Prompts that actually do something
One rule here: every entry is a prompt you can run. No tips that tell you to “leverage AI” without showing you how. Copy the text, paste it into the tool it was written for, and you should have something usable inside a minute.
The Corner Office
draft a board update from my notes
A generic AI prompt produces board-update prose that sounds like every other board update — vague verbs, no numbers, nothing a director could act on. This one forces structure and bans the filler words that make updates unreadable, and it explicitly tells the model to flag gaps instead of guessing.
build me an inbox triage system
Generic inbox-zero advice ignores that everyone's actual failure mode is different — some people over-triage, some under-respond. This prompt makes the model diagnose your specific pattern from your own description before prescribing anything, so the system fits the problem you actually have.
build a hiring scorecard for a role I've never hired before
Off-the-shelf interview questions produce inconsistent scoring because interviewers interpret 'strong communicator' differently — the rubric anchors here fix that by forcing a shared definition of each score. It also builds the scorecard around the specific failure mode you're worried about, which a generic template can't do.
Sales & RevOps
Marketing & Creative
turn a real customer conversation into language I can use
Customer research usually gets summarised into themes, and the summarising is exactly what throws away the useful part — the specific unglamorous phrasing people actually use. Section 5 is the one teams flinch at: discovering that the feature you lead with is the one they skipped past is worth more than any amount of positive feedback.
do a competitive teardown of a rival product
A one-shot 'compare X vs Y' prompt gets you marketing-page fluff repeated back at you. This version forces sourcing and explicitly permits 'unverified' as an answer, which is the difference between a teardown you can defend in a meeting and one that gets torn apart in it.
Finance & Ops
work out what a vendor will actually charge me
Pricing pages are designed to produce one impressive number and defer everything that determines your actual bill. Sections 3 and 4 are where the money hides — the minimum you can really spend, and the thing you assumed was included — and section 7 turns a vague unease into a specific list you can put in front of a salesperson.
work out what I'm actually spending on AI and what to cut
AI spend fragments across per-seat subscriptions, usage-based API bills, and things bundled invisibly into other products, so almost nobody has a real number. Section 5 is the one worth the exercise — usage-based cost tied to a metric you're deliberately growing is a bill that arrives without a decision ever being made.
sanity-check our pricing before we raise it
This beats asking for 'pricing advice' because it anchors every answer to your actual numbers and actual reason for the change, instead of returning a generic pricing-strategy essay. The forced early-warning metric and the customer-facing sentence are the two things people usually forget to prepare before they announce a price change.
skim a contract for red flags before I sign
Most people either skip the contract entirely or read it start to finish without knowing what 'unusual' looks like — this triages severity so a lawyer's time gets spent on the 3s, not the 1s. The explicit ban on inventing legal terminology matters because a confidently wrong answer here is worse than no answer.
turn my KPI dashboard numbers into a written update
A raw dashboard export tells you what happened but not what it means — this prompt forces a single headline instead of a metric-by-metric recap nobody reads past line three. Marking inferred causes as unconfirmed keeps it honest instead of turning correlation into a confident story that's wrong.
Any Role
turn a screenshot into something I can actually act on
Image reading is confident by default and wrong in a specific way — it fills in truncated labels and half-visible figures with plausible values, which is undetectable unless something is explicitly hunting for it. Section 3 is the whole safeguard, and section 2 is the reason it's useful: you took the screenshot because something caught your eye, and naming it is often the actual answer.
understand an automation I inherited before it breaks on me
Inherited automations fail at their weakest external dependency, and the person who knew which one that was has usually left. Section 3 is the map nobody made, and section 6 matters because the honest limits of a static read — you cannot see run history in a JSON export — are what stop you believing you understand something you don't.
give an agent a budget and make it account for what it spent
Agents fail expensively and quietly — the run either stops early and reports as though it finished, or burns through far more than the task warranted with no visible ledger. Forcing a spend-got-next report at each boundary makes the cost visible while you can still intervene, and the closing question about what to budget next time is the only way these estimates ever get better.
turn a long email or chat thread into an actual decision
Long threads persist because the disagreement was never diagnosed — people argue about a conclusion while actually differing on risk appetite or on who has authority. Section 4 is the whole point: naming which kind of disagreement it is usually collapses a week of messages into one reply.
write a one-pager a busy person will actually read
Most internal proposals bury the ask under three paragraphs of context, on the theory that the reader needs to understand before they can decide — but a busy reader stops before reaching it. Leading with the ask and pre-answering the objection reverses that, and the final self-critique catches the one line that will get picked apart.
work out what only I can do, and what I'm doing anyway
Delegation advice fails because it treats the bottleneck as identifying tasks, when the real bottleneck is an over-full 'only I can do this' bucket that feels accurate from the inside. Forcing a fourth bucket for work that should simply stop is what makes this different from a to-do list — and naming the real reason you haven't handed something over is usually more useful than the sorting.
split a big job across several AI workers without getting mush back
Splitting work across parallel workers fails in a specific and predictable way — the pieces come back individually reasonable and collectively inconsistent, because they were never given the same definitions. Sections 3 and 5 exist entirely to catch that, and section 1 is there because the honest answer is sometimes that the work is sequential and the split will make it worse.
get what's in my best person's head into something others can use
The reason expertise never gets documented is that experts genuinely cannot see their own tacit knowledge — asking someone to write down how they do it produces the steps and loses the judgement. Interviewing around a specific recent instance, and refusing to accept 'you get a feel for it', is what surfaces the part that actually matters.
write down a decision so the next person doesn't have to redo it
Decisions get made in meetings and stored in memory, so six weeks later the same debate reopens because nobody can recall what was ruled out or why. Section 3 does the heavy lifting — a record of rejected options is what actually prevents the rerun, and it is the section every template omits.
rehearse a difficult conversation before I have it
Preparing what you'll say is a different skill from saying it under mild stress while someone reacts, and only the second one is what happens in the room. The rules constraining the roleplay matter more than the prompt itself — a model that folds after one good argument teaches you nothing, because the real person won't.
turn a new policy into the questions people will actually ask
Policies are written by people who already know the reasoning and read by people who don't, which is why the same three questions arrive within an hour of every announcement. Section 4 is the one worth the exercise — the clause that reads as an accusation is invisible to whoever wrote it and obvious to everyone receiving it.
write the handover so things don't stall while I'm away
Handovers get written as a list of what you're working on, which is the least useful framing, because the person reading it is looking for what to do when something goes wrong at 4pm on a Thursday. Leading with the break-glass section and being specific about what not to interrupt for is what actually buys you an uninterrupted week.
stress-test a process before I automate it
Automations fail because the written process is a simplified version of the real one, and the simplifications are invisible to the person who wrote it down. Sections 1 and 4 are the two that matter — undocumented human judgement is why it breaks on day one, and silent failure is why nobody notices for a fortnight.
check a claim properly before I repeat it to anyone
The most common failure in fast-moving news isn't invention, it's staleness — a real announcement, accurately reported, that was reversed two days later while the coverage kept circulating. Sections 3 and 4 catch that specific failure, which no amount of general scepticism will, because the thing you're checking genuinely was true once.
read a long announcement and find out what changed for me specifically
Vendor announcements are written to be impressive to everyone, which makes them useless to anyone in particular, and the thing that actually affects you is usually a conditional buried well below the headline. Section 3 is the highest-value part: most disappointment with new features comes from a region, tier or admin gate mentioned once in passing.
compare two tools on my actual use case, not their feature lists
A feature comparison tells you what two tools do; it never tells you which one fits your volume, your constraints and your exit. Section 5 is the honest test — being shown the profile of the person who should choose differently is the fastest way to discover you're that person.
build a repeatable test for an AI task I keep re-running
Models change underneath you constantly — silent alias rebases, new defaults, version bumps — and without a baseline you're left comparing today's output to a memory of how it used to feel. Section 1 is where most homemade evals fail: a test set with no refusal case and no edge case passes every model forever and gives false confidence.
understand something well enough to present it tomorrow
Summaries make you feel informed and leave you unable to survive the first follow-up question, which is the only part of presenting that's actually hard. Sections 5 and 6 are the ones worth the prompt — knowing in advance which question will expose you, and which confident sentence will be wrong, is what separates prepared from rehearsed.
write the message I've been putting off for days
The reason a message sits unsent for four days is almost never that the words are hard — it's that the sender hasn't admitted what they're avoiding. Naming the fear first is what unblocks the draft, and the three-version ending removes the last excuse, which is uncertainty about tone.
find the repeating task buried in my calendar
Everyone believes they know where their week goes, and almost nobody is right, because the expensive pattern is spread across entries with different names. Section 3 is the one that surprises people — the prep and follow-up around a meeting routinely costs more than the meeting, and it never appears on the calendar as anything at all.
turn something that went wrong into a checklist that stops it recurring
Most post-mortems produce a narrative and a resolution to be more careful, neither of which survives a busy week. The three constraints in section 4 are doing the real work — a checklist item you can answer 'yes' to out of habit isn't a control, and section 6 exists because the dangerous thing about any checklist is the coverage people assume it has.
run a pre-mortem before I commit to something big
Asking a model for risks gets a hedged list nobody acts on, because listing risks and inhabiting a failure are different cognitive tasks. Forcing three separate committed accounts surfaces different failure modes each time, and lens 3 does the uncomfortable work — most failures were foreseen by somebody whose warning didn't land, which is a fixable organisational problem rather than an unlucky one.
find out what my AI can already reach, and what I'm not using
Most people use a tool with six live connectors as though it were a blank chatbot, because nothing in the interface ever tells you what it can reach. Asking the model to audit itself surfaces the gap in about thirty seconds, and section 4 is where it earns its place — it turns an abstract capability list into five jobs you could start today.
turn a rambling voice note into three things I'll actually do
Voice notes are where the honest version of your thinking lives, and a normal 'summarize this' prompt destroys exactly what makes them useful by flattening the hesitation out. Sections 3 and 4 do the real work — naming the fake commitment and the thing you keep avoiding is the part you cannot do for yourself, because you are the one avoiding it.
make AI argue against a plan I've already decided on
Ask any assistant to review a plan you've announced you've decided on and it will validate it, because agreement reads as helpfulness. Explicitly assigning it the opposing side removes that pull, and section 4 is the one people flinch at — it names the motive underneath the plan, which is usually where the bad decisions actually come from.
get real questions out of a spreadsheet instead of a summary
Uploading a spreadsheet and asking what it says gets you the columns read back with confident adjectives attached. This forces actual computation and then spends most of its effort on the two things a human misses — the relationship nobody thought to test, and the question the data quietly cannot answer but will be used for anyway.
have one AI audit another one's answer before I trust it
Two models trained differently fail differently, so the second one catches what the first was confident and wrong about — which is exactly the failure a single model cannot self-report. Section 2 is the whole point: a fabricated specific sits inside otherwise-correct text at the same confidence level, and it is nearly invisible unless something is explicitly hunting for it.
make AI remember who I am in every future chat
Most people re-explain their job at the start of every chat, then get generic answers anyway because they under-explained it. Writing the context once into memory means every future conversation starts warm — and the read-back at the end catches the case where the model stored something subtly wrong, which you would otherwise never discover.
make deep research mode actually earn the ten minutes
Deep research modes fail most often because they're pointed at a topic instead of a decision, so they return a well-cited essay that changes nothing. Naming the decision first, and demanding the counter-case and the falsifier, turns ten minutes of compute into something you can act on — and section 6 is the one that saves you, because it tells you where the answer is thin.
turn a messy folder of documents into one briefing
Asking for a summary of many files gets you a blended average that quietly buries the one thing you needed to notice — that two documents disagree. This forces the contradictions to the surface with both sources named, which is the only part of the output a person genuinely cannot produce faster themselves.
get AI to write the reusable instruction file for a task I repeat
The reason repeat tasks never get delegated is that the instructions live in one person's head and writing them down feels like more work than doing the job again. The interview does the extraction for you, and step three is the part almost everyone skips — a set of instructions that has never been followed once is a draft, not a process.
give an AI agent a task without letting it run away
The failure mode of an agent isn't doing the task badly — it's doing forty minutes of confident work in the wrong direction, or taking an irreversible action nobody sanctioned. Forcing the plan and the stop conditions to be written before any work begins turns both into something you can approve or correct in thirty seconds, while it's still cheap.
prep me for a meeting in 5 minutes
Most meeting-prep prompts just summarize — this one forces a single objective and a real objection, which is the part execs actually forget to think through beforehand. It beats winging it because the model catches contradictions in your own notes that you'd miss skimming them thirty seconds before the call.
figure out if a new AI model release actually matters for my business
AI news is optimized for hype, so most 'should I care' prompts just get the hype reflected back with extra confidence. Anchoring the question to what you already use AI for turns a generic capability announcement into a specific yes/no you can act on, and forcing a single tier instead of a hedge stops you from bookmarking it and never deciding.
45 prompts and counting. New one every weekday in the daily brief. Researched as of August 2026.