AI prompts for real work: a map of the tips and guides

Background

Most advice on prompting is about prompting. This page is organised by the job instead. Nobody opens an AI tool to practise prompt engineering. They open it because there's a policy to turn into a quiz, a contract that's come back with changes, or a formula that won't behave.

Every tip below follows the same shape: say what you're handing over, say what you want back, say what to do when something's missing, and then check the result yourself. The details change from job to job. That bit doesn't.

The prompts work in Claude, Copilot or ChatGPT unless a tip says otherwise. Where a feature belongs to one tool, like making a downloadable file, the tip says so and is tagged.

Two habits before you start:

• Check what you're pasting. If it's personal, clinical, commercial or confidential, use a tool your organisation has approved for that data, or redact it first.
• You sign it off, not the tool. Anything with your name on it gets read by you.

Setup

Find the job closest to yours. Each line says what the tip does and the check that goes with it.

Make something

• A PDF you can download: name the layout and give it a fallback. Tool-specific, so test it first.
• A diagram from a process: ask for Mermaid, then render it before you trust it. Listing the steps out of order shows what you forgot.
• A template from your best document, or match the template you already have. Check no real names survive.
• A mail merge letter: find the parts that change. Test three rows first.
• A long document, section by section, with fixed structures per document type.
• Alt text written for what the image is for. A person checks every one.

Clean up and protect

• Redact with placeholders: replace the names yourself, then ask the AI what you missed. See also redacting before an HR case.
• Blank out words for a training exercise: check the answer key against the original.
• Plain English for one named reader: compare dates and obligations line by line afterwards.
• A cleaning plan for a messy spreadsheet: ask for the method, run it on a copy, compare row counts.

Make sense of what you've got

• Messy text into a table: "NOT STATED" for gaps. Count the rows.
• What changed between two versions: quoted differences, with tracked changes as the main source.
• A spreadsheet formula: explain, fix, then test on three rows.
• Meeting notes in a fixed shape and minutes that read the same.
• One itinerary from booking emails: check every time against the originals.

Getting better answers

• Make it ask you questions first.
• Show it examples instead of describing the rule: use a script for long lists.
• Ask it to list its assumptions: a prompt for your review, not a guarantee.
• Give it context every time, and save the good prompts as templates.
• Same prompt, different answer: five things to check in order.

Admin

• A renewal calendar from your contracts and the clauses worth a lawyer's time. Neither is legal advice.
• Answer "can I?" questions from the policy: quote the clause, or say it isn't covered.
• An FAQ from real questions and a handover note before leave.
• Chaser emails as a tone ladder, an SOP from how it's really done, a recurring agenda and triaging a shared inbox.
• A customer reply that ends the thread and a complaint as a neutral timeline.

Operations and ways of working

• Find where a process waits, run a pre-mortem and turn an incident into a checklist.
• Roll up team updates, with evidence and check a rota instead of building one.
• Running a small project without the paperwork, meetings that earn their slot and documentation your team will read.
• A delegation brief, checked for gaps.
• A RACI from a workshop's sticky notes that settles who decides.

IT

• A knowledge base article from a ticket, ticket triage and a runbook from tribal knowledge.
• APIs and webhooks, and why most integrations need both: two prompts for reviewing a design and reading a provider's docs.
• PowerShell with a dry run first: read the script, and know what's reversible.
• Log triage and comparing two environments: the tools find the facts, the AI ranks them.
• An access review and a licence audit.
• A vendor's sales deck, read for what it leaves out, turned into questions to ask.
• Alert noise, incident updates for two audiences and a rollback plan with every change.

Finance: month end, budgets and suppliers

The rule for every finance prompt: the spreadsheet does the sums, the model writes the words.

• Never let the model invent a number and lock the numbers before the commentary.
• Sanity-check a cash-flow forecast and rehearse with a sceptical finance director.
• Bank rec exceptions, accruals against last month and reconcile before you summarise.
• Overlapping suppliers, quotes compared like for like and questions for a new supplier.
• A variance narrative, budget and actuals into a P&L and invoice triage.
• Ask for the boring version of a forecast first, before the confident one.

HR: hiring, reviews and leavers

Strip names before anything goes in (see redacting before an HR case), and a person makes every decision about a person.

• A job description a candidate would recognise, an interview scorecard from the role and rejection emails from the same criteria.
• A performance review that isn't a wall of adjectives.
• Both sides in investigation notes before the summary is final.
• Exit interview themes, never a summary per leaver.

New starters

• An onboarding checklist, not a welcome email.
• A first-week plan, a first-month plan and a glossary of the team's jargon.
• Asking for help in a new job and a one-page "how we work".

Marketing and web

• Check you can measure a campaign before you write it.
• A content brief built on search intent.