Build your own small tools.
You do not have to buy this from anybody — not from us either. Most of what AI is genuinely good at, on an ordinary working day, is small: the fifteen minutes you lose to the same shuffle every morning, the document you re-read four times, the question you answer twice a week. Small things, built by the person who actually does the work, compound faster than anything you can purchase.
Here's how to build them. The plays are free, the prompts are copy-and-paste, and none of it needs a budget approval.
Five rules before you build anything
Start with what you did three times this week
Not the most annoying task — the most repeated one. Repetition is what makes a small tool pay for itself, and it's the only reliable signal you have. If you can't name something you did three times, you're not ready to build; you're ready to watch for a week.
Do it by hand ten times first
You cannot describe a process you've never consciously performed. Doing it manually and paying attention is the mapping phase, and it's where you discover the exception nobody mentions. Skipping it is why most personal tools get abandoned in week two.
Write the stop condition before the first line
Decide up front what makes this thing wrong, and what you'll do when it is. "If it can't find the number, it says so instead of guessing." A tool without a stop condition doesn't fail — it fails quietly, which costs more.
Keep yourself on anything that leaves the building
Drafting is the machine's job. Sending is yours. Anything touching a customer, a supplier, money, or a commitment gets your eyes on it before it goes — not because the draft is bad, but because your name is on it.
A prompt you reuse is a tool. Save it.
The moment something works, put it in a file. Give it a name and a date. That file becomes your personal toolkit faster than you'd believe — and in a year it's worth more than any software you subscribed to in the same period.
The plays, in the rhythm you already work
Eleven small builds. Open one, follow the steps, copy the prompt. Nothing here needs anything beyond a chat assistant and the files you already have.
Daily · 5 minutes · updated Aug 2026
The end-of-day handoff
Close the day by telling the machine what happened. Open tomorrow already briefed instead of reconstructing.
What it does: turns five minutes of unstructured talking into a clean note — what moved, what's stuck, what needs you first thing. You stop starting each morning by remembering.
- Voice-to-text (already on your phone)
- A chat assistant
- One running note file
- 5 minutes at the end of the day
- At the end of the day, talk for two or three minutes into voice-to-text. Don't organize it. Ramble.
- Paste the mess into your assistant with the prompt below.
- Keep the output in one running file — one heading per day.
- Tomorrow, paste yesterday's note back in and ask: "what did I say I'd do that I haven't?"
Below is an unstructured end-of-day dump from my work. Turn it into a short handoff note with exactly four sections: MOVED — what actually progressed today STUCK — what is blocked, and on whom or what TOMORROW, FIRST — the two or three things that must happen first, in order WATCH — anything that smells like it will become a problem Rules: use my own words where you can. Do not invent anything I did not say. If something is ambiguous, put it under WATCH with a question mark rather than guessing. DUMP: [paste]
Daily · 10 minutes · updated Aug 2026
The triage pass
Sort the inbox by what actually needs your judgment, then let the machine draft the rest for your signature.
What it does: separates the messages that need you from the ones that need a reply. Most inboxes are 80% the second kind, and that 80% is what eats the morning.
- Your inbox, before you start replying
- A chat assistant
- 10 minutes
- Copy the subject lines and first lines of everything unread into one block.
- Run the prompt. You'll get three buckets back.
- Handle the "needs you" bucket yourself, properly, with a clear head.
- For the routine bucket, paste each one in and ask for a draft in your voice — then read it, fix it, send it. Never send unread.
Here is a list of unread messages (subject + opening line). Sort every item into exactly one bucket: NEEDS ME — requires my judgment, a decision, a relationship, or money ROUTINE — a reply is needed but the answer is predictable NOISE — no reply needed; note it and move on For each NEEDS ME item, add one line: what decision is actually being asked of me. Do not draft anything yet. Do not guess at context you were not given. MESSAGES: [paste]
Daily · as needed · updated Aug 2026
Interrogate the document
Stop reading contracts, invoices and reports front to back. Ask them questions instead.
What it does: turns a document you'd skim into one you actually understand, in less time. Works on anything dense: agreements, statements, policies, long reports, technical quotes.
- The document, as a file or pasted text
- A chat assistant that accepts uploads
- 5 minutes per document
- Paste or upload the document.
- Run the four questions below — they work on almost any document type.
- Then ask your own real question, the one you opened the file for.
- Verify anything that matters against the document itself. The answer tells you where to look; it isn't the source of truth.
Read the attached document and answer these four questions, quoting the exact line you got each answer from: 1. What am I agreeing to or being told, in plain language, in under 100 words? 2. What are the dates, amounts and deadlines? List every one. 3. What is unusual here compared to a standard version of this kind of document? 4. What would cost me money or time if I missed it? If the document does not answer one of these, say "not stated" — do not infer.
Daily · before the call · updated Aug 2026
The ten-line brief
Never walk into a conversation cold again. Ten lines of context, generated from what you already have.
What it does: turns scattered history — the email thread, your notes, the last invoice — into a single page you can read in the two minutes before a call. This is the single highest-return thirty seconds of prep available to anyone.
- Whatever history you have: the thread, notes, numbers
- A chat assistant
- 2 minutes before the call
- Gather whatever you have on the other party: the thread, past notes, numbers.
- Run the prompt.
- Add one line yourself: what you want out of the conversation. That part is never the machine's job.
From the material below, write me a ten-line pre-call brief: - Who they are and our history in two lines - The last three things that happened between us, with dates - Anything unresolved or promised and not delivered — by either side - Two questions I should ask - One thing likely to come up that I should have an answer ready for Only use what is in the material. Mark anything you are unsure about with [check]. MATERIAL: [paste]
Weekly · 20 minutes · updated Aug 2026
The recurring-question catcher
Every question you answer twice is a missing procedure. Catch them, and you find the gaps in your own operation.
What it does: this is the highest-leverage habit on this page. Keep a one-line log of questions people ask you. At the end of the week, the repeats tell you exactly what to write down — and what to build next.
- One notes app you will actually open
- A week of one-line discipline
- 20 minutes on Friday
- All week, when someone asks you something, drop the question into a note. One line. No formatting.
- Friday, paste the list into the prompt.
- Write a short answer to the top repeat — once, properly — and put it where people can find it.
- Next week, that question stops arriving. Repeat until your week is quiet.
Here is a raw list of questions people asked me this week. 1. Group them by what they are really asking about, even when worded differently. 2. Rank the groups by how often they came up. 3. For the top three, tell me what document or decision is missing that would stop the question from being asked at all. A recurring question is a gap in the procedure, not a gap in the person — treat it that way. QUESTIONS: [paste]
Weekly · 30 minutes · updated Aug 2026
Map one process a week
Ten weeks from now your operation is documented, and you never sat down to "do documentation".
What it does: converts one undocumented process per week into something written, using conversation instead of a blank template. The machine interviews you — that's the trick. Answering questions is far easier than facing an empty page.
- Voice-to-text
- 30 quiet minutes
- The process you would hate to explain from a sickbed
- Pick the process you'd most hate to explain if you were off sick.
- Run the prompt and answer the questions out loud into voice-to-text.
- Ask for the result as a numbered procedure with the exceptions listed separately.
- File it. Next week, the next process.
You are going to interview me about one process in my work so it can be written down properly. Ask me ONE question at a time, and wait for my answer before the next. Cover: what triggers it, every step in order, who owns each step, what systems are touched, what the handoffs are, what usually goes wrong, and what the exceptions are. Keep going until you could hand this to a new person. Then write it up as a numbered procedure, with a separate section for exceptions and one for "what to do when it goes wrong." The process is: [name it]
Weekly · 15 minutes · updated Aug 2026
The template that writes itself
You've already written the perfect version of that email. Five times. Turn them into one reusable thing.
What it does: extracts your own voice and structure from work you already did, instead of adopting somebody else's template. The output sounds like you because it came from you.
- 4\x{2013}5 past examples from your sent folder
- A chat assistant
- 15 minutes
- Find four or five examples of the same kind of message you send often — quotes, follow-ups, updates, rejections.
- Run the prompt.
- Save the template with its variables marked. Reuse, adjust, done.
Below are several messages I have written for the same purpose. 1. Identify the structure they share — the moves I make, in order. 2. Note the tone: sentence length, formality, how I open and close. 3. Produce one reusable template that matches how I actually write, with [VARIABLES] marked clearly. 4. Tell me which of my examples was strongest and why. Do not make it more formal or more polished than my originals. Match me. EXAMPLES: [paste]
Weekly · 10 minutes · updated Aug 2026
The honest weekly review
Not a status report for someone else. A short, unflattering read of your own week.
What it does: uses your daily handoff notes as raw material, so the review costs almost nothing to produce. The value is in the third question — the one about what keeps reappearing.
- Your daily handoff notes from this week (play one)
- A chat assistant
- 10 minutes on Friday
- Paste the week's daily notes in together.
- Run the prompt.
- Act on exactly one thing from the output. One. Ten weeks of one thing beats a list you never touch.
These are my daily notes from one week. Give me a short, honest review: 1. What actually moved forward, as opposed to what stayed busy? 2. What did I say I'd do and not do — and does that pattern repeat? 3. What appears in more than two days' notes? That thing is either important or unresolved. Which is it? 4. If I could only fix one thing next week, what should it be? Be direct. Do not soften it and do not congratulate me. NOTES: [paste]
Monthly · 45 minutes · updated Aug 2026
The pattern read
One month of records, read at line level, looking for what a single document can never show you.
What it does: finds the repeats and outliers hiding across a month of invoices, orders, tickets or jobs. This is the thing humans genuinely cannot do by eye, because the pattern only exists in aggregate.
- One month of records exported to a spreadsheet \x{2014} messy is fine
- A chat assistant that accepts file uploads
- 45 minutes
- Export the month to a spreadsheet — whatever you already have is fine, it does not need to be clean.
- Upload it and run the prompt.
- Verify every flagged item against the original record before you act. Every one.
Attached is one month of records. Analyse at line level and report: 1. Anything that looks like a duplicate — same item, same party, close dates 2. Prices or amounts that differ from the pattern for that item or party by more than 20% 3. Anything that repeats more often than it should for its type 4. The five largest items, and whether each looks normal for its category For every flag: show me the exact rows and say what made it suspicious. Where you are unsure, say so — I would rather check ten than miss one. Do not calculate totals I did not ask for.
Monthly · 20 minutes · updated Aug 2026
The repetition audit
What did you do more than four times last month? That's your build list, written by your own behaviour.
What it does: stops you guessing at what to build next. You don't choose the project — the month chooses it for you, and repetition is the only vote that counts.
- Last month: daily notes, sent items, calendar
- A chat assistant
- 20 minutes
- Pull together the month's daily notes, sent items, and calendar.
- Run the prompt.
- Take the top item and build a small tool for it — usually just a saved prompt.
- Next month, it shouldn't be on the list. If it is, your tool didn't work; find out why.
From the material below, list every task or request that happened four or more times last month. For each one give me: - roughly how often it occurred - how long it takes each time, if you can tell - whether the steps are the same every time or vary - whether it needs judgment, or just needs doing Then rank them by "same steps every time × happens most often" — that ranking is my build list. Explain the top choice in two lines. MATERIAL: [paste]
Monthly · 15 minutes · updated Aug 2026
Tend the prompt library
The file where your working prompts live is the actual asset. Once a month, prune it.
What it does: keeps the toolkit sharp instead of sprawling. Most people's prompt collections die of clutter, not of failure — they simply get too messy to open.
- Your prompt file
- 15 minutes, once a month
- Ruthlessness
- Open the file. Delete anything you haven't used in two months. Be ruthless.
- For the ones you use constantly, run the prompt below to tighten them.
- Add one line to each: what it's for, and when it fails.
- Keep it in plain text somewhere you'll actually find it. Not in a chat history.
Here is a prompt I use regularly. Improve it without changing what it does: 1. Remove anything vague that could be read two ways 2. Add explicit instructions about what to do when information is missing 3. Make the output format unambiguous 4. Tell me the most likely way this prompt gives me a wrong answer, and add one line that guards against it Give me the rewritten prompt, then a two-line note on what you changed and why. PROMPT: [paste]
Nothing filed under that rhythm yet.
What you actually need — and what you don't
Roughly everything on this page runs on the first two. The rest are for later, if ever.
A chat assistant
The workbench itself. Any of the serious ones will run every play here. Pick one and learn it properly rather than sampling five — fluency in one beats a shallow tour of the lot.
Voice-to-text
Already on your phone and your machine. It is the single biggest unlock on this page: you can talk through a messy situation five times faster than you can type it, and messy input is fine — that's what the assistant is for.
A plain text file
Where your prompts live. Not a chat history you'll never scroll back through, and not a fancy app. A file with a name, in a folder you open. Boring beats clever here.
A spreadsheet
Still undefeated for anything with rows. The pattern read works because your records are in a grid — if yours aren't, getting them there is more valuable than any AI you could point at them.
A workflow runner
The tools that connect apps and run things on a schedule — n8n, Zapier, Make and the like. Genuinely powerful, and genuinely premature for most people. Do not reach for one until you've run a workflow manually ten times and it hasn't changed. Automating an unstable process just makes the instability faster.
A document notebook
Tools that let you load a body of your own documents and ask questions across all of them. Worth it once "which document said that?" becomes a weekly question. Before that, uploading a file to a chat covers you.
Notice what isn't on this list: anything you have to buy from a consultant. If somebody's first answer to a fifteen-minute problem is a platform, they're selling, not helping.
Five ways this goes wrong
Building for the exception
You'll be tempted to handle the weird case first, because it's interesting. Build for the ordinary case, handle the exception yourself, and notice how rarely the exception actually shows up.
Trusting the confident answer
The dangerous output isn't the obviously wrong one — it's the plausible one you stop checking around week four. Decide in advance which outputs you always verify, and keep verifying those.
Letting it invent
If you don't say "tell me when you don't know", you will eventually get a number that looks right and isn't. Put that instruction in every prompt that touches facts. It's in all of the ones above for exactly this reason.
Tool-shopping instead of building
An afternoon comparing apps feels productive and produces nothing. The first version of every play here should be built with what you already have open.
Trying to fix everything at once
One play, run for a month, beats eleven tried for a week. The compounding comes from the habit, not the cleverness.
Why we're giving this away
Because it's true, and because you'd find it out anyway. Most of what people pay for in the name of AI is work they could have done themselves in an afternoon, and the ones who learn to do it are the ones who eventually recognise the problems genuinely worth outside help — the ones that cross departments, touch money, or need something running when nobody's watching.
If the workbench outgrows itself, that's the conversation to have. Not before.