Prompt & skill library

Setup

Lead with who the answer is for and what it must achieve. The audience and goal shape every line that follows.

Setup

Specify the audience before the format. "For a busy engineer who will skim" beats "use bullets" alone.

Setup

When stuck, paste one example of output you would happily use. Models copy the level of a good example better than they follow adjectives.

Setup

Separate instructions from material: put pasted documents inside clear markers like <document>…</document> so the model cannot mistake them for your request.

Setup

Put long documents first and your actual question last, so the instruction is the freshest thing the model reads.

Setup

State the date and your location when it matters: "Today is 1 November 2026, I am in the UK." The model does not know your date, currency or laws unless you say.

Setup

Give the model the facts it cannot know — prices, policies, names, numbers. Missing facts are exactly where invented details creep in.

Setup

Say what you will do with the answer: send it, print it, paste it into code. The right format follows from the destination.

Setup

Mark the details that change each time as [BLANKS]. A prompt with clear blanks is reusable; one with buried specifics is not.

Setup

If a task has several parts, number them. Numbered steps are easier for the model to complete in order and easier for you to check.

Setup

Tell the model what to do when information is missing: "Ask me one question" or "mark the gap as [TO CONFIRM]". Otherwise it may guess.

Setup

Put what never changes (tone, audience, banned words) in custom instructions or a project, and keep each message for this task only.

Setup

Keep one task per conversation. Long chats that switch topics drag old instructions into new answers.

Setup

Start a fresh chat when the model keeps repeating a mistake. A clean context often fixes what more corrections cannot.

Patterns

Show two or three examples of the exact output you want (few-shot). Pick examples that differ from each other so the model learns the pattern, not one case.

Patterns

Include one counter-example: "Not like this:" plus a bad sample and why it is bad. It shows the boundary faster than a rule.

Patterns

Ask for a draft, then a critique against your criteria, then a revision. A self-review pass catches mistakes, at the cost of extra time.

Patterns

For agent loops, cap iterations explicitly: "Stop after 5 attempts and report what is still failing."

Patterns

End prompts with a checklist: "Before answering, check: every [BLANK] handled, no invented numbers, under 200 words."

Patterns

Give the model a weighted rubric whenever you ask it to judge or score. Without one, scores drift from run to run.

Patterns

Ask for three options that differ on one thing — shorter or longer, safe or bold — then pick. Choosing is faster than describing.

Patterns

When a prompt works, turn it into a template: swap the specifics for [BLANKS] and save it. Your best prompts are assets.

Patterns

Chain prompts for big jobs: outline, then section drafts, then an edit pass. Each step is small enough to check.

Patterns

For classification, give the full list of allowed labels plus "other". Open-ended labels multiply.

Patterns

Role-play the reader: "You are a sceptical CFO reading this. List your three objections." Then fix those.

Patterns

Ask the model to restate the task in one sentence before it starts. If the restatement is wrong, your prompt is unclear.

Reasoning

Ask for step-by-step reasoning on maths and logic problems. Chain-of-thought prompting took PaLM 540B from 18% to 57% on the GSM8K maths benchmark (Wei et al., 2022).

Reasoning

Ask for reasoning once, clearly. Repeating "think step by step" makes answers longer, not better.

Reasoning

For tasks with more than five steps, ask for the plan first, check it, then ask for the execution in a second message.

Reasoning

Follow up with "What could be wrong with this answer?" It is a cheap second look before you rely on it.

Reasoning

For ambiguous tasks, ask the model to list the possible interpretations, pick one and say why before it proceeds.

Reasoning

When you want an opinion, force a decision: "Pick one and defend it in three sentences" beats "weigh the options".

Reasoning

Ask for assumptions as a separate list. A wrong assumption takes seconds to correct; a wrong conclusion takes longer to spot.

Reasoning

For estimates, ask for a low, likely and high figure with the reasoning for each. A single number hides the uncertainty.

Reasoning

Ask the model to argue against its own recommendation, then give a final view. It surfaces risks you would otherwise miss.

Reasoning

For calculations, ask to see every step or use a tool that runs code. Language models can slip on long arithmetic.

Reasoning

Ask the model to label each point "from your input" or "my inference". It separates facts from guesses.

Reasoning

If answers change a lot between runs, the question is underspecified. Add constraints instead of re-rolling.

Format

If you need JSON, give the exact schema — field names, types, required fields — and one valid example.

Format

Ask for a number of bullets or lines rather than words. You can check it at a glance.

Format

Ask for a table whenever you compare more than two options on more than two criteria.

Format

For structured output, validate it against a schema in code and re-ask with the error message when it fails.

Format

List the headings you want, in order. An explicit output format is how you get the same shape every time.

Format

Ask for the bottom line first for anything a busy person will read.

Format

Ask for plain text when you will paste into a tool that does not render markdown, or the asterisks come along.

Format

For emails, specify the subject line, length and the one action you want. "Reply-ready" is clearer than "professional".

Format

For slides, ask for one message per slide plus speaker notes, not paragraphs on the slide.

Format

Give character limits for anything going into a limited field: ad headlines, meta descriptions, text messages.

Format

Ask for CSV with a header row when the output is going into a spreadsheet.

Format

For long documents, ask for a table of contents first, approve it, then generate one section at a time.

Voice

Describe tone through the relationship with the reader: "direct, peer-to-peer" works better than "speak clearly".

Voice

Ban the phrases you never want by name — "delve", "in today's fast-paced world", "unlock". Specific bans work better than "avoid clichés".

Voice

Ask it to drop hedging words ("might", "could", "may") where you are sure — and keep them where the uncertainty is real.

Voice

Paste a paragraph of your own writing and ask the model to match its sentence length, word choice and rhythm.

Voice

Specify register, not formality: "casual peer" versus "polite stranger" gives sharper differences than formal versus informal.

Voice

Ask for a reading level — "plain English, short sentences" — when writing for customers or the public.

Voice

Say who is speaking: "Write as the owner of a small bakery, first person, warm." A voice needs a speaker.

Voice

Ask for a version 30% shorter. Most first drafts get clearer when cut.

Voice

Drop exclamation marks and superlatives unless your brand really talks that way. They make text read as machine-written.

Voice

For translations, ask for natural phrasing for a native speaker in a named country, not a word-for-word translation.

Facts & research

Paste the source and say "answer only from this text". Ask the model to quote the sentence it relied on.

Facts & research

Never trust a citation you have not opened. Models can produce realistic references that do not exist.

Facts & research

Ask the model to mark anything it is unsure of as [VERIFY], then check those first.

Facts & research

For current events, prices or laws, use a tool with web search and check the date on every source.

Facts & research

Ask for the date behind each figure. A correct number from 2019 can be the wrong number today.

Facts & research

Ask "what would change this conclusion?" to find out which facts are worth verifying.

Facts & research

When summarising a study, ask for sample size, method and limitations — not just the headline result.

Facts & research

Ask for the strongest counter-evidence too. A one-sided summary is convincing and misleading.

Facts & research

For legal, medical or money questions, use AI to understand the terms and prepare questions, then confirm with a qualified professional.

Facts & research

Number your sources and ask the model to cite by number. Every claim stays traceable.

Facts & research

Ask for definitions of the key terms before any analysis. Many disagreements are two different definitions.

Facts & research

Distrust precise-sounding statistics with no named source. If you cannot trace it, do not publish it.

Agents & tools

Give an agent a clear stopping condition and a budget — steps, time or money — before it starts.

Agents & tools

Let agents read freely, but require your confirmation before anything that sends, deletes, buys or publishes.

Agents & tools

Write tool descriptions as carefully as prompts: what the tool does, when to use it and what it returns.

Agents & tools

Ask the agent to keep its plan in a checklist and tick items off. It is easier to follow, and to resume after a failure.

Agents & tools

Treat what an agent reads on web pages or in emails as data, not instructions. Prompt injection hides commands inside content.

Agents & tools

Give agents the smallest permissions that work: read-only keys, sandbox folders, test accounts.

Agents & tools

Log every tool call an agent makes. When something goes wrong, the log shows where.

Agents & tools

Start agents on work you can check quickly. Verifiable tasks first, judgement calls later.

Agents & tools

For research tasks, have the agent collect sources first, then summarise only from those sources.

Agents & tools

Use a separate pass, or a separate agent, to review what another wrote. A fresh reviewer misses less than the author.

Agents & tools

Keep reusable instructions in a skill or project file instead of retyping them in every chat.

Agents & tools

When an agent fails, read its trace before rewriting the prompt. The fix is often a missing fact or a missing tool.

Coding with AI

Give the model the relevant files, the exact error and what you already tried. Partial context produces plausible but wrong fixes.

Coding with AI

Ask for a plan and the list of files it will change before it writes any code.

Coding with AI

Ask for tests alongside the code, then run them. Code that has not run is a draft.

Coding with AI

State your versions — language, framework, libraries. APIs change, and old patterns still look correct.

Coding with AI

Ask for the smallest change that fixes the bug, not a rewrite. Small diffs can be reviewed.

Coding with AI

Paste the full stack trace, not a summary. The useful line is often the one you would have cut.

Coding with AI

Never paste secrets, API keys or customer data into a prompt. Use placeholders.

Coding with AI

Ask the model to explain the code back to you before you merge it. If you cannot follow it, do not ship it.

Coding with AI

For refactors, ask it to keep behaviour identical and to list anything that might change.

Coding with AI

Check that a suggested package exists and is maintained before installing it. Plausible names can be typo-squatted.

Coding with AI

Run a separate security review prompt: injection, authorisation checks, secrets, input validation.

Coding with AI

Keep a project instructions file — conventions, commands, gotchas — that coding agents read first.

Images & video

Describe subject, setting, lighting, camera angle and style, in that order. The order shows you what is missing.

Images & video

For photoreal images, name the lens and the light: "50mm, soft window light, shallow depth of field".

Images & video

When editing an image, ask for one change at a time. Several changes at once often break what was right.

Images & video

Put any text you need in the image in quotes and keep it short. Long text inside images is still error-prone.

Images & video

Reuse one style line across a series so the images look like one brand.

Images & video

For product images, never let AI change the product itself. Shape, logo and colour must match reality.

Images & video

Check usage rights and platform rules before using AI images in ads. Some platforms require you to label AI-generated or altered content.

Images & video

For video, describe camera movement separately from the scene: "slow push-in", "handheld follow".

Images & video

Generate several variations and pick one, instead of perfecting a single prompt for an hour.

Images & video

Do not create realistic images of real people without consent, or anything that could mislead viewers.

Data & spreadsheets

Paste the column names and three sample rows before asking for a formula. Formulas depend on your exact layout.

Data & spreadsheets

Ask for the formula plus a plain explanation of what each part does.

Data & spreadsheets

Say what each column means and its units. "amount" in cents or in euros changes every answer.

Data & spreadsheets

Ask for a data quality check first: blanks, duplicates, odd values, mixed formats.

Data & spreadsheets

Use a tool that runs code for anything beyond a few hundred rows. Do not ask a chat to eyeball big tables.

Data & spreadsheets

Ask for SQL with comments and a note of which tables and filters it assumes.

Data & spreadsheets

Spot-check AI calculations against two or three rows you work out by hand.

Data & spreadsheets

Ask for the chart type with a reason: trends as lines, comparisons as bars, parts of a whole only when there are few parts.

Data & spreadsheets

Remove personal data before pasting a spreadsheet, or replace names with IDs.

Data & spreadsheets

Ask for a written definition of every metric in a report so everyone reads it the same way.

Business & sales

Paste the customer's actual words and ask for a reply that answers each point they raised.

Business & sales

For proposals, give scope, price, timeline and what is excluded. The model cannot price your work for you.

Business & sales

Ask for the three objections a buyer will raise and an honest one-line answer to each.

Business & sales

Turn call notes into decisions, owners and dates — not a summary of who said what.

Business & sales

For pricing, give your costs and target margin and ask to see the maths. Never let it invent your numbers.

Business & sales

In cold emails, give one real, specific reason you are contacting this person. Generic personalisation reads as spam.

Business & sales

Ask for a 20-second version of any pitch. If it does not work short, it will not work long.

Business & sales

Ask for the worst case in any plan, and what you would do if it happens.

Business & sales

Decide your walk-away point before asking for negotiation scripts.

Business & sales

Use AI to draft and explain contracts and policies, and have the important ones reviewed by a professional.

Business & sales

End every plan with next steps, each with an owner and a date.

Business & sales

Before writing a job advert, write what the person must achieve in their first 90 days.

Marketing & ads

Give the product, the specific customer and the one problem it solves. "For everyone" produces copy for no one.

Marketing & ads

Ask for three angles — problem, outcome, proof — and test them against each other.

Marketing & ads

Every claim in an ad needs evidence you could show a regulator. Ask the AI to mark the claims that need proof.

Marketing & ads

In the US, the FTC rule on fake reviews (in force since October 2024) bans buying reviews, fake or AI-written reviews, and undisclosed insider reviews.

Marketing & ads

Ask for a clear disclosure line in every influencer brief for paid or gifted content.

Marketing & ads

Match the ad's promise to the landing page headline. A mismatch wastes the click you paid for.

Marketing & ads

Ask for a version sized for each platform's limits instead of trimming one long version by hand.

Marketing & ads

Write the measurement plan before the campaign: the metric, the target and the date you will decide.

Marketing & ads

Ad platforms count conversions their own way. Compare their numbers with your own sales data before scaling spend.

Marketing & ads

In creative tests, change one thing at a time — the hook or the image, not both.

Marketing & ads

Advertising to the EEA or UK with Google Ads? Google has required Consent Mode v2 since March 2024 to keep features such as remarketing and conversion modelling working.

Marketing & ads

Turn one strong piece into many formats: a blog post becomes a carousel, a short video and an email.

Startups & investors

Before an investor meeting, look up their recent investments and check for portfolio companies that compete with you.

Startups & investors

Answer the question the investor asked, then stop. Long answers to short questions lose the room.

Startups & investors

Know your numbers cold — revenue, growth, burn, runway, retention. Ask an AI to quiz you on them.

Startups & investors

Most pre-seed rounds are raised on SAFEs — Carta reported 93% in Q2 2026. Model your dilution before you sign one.

Startups & investors

Run fundraising as a process: a target list, a short window for first meetings, and a tracker for every conversation.

Startups & investors

Send a short follow-up within a day of each investor meeting, with any answers you promised.

Startups & investors

Do reverse due diligence: talk to founders an investor backed, including ones whose companies struggled.

Startups & investors

Send monthly investor updates. Share bad news early, together with what you are doing about it.

Startups & investors

Use AI to draft data room answers, then have the legal and financial documents reviewed.

Startups & investors

Ask an AI to play your toughest investor and grill you on competition, defensibility and "why now".

Startups & investors

Product Hunt's homepage features are chosen by its editors, so prepare your own audience and launch assets — upvotes alone are not a plan.

Startups & investors

Never make AI claims your product cannot back up. The FTC has taken action against companies over inflated AI marketing claims.

Customer service

Paste the customer's message and your policy, then ask for a reply that applies the policy kindly.

Customer service

For angry customers, ask for a calm reply that acknowledges, explains and offers one clear next step.

Customer service

Ask what not to say in a tricky reply. It is often more useful than what to say.

Customer service

Paste five tickets about the same problem and ask for one clear help article that answers them all.

Customer service

Tell the model your refund and replacement limits so it never promises more than policy allows.

Customer service

Ask for a two-line and a step-by-step version. Some customers want the short answer, others want every step.

Customer service

If you deploy a chatbot, always give customers a clear way to reach a person.

Customer service

Never paste customers' full card numbers or ID documents into an AI tool.

Teaching & learning

Ask for an explanation at three levels — child, student, expert — then keep the one you need.

Teaching & learning

Ask for quiz questions with answers, plus why each wrong answer is wrong.

Teaching & learning

Use the AI as a tutor: ask it to question you instead of giving you the answer.

Teaching & learning

For a lesson plan, give the age, what they already know, the time available and the one thing they should be able to do afterwards.

Teaching & learning

Check every fact in teaching material before it reaches students.

Teaching & learning

Ask for the common misconceptions about a topic, then teach directly against them.

Teaching & learning

Turn notes into flashcards and review them over several days. Spacing beats cramming.

Teaching & learning

Ask for a worked example followed by similar practice problems.

Privacy & safety

Never paste passwords, card numbers, API keys or government IDs into a chat.

Privacy & safety

Remove names and contact details before asking for a summary, unless you are allowed to share them.

Privacy & safety

Check whether your AI tool uses your chats for training, and switch it off for sensitive work where you can.

Privacy & safety

Use work-approved tools for work data. A personal account may not meet your employer's rules.

Privacy & safety

Treat AI output about health, money or law as a starting point, never as the decision.

Privacy & safety

Watch for instructions hidden in documents or web pages you ask an AI to read. They can try to redirect it.

Privacy & safety

Do not upload other people's photos or private messages without their consent.

Privacy & safety

Put a human review step before anything is published. Mistakes with your name on them are yours.

Privacy & safety

In the EU, AI Act transparency duties have applied since 2 August 2026 — for example, telling people when they are talking to an AI system. Check which apply to you.

Privacy & safety

Keep a record of the prompts and sources behind important decisions, so you can explain them later.

Workflow

Save prompts that work, with [BLANKS], so next time you fill in details instead of starting again.

Workflow

Batch similar jobs: write ten product descriptions in one session with one set of instructions.

Workflow

Use projects or custom assistants for recurring work so the background details are always loaded.

Workflow

Draft with AI, finish by hand. The last part — your judgement — is what makes it yours.

Workflow

Ask for a checklist version of any process you repeat.

Workflow

Time-box it: if three attempts do not work, change the approach or do it yourself.

Workflow

Keep a facts file about your business — prices, policies, tone — to paste at the start of a chat.

Workflow

Paste meeting notes and ask for the follow-up email to attendees, with decisions and owners.

Workflow

Once a week, review where AI saved you time and where its work needed redoing. Adjust your prompts accordingly.

Workflow

Use voice input to brain-dump your thoughts, then ask the AI to structure them.

Cost & models

Use small, fast models for simple jobs like sorting and short rewrites, and larger models for complex reasoning.

Cost & models

A clear prompt that works first time is cheaper than a short one that needs three retries.

Cost & models

Where your platform supports prompt caching, keep long fixed instructions at the start so repeated context can be reused.

Cost & models

For important tasks, try the same prompt on two models. Strengths differ by task.

Cost & models

Cap output length on API calls so a runaway answer cannot cost more than you expect.

Cost & models

Build a small test set and measure quality on it before switching models to save money.

Cost & models

Send non-urgent API jobs as batches where your provider offers discounted batch processing.

Cost & models

Track cost per useful output, not cost per call.

Compliance

For regulated outputs, require the model to cite the source document and section for every claim — then check each one.

Compliance

Ask the AI to flag sentences that could be read as financial, legal or medical advice.

Compliance

Keep a human in the loop for decisions that affect people's jobs, credit, housing or health.

Compliance

When AI helps write public content, keep the claims you can prove and cut the rest.

Compliance

Label AI-generated images or audio where platforms or the law require it.

Compliance

Ask for alt text for images and a plain-language summary of complex documents. Accessibility is part of the job.

Compliance

Do not ask AI to reproduce song lyrics, book passages or paid content you do not have rights to.

Compliance

Keep consent records for marketing messages. AI can draft the message; it cannot create the consent.