-layer- What the user says.
tool-layer- Tool definitions and results.
retrieval-layer- Documents fetched for the task.
skill-layer- Skills loaded for the task.
collision- Two instructions that contradict each other.
overlapping-skills- Skills whose triggers overlap, so the wrong one may load.
instruction-conflict- When layers disagree; decide in advance which wins.
attention-budget- How much of the model’s focus a section is likely to get — long clutter dilutes it.
untrusted-context- Content from web pages, emails or users that may contain injected instructions.
wrap-retrieved- Put fetched text in tags and tell the model to treat it as data, not instructions.
citation-span- The exact passage that supports a claim.
quote-then-reason- Have the model quote the relevant passage first, then answer from it.
Agents, MCP & skills 79
The vocabulary of AI that takes actions: agents, tools, the Model Context Protocol and installable skills.
agent- A model that plans, uses tools and takes steps toward a goal, not just one reply.
harness- The software around a model that runs the loop, tools and permissions.
subagent- A helper agent given one piece of a bigger job.
orchestrator- The agent or code that assigns work to others and combines results.
peer-to-peer- Agents that coordinate as equals rather than through a boss.
hierarchical-multi-agent- Agents arranged with a lead and workers.
trajectory- The full sequence of an agent’s thoughts, tool calls and results.
trajectory-eval- Judging the path an agent took, not only its final answer.
tool-call- The model asking to run a tool with specific inputs.
tool-choice- Which tool the model picks — or a setting that forces one.
function-calling- The API feature that lets models call your functions with structured arguments.
parallel-tools- Calling several tools at once.
stop-condition- When the agent must stop: done, stuck, or out of budget.
loop-limit- A cap on how many steps an agent may take.
human-gate- A point where a person must approve before the agent continues.
human-in-the-loop- A person reviews or approves key steps.
approval-flow- The steps and people needed to approve an action.
agent-skills- Installable folders of instructions an agent loads when a task matches.
agentskills.io- The public format for Agent Skills.
SKILL.md- The file at the heart of an Agent Skill: a name, a description, then the instructions.
reference.md- An optional extra file a skill loads only when it needs the detail.
frontmatter- The block of settings (like name and description) at the top of a Markdown file.
kebab-case- Lowercase words joined by hyphens, e.g. salon-rebooking-desk.
auto-invoke- When an agent loads a skill by itself because the task matches its description.
trigger-description- The skill description written so the agent knows when to use it — and when not.
name-collision- Two skills or tools with the same name.
prefix-house-skill- Prefixing your own skills’ names so they never clash with others.
MCP · model-context-protocol- Model Context Protocol — an open standard for connecting AI apps to tools and data.
JSON-RPC- The message format MCP uses under the hood.
host- The app the person uses (a desktop chat app, an IDE) that connects to MCP servers.
client- The part of the host that talks to one MCP server.
server- A program that offers tools, resources or prompts over MCP.
MCP-tool- An action an MCP server offers, like “create_invoice”.
MCP-resource- Data an MCP server exposes for reading, like a file or record.
MCP-prompt- A reusable prompt template an MCP server offers.
scope- What a token or tool is allowed to touch.
least-privilege- Give each tool only the access it truly needs.
Smithery- A directory of MCP servers.
connector- A ready-made link between an AI app and another service.
localhost-server- An MCP server running on your own machine.
OAuth- The standard sign-in flow that grants an app limited access without sharing your password.
session-token- A temporary credential for one session.
Claude-Code- Anthropic’s command-line coding agent.
Codex-CLI- OpenAI’s command-line coding agent.
Cursor- An AI code editor.
Windsurf- An AI code editor.
Copilot-agent- GitHub Copilot’s agent mode for multi-step coding tasks.
DSPy- A framework that builds and optimises prompts as programs, scored by a metric.
signature- In DSPy, the declared inputs and outputs of a step.
module- In DSPy, a reusable step such as Predict or ChainOfThought.
Predict- DSPy’s basic module: inputs in, outputs out.
dspy.ChainOfThought- DSPy’s module that adds a reasoning step before the output.
GEPA- A prompt optimiser that improves prompts from written feedback on failures.
MIPROv2- A DSPy optimiser that searches instructions and examples together.
reflection-LM- The model used to reflect on failures and propose better prompts.
Pareto-frontier- The set of options where you can’t improve one score without hurting another.
textual-feedback- Written critique (not just a score) fed back to improve a prompt.
compile-program- In DSPy, running the optimiser to produce tuned prompts.
eval-harness- The code that runs test cases and scores outputs.
gold-set · golden-set- Test cases with trusted expected answers.
smoke-test- A quick check that nothing is obviously broken.
slice- A subset of test cases, e.g. only tricky inputs.
hazard-tag- A label on a test case for the risk it probes.
faithfulness-rubric- A scoring guide for whether answers stick to the sources.
schema-valid- Output that matches the required structure exactly.
JSON-schema- A standard way to describe the shape of JSON data.
pydantic- A Python library for defining and validating data shapes.
trace- The recorded path of one request through a system.
span- One timed step inside a trace.
telemetry- Data a system sends about its own behaviour.
OpenTelemetry- An open standard for traces, metrics and logs.
langsmith-like- Tracing and evaluation tools for LLM apps.
prompt-injection-via-resource- Injected instructions arriving through a file or tool result.
tool-misuse- An agent using a tool wrongly or for the wrong purpose.
exfiltration- Data being sent somewhere it shouldn’t go.
read-only-first- Start agents with read access; add write actions once they’re proven.
side-effect- A change in the world caused by an action — sending, paying, deleting.
idempotent- Safe to run twice: the second run changes nothing.
pagination-cap- A limit on how many pages of results a tool fetches.
Measuring output: evals & quality 69
How people check whether prompts and AI features actually work — and the numbers they watch.
eval · evaluation- A repeatable test of an AI output against criteria.
case-id- A unique ID for each test case.
pass-bar- The score an output must reach to count as a pass.
offline-eval- Testing on a fixed set of cases before release.
online-eval- Measuring real usage after release.
shadow-eval- Running a new version silently alongside the old to compare.
canary- Releasing to a small share of users first.
LLM-as-judge- Using a model to grade outputs against a rubric — useful, but check its bias.
pairwise-judge- A judge that compares two answers and picks the better.
rubric-score- A score from written criteria.
Brier- The Brier score for probability forecasts: lower is better.
calibration- Whether “80% sure” is right about 80% of the time.
ECE- Expected calibration error — the average gap between confidence and accuracy.
reliability-diagram- A chart of stated confidence against actual accuracy.
precision- Of the items flagged, how many were right.
recall- Of the items that should be flagged, how many were caught.
F1- A single score balancing precision and recall.
exact-match · EM- Output identical to the expected answer.
ROUGE- A word-overlap score often used for summaries.
BLEU- A word-overlap score from machine translation.
BERTScore- A similarity score based on meaning rather than exact words.
factuality- Whether claims are true.
regression- Something that used to work and now doesn’t.
drift- Gradual change in inputs or outputs over time.
distribution-shift- Real inputs differ from the ones you tested on.
red-team- Deliberately attacking a system to find weaknesses.
purple-team- Attackers and defenders working together.
table-stakes- The basics something must do just to be considered.
severity- How serious a problem is.
critical- Must be fixed before anything ships.
major- Serious, but not a blocker.
minor- Small; fix when convenient.
false-refusal- The model declined something it should have done.
false-comply- The model did something it should have declined.
over-cite- Citing sources that don’t support the claim, or too many.
under-cite- Claims without the sources they need.
latency- How long a response takes.
cost-per-pass- What it costs to get one acceptable output.
tokens-in- Tokens sent to the model.
tokens-out- Tokens the model generated.
cache-hit- A request answered from cache instead of recomputed.
prompt-cache- Reusing the processed start of a prompt to save time and cost.
prefix-cache- Caching a shared prompt beginning across requests.
sensitivity- How reliably a check catches real problems.
sharpness- How confident (far from 50/50) forecasts are, regardless of accuracy.
resolution- How well forecasts separate things that happen from things that don’t.
coverage- The share of cases or topics a test set includes.
recall-at-k- Whether the right item appears in the top k results.
hit-rate- How often the thing you wanted was found.
miss-rate- How often it wasn’t.
latency-p50- The median response time.
latency-p95- The time 95% of responses beat — shows slow cases.
latency-p99- The time 99% beat — the worst typical case.
cost-per-1k- Cost per thousand tokens or requests.
cost-per-task- Total cost to complete one job end to end.
window-waste- Context spent on content that didn’t help the answer.
skill-hit- A skill loaded when it should have.
skill-miss- A skill that should have loaded but didn’t.
skill-collision- The wrong one of two similar skills loaded.
retrieval-hit- Retrieval fetched the passage that answers the question.
retrieval-miss- It didn’t.
retrieval-junk- Retrieved text that is irrelevant and just takes space.
tool-hit- The right tool was called with the right inputs.
tool-miss- A needed tool wasn’t called.
human-edit-rate- How much people change AI drafts before using them.
first-pass-accept- Share of outputs used without edits.
rewrite-cycles- How many rounds of revision a draft needs.
minutes-to-draft- Time from request to a usable draft.
open-rate- Share of recipients who open an email.
Models & how they work 26
Plain meanings for the words used about how today’s AI models are built and run.
transformer- The neural-network design behind today’s language models.
attention- The mechanism that lets a model weigh which earlier words matter for the next one.
KV-cache- Stored intermediate results that make long generations faster.
mixture-of-experts · MoE- A model made of many sub-networks, only some used per word — cheaper to run.
routing- Choosing which expert, model or tool handles a request.
reasoning-model- A model trained to think at length before answering hard problems.
chain-model- A setup where one model’s output feeds the next.
speculative-decoding- A small model drafts words and a large model checks them — faster output.
draft-model- The small model in speculative decoding.
multimodal- Handles more than text — images, audio, video.
vision-language- A model that understands images and text together.
image-brief- A structured prompt for an image model: subject, setting, light, palette, aspect ratio.
voice-agent- An AI you talk to by voice that can take actions.
realtime-api- An interface for live, low-delay voice or text conversations.
long-context- Models that take very long inputs — whole books or codebases.
needle-in-haystack- A test of whether a model finds one fact in a very long input.
synthetic-data- Training or test data generated by models.
alignment- Making models behave as intended and safely.
RLHF- Reinforcement learning from human feedback — training on people’s preferences.
RLAIF- The same, with AI-generated feedback.
DPO- Direct preference optimisation — a simpler way to train on preferences.
constitutional-principles- Written rules a model is trained to follow when judging its own answers.
toolformer- Early research on models teaching themselves to use tools.
functionary- Open models tuned for function calling.
computer-use- An AI that operates a computer screen — clicking and typing.
browser-use- An AI that navigates web pages to complete tasks.
Safety, legal & honesty 18
Words for keeping AI output honest and within the rules. Not legal advice — check with a professional.
AI-Act- The EU’s law on artificial intelligence, with duties that depend on risk level.
transparency-note- A plain statement of what an AI feature does, its limits and the data it uses.
provider-identity- Which company’s model sits behind a product.
output-may-be-wrong- The honest warning that AI output needs checking.
high-risk-not-declared- A reminder not to use AI for regulated high-risk decisions without the required steps.
training-on-user-data- Whether a service uses what you type to train its models — check the settings.
PII- Personally identifiable information — names, addresses, IDs. Redact before pasting.
credentials- Passwords, keys and tokens — never paste them into a chat.
unpublished-financials- Confidential numbers that shouldn’t go into public tools.
legal-conclusion- A definitive legal answer — AI can prepare questions, a lawyer gives conclusions.
medical-advice- Diagnosis or treatment — AI can help you prepare; a clinician decides.
employment-decision- Hiring or firing — AI can support, but people decide and must check fairness rules.
credit-decision- Lending decisions — regulated; AI output needs human review.
acceptable-use- A service’s rules for what you may use it for.
refund-window- The period in which a purchase can be refunded.
SOC2-only-if-report- Only claim SOC 2 compliance if you have the audit report.
copyright-summary- Summarise protected text in your own words instead of copying it.
not-verbatim- An instruction to paraphrase rather than reproduce.
Teaching with AI 29
Terms teachers use when they bring AI into assignments honestly.
integrity-trace- Evidence of how a student produced work — drafts, notes, version history.
process-artifact- A piece of work showing the steps, not just the final product.
oral-defense- A short conversation where students explain their work.
disclosure-line- A sentence where students state how they used AI.
allowed-tools- What students may use for a task.
required-use- A task where using AI is part of the assignment.
forbidden-use- Where AI is not allowed.
rubric- The criteria and levels used to grade work.
criterion- One thing the rubric assesses.
descriptor- The description of what each level looks like.
portfolio- A collection of work over time.
in-class-write- Writing done in the room, often to see a student’s own voice.
unseen-task- A task students haven’t prepared for in advance.
watermark- A hidden signal in AI output — unreliable as proof on its own.
detector- Software claiming to spot AI writing — error-prone; never the only evidence.
false-positive- Wrongly flagging honest work as AI-written.
syllabus-paragraph- The course policy statement on AI use.
department-chair- The person who sets or approves department policy.
Socratic- Teaching by asking questions that lead the student to reason.
Feynman- Explaining a topic simply, as if to a beginner, to test understanding.
worked-example- A fully solved example students study before trying.
cognitive-apprenticeship- Teaching by modelling expert thinking out loud.
formative- Assessment for learning, during the course.
summative- Assessment of learning, at the end.
resit- A second attempt at an assessment.
accessibility- Making material usable for everyone, including disabled learners.
spellcheck-analogy- Framing AI as a tool like spellcheck — helpful, but it doesn’t do the thinking.
citation-of-model-use- Citing an AI tool the way you would a source.
bibliography-line- The reference entry for AI use in a bibliography.
Business, growth & search 27
Words that come up when you use AI for pricing, marketing and being found online.
value-metric- What you charge by — per seat, per project, per use.
JTBD · job-to-be-done- Jobs to be done — the progress a customer is trying to make; buying decisions follow from it.
pricing-page- The page that explains plans and prices.
plan-ladder- Plans arranged from basic to premium.
VAT- Value-added tax, included in consumer prices in the EU and UK.
Stripe- A payments company used for online checkout and subscriptions.
conversion- A visitor doing the thing you want — buying, signing up.
activation- A new user reaching the first real value.
retention- Customers who keep using or paying.
churn- Customers who leave.
NPS- Net Promoter Score — how likely people are to recommend you.
CSAT- Customer satisfaction score.
CES- Customer effort score — how easy it was to get help.
claim-hygiene- Only saying things you can prove.
invented-metric- A made-up number presented as fact — never publish one.
positioning-line- One sentence on who you are for and why you’re different.
voice-bible- A guide to how a brand sounds, with examples.
organic-search- Visitors from unpaid search results.
E-E-A-T- Experience, expertise, authoritativeness, trust — what Google’s raters look for.
schema- Structured data that tells search engines what a page is about.
FAQ-page- A page of questions and answers, marked up with the FAQPage structured-data type for search.
zero-click- A search answered on the results page, with no visit.
AI-overview- Google’s AI-written summary at the top of some results.
snippet- The short text shown under a search result.
paywall- Content that needs payment to read in full.
teaser- The opening part shown before a paywall.
capstone- A final project that pulls a course together.
Tools, companies & web standards 63
Names you will meet when working with AI and the web.
anthropic- The company that makes Claude.
openai- The company that makes ChatGPT.
google- Makes Gemini and runs Google Search.
grok- xAI’s chat model.
gemini- Google’s family of AI models and its chat app.
claude- Anthropic’s family of AI models and its chat app.
sonnet- Claude’s mid-size model line.
opus- Claude’s largest model line.
chatgpt- OpenAI’s chat app.
copilot- Microsoft’s and GitHub’s AI assistants.
replit- An online coding platform with an AI agent.
v0- Vercel’s AI tool that generates web interfaces.
langchain- A framework for building LLM apps.
langgraph- A LangChain library for agent workflows as graphs.
llamaindex- A framework for connecting LLMs to your data.
haystack- An open-source framework for search and RAG apps.
chromadb- An open-source vector database.
pinecone- A hosted vector database.
weaviate- An open-source vector database.
qdrant- An open-source vector database.
milvus- An open-source vector database built for scale.
postgres- A widely used open-source database (PostgreSQL).
sqlite- A small database stored in a single file.
duckdb- A fast in-process database for analytics.
parquet- A column-based file format for data.
iceberg- A table format for large data lakes.
s3- Amazon’s file storage service.
r2- Cloudflare’s file storage service.
minio- Open-source storage compatible with S3.
oidc- OpenID Connect — sign-in built on OAuth.
jwt- JSON Web Token — a signed token carrying identity or permissions.
jwks- The published keys used to check JWT signatures.
sepa- The European system for euro bank transfers.
prisma- A database toolkit for TypeScript.
drizzle- A lightweight TypeScript database library.
nextjs- A React framework for websites and apps.
remix- A React web framework.
astro- A framework for fast, content-heavy websites.
tailwind- A CSS framework of small utility classes.
shadcn- A popular set of copy-in React interface components.
webvitals- Google’s measures of page experience.
INP- Interaction to Next Paint — how quickly a page responds to taps and clicks.
LCP- Largest Contentful Paint — how fast the main content appears.
CLS- Cumulative Layout Shift — how much the page jumps while loading.
TTFB- Time to First Byte — how fast the server starts responding.
sitemap- A file listing a site’s pages for search engines.
robots- robots.txt — the file telling crawlers what they may visit.
llms.txt- A plain-text summary of a site written for AI tools.
jsonld- JSON-LD — the usual format for structured data on web pages.
schema.org- The shared vocabulary for structured data.
Course- The structured-data type for a course.
HowTo- The structured-data type for step-by-step instructions.
YMYL- “Your money or your life” — topics where search engines demand extra trust.
murphy-decomposition- Splitting a Brier score into reliability, resolution and uncertainty.
metaculus- A public forecasting platform.
polymarket- A prediction market.
good-judgment- Good Judgment — a forecasting organisation known for superforecasters.
superforecaster- A person with a strong, measured record of accurate forecasts.
ACLED- A dataset of conflict events worldwide.
UCDP- Uppsala Conflict Data Program — a conflict dataset.
FIRMS- NASA’s near-real-time fire data.
FRED- The St. Louis Fed’s economic data service.
IMF- The International Monetary Fund, a source of economic data.
Blockchain & on-chain terms 51
For briefs about crypto, tokens and on-chain systems. Education, not investment advice.
protocol- The rules and software a blockchain system runs on.
L1- A base blockchain, like Ethereum or Bitcoin.
L2- A network built on top of an L1 to make it cheaper or faster.
rollup- An L2 that bundles transactions and posts them to the L1.
optimistic- A rollup that assumes transactions are valid unless challenged.
ZK- Zero-knowledge — proofs that something is valid without revealing the details.
data-availability- Whether the data needed to check a chain is actually published.
escape-hatch- A way for users to withdraw even if the operator stops working.
sequencer- The party that orders transactions on an L2.
oracle- A service that brings outside data (like prices) on-chain.
wrapper- A token that represents another asset on a different chain or form.
RWA- Real-world asset — a legal claim to something off-chain, represented by a token.
legal-claim- What a token actually gives you a right to — often less than it seems.
restaking- Reusing staked assets to secure more services, which adds shared risk.
slashing- Losing staked funds as a penalty for misbehaviour.
correlated-risk- Risks that fail together, so spreading doesn’t protect you.
intent- A user states the outcome they want; others work out how to do it.
solver- A party that fills an intent.
account-abstraction- Smart-contract wallets with features like recovery and paid-for fees.
bridge- Moves assets between chains — historically a common point of failure.
canonical-token- The official version of a token on a chain.
issuance- How new tokens are created and at what rate.
fee-market- How transaction fees are set by demand.
MEV- Value extracted by re-ordering or inserting transactions.
PBS- Proposer-builder separation — splitting block building from block proposing.
proposer- The validator that proposes a block.
builder- The party that assembles a block’s transactions.
governance- How a protocol makes decisions and changes.
token-vote- Voting weighted by tokens held.
multisig- A wallet that needs several signatures to act.
timelock- A delay before a change takes effect, giving users time to react.
pause-guardian- A party allowed to pause a protocol in an emergency.
upgrade-key- Who can change the contract’s code.
proxy- A contract pattern that allows upgrades.
finality- When a transaction can no longer be reversed.
reorg- When recent blocks are replaced by a different chain history.
liveness- Whether the system keeps producing blocks.
correctness- Whether what it produces is valid.
TVL- Total value locked — deposits in a protocol; not a measure of safety.
not-a-design-proof- Popularity or deposits don’t prove a design is sound.
APY-not-evidence- A yield number says nothing about how safe a protocol is.
explorer-check- Verify claims on a block explorer rather than trusting posts.
contract-verify- Check that published code matches what is deployed.
incident-class- The kind of incident — exploit, outage, oracle fault, UI-only.
UI-only- The website is down but the contracts and funds are fine.
custody- Who actually holds the assets.
oracle-fault- Wrong outside data causing wrong on-chain behaviour.
status-page- The public page for incident updates.
next-update-time- Always tell users when the next update will come.
do-not-tweet- Don’t post details publicly before they are confirmed.
funds-unproven- Don’t claim funds are safe until it has been verified.
Build your own labels
Long briefs read better when their steps have short, consistent names. Combine the words below — nothing magic, just clear labels you and the model can both refer back to.
Verb + part
define · apply · avoid · measure · teach · ship · gate · test · document · retire + role · context · task · constraints · output · example · audience · tone · format · source · claim · risk
e.g. define-role · test-output · gate-context · document-constraints
Subject + attribute
prompt · context · skill · agent · tool · eval · trace · token · schema · brief · role · task · output · claim · model · image · risk · policy · rubric + audit · budget · check · drift · gate · limit · log · map · note · order · plan · policy · risk · rule · score · spec · test · trace · version
e.g. prompt-audit · context-budget · output-spec · claim-check
Prefix + subject
pre · post · anti · non · re · co · sub · multi · cross · over · under · mis · auto + brief · claim · eval · skill · agent · tool · trace · token · plan · risk · prompt · context · schema · score · run
e.g. pre-brief · post-eval · anti-claim · re-run · over-cite
Three habits the phrases don’t replace
- Say what the output is for and who acts on it. A model can’t optimise for a goal it wasn’t told.
- Give it the material. A pasted draft, a table or three real examples beat any adjective.
- Review what comes back. A structured prompt is a professional starting point, not a guarantee — see why models invent things.
Prompts that use these phrases (free to read)
- Staff-Engineer Code ReviewReviews a PR like a senior who optimizes for the next reader, not style nits.Free
- Calm + Concrete Investor UpdateMonthly update that earns trust by being honest about misses.Free
- Cold Outbound — Peer-to-PeerReach a busy founder without AI-slop tone. Four lines, one ask, no hedging.Free
- Sales Discovery — 7 Questions That Surface PainAsk the questions that surface pain, not features. Seven questions, no leading the witness.Free
- Rubric Design — Creative WorkDesign a rubric for grading creative work that doesn't kill the creativity.Free
This sheet and the prompts above are free to read. The free prompt set is free with an account; the rest of the library comes in ready packs, one payment each.
Questions
What is an AI prompt cheat sheet?
A short reference of phrases that reliably change what a model returns — how to set a role, give context, fix the length, add rules, ask for a format and add a self-check — with wording you can paste.
Are there secret codes that unlock better AI answers?
No. There are no hidden commands. What works is clear structure: who the model should be, the situation, one task, the rules and the shape of the answer. The phrases on this page are simply the clearest ways to say those things.
Does this work in ChatGPT, Claude and Gemini?
Yes. The phrases are plain language and the structure is the same for every chat model. Exact results differ by model, so check the output — especially numbers and facts.
What is the most useful single line?
Usually the audience and the material: who it is for and what you already have. After that, “If you don’t know, write UNKNOWN” removes a lot of confident guessing.