Not theory — specific, everyday tasks with a real example prompt, what a good output looks like, and exactly what to check before you trust it. Each one names the human-review step explicitly, because that step is not optional.
Work & Communication
Meeting notes and follow-up
Who this is for: Anyone who sits in recurring meetings and owes a recap or action list afterward.
Problem it solves: Meeting notes are the first thing skipped under time pressure, and the first thing people wish existed a week later.
Inputs needed: A transcript or your own rough notes from the meeting.
Suggested tools: Any chat model, or a dedicated notetaker — see the Meetings & Knowledge tools page for a full comparison.
Workflow:
- Paste the raw transcript or notes, unedited.
- Ask for three sections: decisions made, action items with owners, and open questions.
- Ask it to flag anything ambiguous rather than guessing who owns what.
Example prompt:
Here are my raw notes from today's planning meeting: [paste notes]
Summarize as: (1) Decisions made, (2) Action items — with owner and rough
deadline if mentioned, (3) Open questions nobody resolved. If an action
item's owner is unclear, list it under "unassigned" instead of guessing.
Human review — not optional: Check every named owner and deadline against your own memory of the meeting — models will occasionally assign an action to the wrong person if the transcript is ambiguous.
Email drafting and triage
Who this is for: Anyone facing an inbox where most replies are variations of a few recurring types.
Problem it solves: Drafting is usually the bottleneck, not the underlying decision — you already know what you want to say.
Inputs needed: The email you're replying to, plus a one-line note on your intended answer or tone.
Suggested tools: Any chat model; ChatGPT and Claude both handle this well.
Workflow:
- Paste the incoming email in full.
- State your intended answer in plain language — don't ask the model to decide for you.
- Specify tone and length ("brief and warm," "formal, three sentences").
Example prompt:
Here's an email I received: [paste email]
My answer is: yes, but I need until Friday, not Wednesday.
Draft a brief, warm reply in three sentences or fewer.
Human review — not optional: Read the full draft before sending — never send an AI-drafted reply unread, especially anything touching commitments, dates, or pricing.
Performance review and feedback writing
Who this is for: Managers writing performance reviews, or anyone giving structured feedback to a peer or direct report.
Problem it solves: Turning a scattered list of observations into fair, specific, well-organized feedback is time-consuming and easy to write unevenly across a team.
Inputs needed: Your actual, specific observations about the person's work — not a request to invent feedback from nothing.
Suggested tools: Any chat model; keep sensitive personnel details as anonymized as your workplace policy requires.
Workflow:
- List your real, specific observations — concrete examples, not vague impressions.
- State the review's structure or template if your company has one.
- Ask for balanced, specific language — not generic praise or generic criticism.
Example prompt:
Here are my specific observations about [role, no need for real name]
over this period: [list concrete examples — strengths and growth areas].
Write this into a performance review section following a strengths /
growth areas / goals structure. Keep it specific to what I described —
don't add generic filler praise or criticism I didn't mention.
Human review — not optional: Verify every specific claim and example against your own records before finalizing — and check tone carefully, since performance reviews carry real consequences for the person receiving them.
Research & Learning
Research synthesis
Who this is for: Anyone who needs to turn several sources into one clear picture — students, analysts, consultants, curious readers.
Problem it solves: Reading five sources and holding the whole picture in your head is slow; a model can hold the structure while you verify the substance.
Inputs needed: The source documents or links themselves, not a request to "research this" from scratch with no sources provided.
Suggested tools: Perplexity, NotebookLM, Elicit, or Consensus for grounded, cited research — see the Research tools page for a full comparison.
Workflow:
- Provide the actual source material — paste it in, or use a tool with real-time web access.
- Ask for a structured synthesis: key claims, where sources agree, where they disagree.
- Ask it to keep every claim attributed to which source it came from.
Example prompt:
Here are three articles on [topic]: [paste or link].
Synthesize them into: (1) Where they agree, (2) Where they meaningfully
disagree, (3) Any claim only one source makes. Attribute every claim to
its source by name.
Human review — not optional: Spot-check the highest-stakes claims against the original source directly — synthesis can flatten a hedge ("may cause") into a certainty ("causes") if you don't ask it to preserve that language.
Learning a new topic quickly
Who this is for: Anyone starting from zero on an unfamiliar subject and needing an honest, structured on-ramp.
Problem it solves: Generic search gives you fragments; a good tutor-style prompt gives you a sequenced explanation you can push back on.
Inputs needed: Your actual starting knowledge level, stated honestly — this page's own 5-layer context pattern applies directly here.
Suggested tools: Any strong chat model — this is a conversational, iterative task rather than a single-shot one.
Workflow:
- State your real starting level ("I know basic algebra, nothing about this field").
- Ask for a short explanation, then ask it to check your understanding before going further.
- Ask explicitly for where experts disagree, not just the settled consensus.
Example prompt:
Teach me [topic] assuming I know [your actual background].
Explain the core idea in plain language, then ask me one question to
check I understood it before moving to the next layer. Tell me
explicitly where this field still has real disagreement among experts.
Human review — not optional: Treat specific facts, dates, and numbers as claims to verify separately — a model can explain a concept's logic correctly while still getting a specific figure wrong.
Document comparison
Who this is for: Anyone reviewing two versions of a contract, policy, proposal, or draft.
Problem it solves: Manually diffing two long documents for meaningful changes — not just wording — is tedious and error-prone.
Inputs needed: Both full documents, pasted in or uploaded.
Suggested tools: Claude or ChatGPT with document upload support; for anything with legal or financial consequences, this speeds up review, it doesn't replace it.
Workflow:
- Provide both versions in full.
- Ask specifically for substantive changes — dates, dollar amounts, obligations, deleted clauses — not stylistic edits.
- Ask it to quote the exact before/after text for each change it flags.
Example prompt:
Compare Document A and Document B [paste both].
List only substantive changes: dates, amounts, added/removed
obligations, deleted or added clauses. For each, quote the exact
before and after text. Ignore purely stylistic wording changes.
Human review — not optional: For anything legally or financially binding, a qualified professional should verify the comparison directly against both source documents — this is a drafting aid, not a substitute for that review.
Creating study guides and flashcards
Who this is for: Students, or anyone preparing for a certification exam or teaching themselves a technical subject.
Problem it solves: Turning a textbook chapter or lecture notes into an efficient study format (summary, flashcards, practice questions) is repetitive busywork a model handles well.
Inputs needed: Your actual course material — notes, a textbook excerpt, or a syllabus — not a request to generate content from a topic name alone.
Suggested tools: Any chat model; NotebookLM is particularly well-suited since it grounds its output in the source material you upload — see the Research tools page.
Workflow:
- Provide the actual material to study from, in full.
- Ask for a structured summary plus a set of flashcards (term/definition or Q&A pairs).
- Ask for a short set of practice questions, and to flag which topics they draw from.
Example prompt:
Here's my lecture material on [topic]: [paste notes/excerpt].
Create: (1) a one-page summary of the key concepts, (2) 15 flashcards
in question/answer format covering the material, (3) 5 practice
questions with answers, noting which section each one tests.
Human review — not optional: Verify factual claims and definitions against your actual course material or textbook — a model summarizing your notes can still introduce a subtly wrong definition, especially for technical or field-specific terms.
Building & Technical
Code help and debugging
Who this is for: Anyone writing or maintaining code, from complete beginners to working developers.
Problem it solves: Getting unstuck on a specific error or unfamiliar syntax is one of the highest-leverage uses of a chat model.
Inputs needed: The actual error message and the relevant code — not a vague description of the problem.
Suggested tools: Claude, ChatGPT, or an AI-native coding tool like Cursor — see the Coding tools page and the Techniques page's vibe-coding tab.
Workflow:
- Paste the exact error message in full, not a paraphrase.
- Include the relevant code, and enough surrounding context to reproduce the issue.
- Ask for an explanation of the cause, not just a fixed snippet — so you can catch it yourself next time.
Example prompt:
I'm getting this error: [paste exact error].
Here's the relevant code: [paste code].
Explain what's causing it, then give me the fix. I want to
understand the cause, not just copy a patch.
Human review — not optional: Run the suggested fix yourself and read the diff before committing — never paste in code you haven't read, especially anything touching authentication, payments, or data deletion.
Brainstorming and ideation
Who this is for: Anyone stuck at the blank-page stage of any project — naming, planning, structuring, or problem-solving.
Problem it solves: Generating a wide first pass of options is exactly what a model is good at; picking the right one is exactly what it isn't.
Inputs needed: A clear constraint set — audience, tone, length, what's already been tried and rejected.
Suggested tools: Any chat model; higher creativity/temperature settings if the tool exposes that option.
Workflow:
- State the real constraints up front, not just the open-ended goal.
- Ask for a wide first pass — 10-15 options, not 3.
- Pick your own favorites, then ask it to refine only those.
Example prompt:
I need 12 name options for [thing]. Audience: [who].
Tone: [tone]. Already rejected: [list]. Give me a wide, varied
first pass — don't converge on one style. I'll pick my favorites
and ask you to refine from there.
Human review — not optional: Treat the output as raw material, not a final answer — the value is in generating options fast, and the judgment on which one is right stays entirely yours.
Extracting structured data from documents and images
Who this is for: Anyone who needs to pull specific data out of a receipt, scanned form, screenshot, or PDF into usable, structured text.
Problem it solves: Manually retyping data from images or scanned documents is slow and error-prone; a model with vision input can read and structure it directly.
Inputs needed: The actual image or document, plus a clear spec of exactly which fields you need extracted.
Suggested tools: Claude, ChatGPT, or Gemini, all of which accept image/PDF uploads directly in their free or low-cost tiers.
Workflow:
- Upload the actual image or PDF — don't retype it yourself first.
- List the exact fields you need (e.g. date, vendor, amount, line items) rather than asking for a general summary.
- Ask it to output in a structured format (a table or CSV-style list) and to flag any field it couldn't read clearly.
Example prompt:
[Upload receipt/form/document]
Extract the following fields into a table: date, vendor name, total
amount, and each line item with its price. If any field is unclear
or unreadable in the image, write "unclear" rather than guessing.
Human review — not optional: Manually verify every extracted number against the original image, especially amounts and dates — OCR-style extraction from models can misread similar-looking characters (e.g. 0 vs O, 1 vs 7) and won't always flag its own uncertainty.
Business & Client-Facing
Customer support draft responses
Who this is for: Anyone handling recurring customer or client questions, from solo founders to support teams.
Problem it solves: Most support replies are variations on a small set of recurring issues — the answer is known, drafting a clear, on-brand version of it each time is the bottleneck.
Inputs needed: The customer's actual message, your policy or the correct answer, and your brand's tone.
Suggested tools: Any chat model for drafting; see the Writing tools page for tone/style-specific options.
Workflow:
- Paste the customer's message in full, including any frustration or specific details.
- State the correct resolution or policy — don't ask the model to invent a policy.
- Specify tone (empathetic, concise, formal) and whether to offer anything beyond the stated resolution.
Example prompt:
Customer message: [paste message]
Our policy/resolution: [state the actual answer, e.g. "we offer a
refund within 30 days, this order is on day 25"]
Draft an empathetic, concise reply. Acknowledge their frustration
in one sentence before the resolution. Do not offer anything beyond
what I've stated.
Human review — not optional: Never let a model invent a policy, discount, or promise on its own — read every reply before sending, especially anything involving money, refunds, or a commitment to the customer.
Data and spreadsheet help
Who this is for: Anyone working in spreadsheets who needs a formula, a cleanup, or a first-pass analysis.
Problem it solves: Remembering exact spreadsheet formula syntax or writing a one-off data-cleaning script is a common, low-stakes time sink.
Inputs needed: A description of your data's actual structure (column names, what's in them) and the exact outcome you want.
Suggested tools: Any chat model can write formulas and analysis logic; see Coding tools if the task grows into an actual script.
Workflow:
- Describe your columns and what's actually in them, including messy edge cases.
- State the exact result you want, not just the general goal.
- Ask for the formula or code plus a plain-English explanation of what it does.
Example prompt:
I have a spreadsheet with columns: Date, Customer, Amount, Status.
Status can be "Paid," "Pending," or blank. I want a formula that
sums Amount for rows where Status is "Paid" and Date is in the
current month. Give me the formula and explain what each part does.
Human review — not optional: Test any formula or script on a copy of your data first, and manually verify the result against a few rows you can check by hand — don't trust an aggregate number you haven't spot-checked.
Presentation and deck outlining
Who this is for: Anyone who needs to turn scattered points into a structured, presentable narrative.
Problem it solves: The hard part of a deck is usually the structure and narrative flow, not the slide design — that's exactly where a model helps most.
Inputs needed: Your raw points or notes, your audience, and how much time you have to present.
Suggested tools: Any chat model for the outline; see the Resource Library for free presentation-generation tools if you want AI-assisted slide design too.
Workflow:
- Provide your raw points, unorganized, and your audience.
- Ask for a slide-by-slide outline with one core idea per slide, not a wall of text per slide.
- Ask it to flag where you need a chart, data point, or example you don't have yet.
Example prompt:
Here are my raw points on [topic]: [paste points].
Audience: [who]. Time: [X minutes].
Give me a slide-by-slide outline — one core idea per slide, a
suggested visual for each, and flag any slide where I'd need a
data point or example I haven't provided.
Human review — not optional: Check that every data point or statistic in the final deck traces back to something you can actually source — a model can suggest a plausible-sounding stat that needs to be replaced with a real one before you present it.
Translation and localization
Who this is for: Anyone communicating with an audience in a language they don't speak fluently, or adapting content for a different region.
Problem it solves: Literal machine translation often misses tone, idiom, and cultural context — the useful version goes beyond word-for-word conversion.
Inputs needed: The source text, the target language, and context about tone and audience (formal/informal, regional dialect).
Suggested tools: Any strong general chat model handles this well for common languages; verify with a native speaker for anything high-stakes.
Workflow:
- Provide the full source text and target language.
- State the register — formal business letter vs. casual social post — and target region if it matters (e.g. Latin American vs. European Spanish).
- Ask it to flag any idiom or cultural reference that doesn't translate directly, rather than silently guessing.
Example prompt:
Translate this into [language, target region] at a [formal/casual]
register: [paste text].
Flag anything — idioms, cultural references, humor — that doesn't
translate directly, and explain your alternative choice for each.
Human review — not optional: For anything customer-facing, legal, or high-stakes, have a native speaker review the translation before it goes out — a model can be fluent and still miss cultural nuance or make a false-friend error.
Job postings and interview questions
Who this is for: Anyone hiring, from a solo founder making their first hire to a small team's hiring manager.
Problem it solves: Writing a job posting that's specific enough to attract the right candidates — and interview questions that actually test for the role — takes longer than it should for something you may only do occasionally.
Inputs needed: The real responsibilities of the role, must-have vs. nice-to-have skills, and anything specific about your team or work style.
Suggested tools: Any chat model handles this well; see Writing tools for tone-specific options.
Workflow:
- Describe the role's actual day-to-day responsibilities, not just a title.
- Separate must-have requirements from nice-to-haves explicitly.
- Ask for the posting plus 5-8 interview questions mapped to specific requirements, not generic questions.
Example prompt:
I'm hiring for [role]. Day-to-day responsibilities: [list].
Must-haves: [list]. Nice-to-haves: [list]. Our team's work style:
[e.g. "fully remote, async-first, small team"].
Write a job posting, then give me 6 interview questions, each one
mapped to a specific requirement it's meant to test for.
Human review — not optional: Review job postings for language that could unintentionally discourage qualified candidates (age-coded phrases, unnecessary requirements), and confirm every stated requirement is genuinely necessary for the role before publishing.
Resume and cover letter tailoring
Who this is for: Anyone job-hunting who needs to adapt one resume and cover letter to fit a specific role, rather than sending a generic version.
Problem it solves: Tailoring a resume to each job posting's actual language and requirements measurably improves response rates, but doing it by hand for every application is slow.
Inputs needed: Your real resume/work history and the actual job posting you're applying to — not a request to invent experience you don't have.
Suggested tools: Any chat model; be explicit that it must only use experience you actually provide.
Workflow:
- Paste your real resume/experience and the specific job posting.
- Ask it to identify which of your existing experiences best match the posting's stated requirements.
- Ask for a tailored resume summary and cover letter draft, using only what you actually provided.
Example prompt:
Here's my resume/experience: [paste].
Here's the job posting I'm applying to: [paste].
Identify which parts of my real experience best match this role's
stated requirements, then draft a tailored resume summary and a
cover letter. Use only the experience I've given you — don't invent
or exaggerate anything.
Human review — not optional: Read every line before submitting and confirm nothing has been embellished or invented beyond your actual experience — misrepresenting qualifications on a resume carries real consequences, and this step is non-negotiable.
Next: Tools
Now — which tool should you actually use for this?
ChatGPT vs Claude vs Gemini vs Perplexity, compared honestly, plus how to build a lean stack.
Continue to Tools →