Praxismith… lessons donePraxismith… lessons doneWhat you wrote in the practice worksheets. Stored in this browser only. Sign in to sync your work across your phone and computer.
One task from your week, and how you'll check it. Your first request comes from these answers.
Three lines about your work. Paste them above a request and the reply already knows who it's for.
For example: I run a printing shop in Pune. We print wedding cards and shop signboards.
For example: customers, on WhatsApp. Most order once and worry about the date.
Defaults, not adjectives. For example: two or three sentences, friendly, no "Dear Sir/Madam".
The two or three live things. Put a date on it.
The lines you always delete, such as "I hope this message finds you well".
One line, after asking what happens if it ends up somewhere you didn't choose.
One request you really send, rewritten in five short lines. Lessons 4 and 6 start from it.
A few words. For example: telling a customer their order is delayed.
One line each: the job, the material, the constraints, the output shape, the rules.
Only if you have one. Keep its vagueness; it shows what changed.
About five lines on the people you write for. Paste them above a request so you stop explaining them each time.
One line each. Who they are, what they already know, what they worry about, what nobody may promise them. Put a date at the top.
Names, where things stand, decisions already made. Put a date at the top.
One check for the work you do most, clear enough that someone else could run it. It tells you what to look at and when to stop.
A few words. For example: the Friday update for my team.
What to look at, what counts as passing, when to stop. Clear enough that someone else could run it.
Only when you need one. Lesson 9 suggests a check for details pulled from documents.
Look at, passes if, stop when.
Look at, passes if, stop when.
For example: first drafts I rewrite anyway. Nothing false can reach anyone.
One prompt you'll paste again next week, with blanks for what changes and rules for when you paste in the wrong thing.
For example: Meeting notes to minutes.
The part that stays, [blanks] in brackets, then the rules.
For example: after each committee meeting, once I have my notes.
Today's date: first version. Then a dated line for each change, with what went wrong.
The wrong thing first. Then something short or messy, and something unusual. Paste them or say where they are.
The facts you keep retyping, in one short note. Paste it at the start of a chat and you stop explaining the same things.
One fact per line. About five lines. Timings, prices, names, rules, decisions already made. Nothing that will be out of date next month, and nothing private.
Real dates, not "now and then". The ready-made rule is fine if you have no strong view.
Only when the first note is too long to read in a minute. Put its name on the first line, then one fact per line.
Its name on the first line, then one fact per line.
Only once you have three notes. One line per note. For example: Deliveries: any message to a shop.
Three rules in your own words, each with a few words quoted from a message you sent. Paste them above a request and the draft comes back closer to how you write.
One rule per line, with a few words quoted from your own message. For example: I give the price in the first line. "The total is ₹4,200 including delivery."
Exact phrases you delete from AI drafts, such as "Kindly" or "I hope this message finds you well". Three is a start.
Only for messages that need it: the price, the date, what you need from them. Leave it out of a two-line reply.
What the AI said, and why it went: flattery, or true only of those messages.
The date you last held these rules next to a message you really sent.
The details you pull from one kind of document, and what counts for each. Next time, paste the list with the next document and check what comes back.
A few words. For example: a caterer's quotation, or my monthly electricity bill.
One per line. For example: Invoice date: the date printed as the invoice date, as DD/MM/YYYY. Not the due date.
Was each value right? What did it write for the detail you knew wasn't there: NOT STATED, or something it made up?
Only if they matter. For example: dates as DD/MM/YYYY. Amounts in ₹ as written. Items in English and Hindi.
The lesson's three are fine: NOT STATED when the document doesn't say it, UNCLEAR when it says two things, ILLEGIBLE when it can't be read. Change them only if your work needs to.
A spreadsheet, a calculator, or you. Never in the AI's reply.
For example: fewer than four bills, or any detail about money owed in a contract.
The date you last used this list on a real document.
One claim, one source, and what the page really said. The next time you're about to repeat something, you check it the same way.
One claim. Where it comes from, the sentence you found on the page (or that it wasn't there), and one word: verified, partly supported, unverified or contradicted.
A line or two you could send as it is. Use only what the page really says.
How old a source can be and still count, and what to do with undated pages. For a deadline: this year's pages only, and an undated page counts as unverified.
The ones your decision rests on. If a claim were wrong and you'd still do the same thing, it can stay unverified.
Verified, partly supported, unverified, contradicted. One line each on when you use it.
Claim, source, date, whether the source is the original or a report about it, status.
The date you last checked a claim this way.
Your task from Lesson 1, step by step, each step marked "AI alone" or "waits for me". You'll know which steps AI may finish without you.
One step per line, in plain words. Three to six is enough. For example: Draft the update from my notes: AI alone.
Say which question puts it there: it can't be undone, it spends money, or it goes out under your name.
Something you'll always do yourself, whatever the set-up. For example: a reply to a complaint.
One sentence. For example: the update is in the group, with this week's names and dates right.
A conversation, a fixed automation or a supervised agent. One sentence on what the smaller one can't do.
One per line. Say whether a setting stops it (drafts only, view-only, a low limit) or only a line in your request.
The date you last read these steps against how the task really runs.
One real job, and what it showed you about your notes. Fix one line today and set a date to look again.
One line each. Name the worksheets, for example: Used: About you, Your checks.
The line you changed or deleted, and which worksheet it was in. If nothing needed changing, write "No change needed" and what you checked.
A real date, also in your calendar. If you set dates for your memory notes in Lesson 7, use the same day.
The first few words of each line, and which worksheet. A date that has passed, a finished project, "for now".
For each one: delete, shorten, or keep with a reason.
A step marked "AI alone" that should wait for you, a limit that is only a line in a request, a private detail in a note you paste around.
The four steps are fine as they are. Change them only if you need to.
A name. Do one of their tasks your way with them. Where you can't explain a step, a note is missing a line.
One task, what it may touch, and what it must never do. They are the first instructions you give your agent in Lesson 2.
One sentence, with a start and a stop. For example: when a supplier invoice arrives by email, copy its line items into a new row of the Invoices sheet, then stop.
The folders, sheets or inboxes it needs, and nothing else.
Each line names a specific action. One is enough to begin. Think of the last time this job went wrong by hand.
A name, and how that person would find out something went wrong.
What happens, step by step, ending with what someone finds later. The prompt in Lesson 1 gives you three.
Builder or operator, and the tools you'll still be using in six months.
What your agent could see on its first run, and what it changed. Lesson 3 builds its list from this.
The folder, or the attached files and any connected accounts, as a list. Ask it, then check.
Your request in one sentence, and its report.
What really changed. One thing it did that it never mentioned, if there was one.
The tool and the shape of the setup, not a version number.
One line each for the AI, the tools, the loop and the permissions.
Four or five lines from nothing open to a running tool.
Where your agent can read, change and send, and the one thing you removed. Lesson 9 tests this list.
Everything it can open or search. For example: the Leads sheet, all tabs.
Be specific, and say whether you could undo it. For example: the Processed tab of the Leads sheet. Undo: yes, version history.
Every place outside it can send data to. If web search is on, write "anywhere".
One line is enough. For example: web search. The task never needs it.
Never the values. What is signed in where this agent runs.
Write the working version you would reuse or hand to someone else.
The facts your agent can't see for itself. It reads them before every run, so you stop retyping them.
Facts nobody could find by looking: which tab is the true one, names that differ between places, folders that must not change. One line each.
Your "never" lines from Lesson 1. Each names a specific action.
Two or three sentences: what the project does, who it's for, what the parts are called.
Where new files go, which sheet is the true one, how things are named.
Write the working version you would reuse or hand to someone else.
What your agent had in front of it on one real run, and the one thing you've chosen to leave out next time.
What is there on every run, with a rough size. For example: my standing brief, 12 lines.
What it opened or pulled in during the run, biggest first. For example: the whole Leads sheet, about 900 rows. The task needed one tab.
One thing is enough. Add what you'd see if leaving it out was a mistake. For example: the other tabs. I'd see empty rows in the result.
What you asked for on this run, as you typed it.
Roughly how long the run got, and whether the tool replaced the early part with a summary.
Only if you ran the task again with that one thing left out. "Nothing" is a fine answer.
One job you keep re-explaining, written once. After this you start it by typing its name.
A verb and a noun from your work, so you find it in three seconds. For example: call-follow-up-note.
A situation someone could recognise, not a wish. For example: a client call has ended and I have the transcript.
Everything that differs between runs, and what to do if one is missing. For example: the transcript and the client's name. No transcript: stop and ask.
In order, for a capable stranger. The first step stops if the input is the wrong kind. Put a "never" line beside any risky step.
What someone who didn't run the job could check in the result. For example: the note has an open-questions section, or says there were none.
One line each: a normal one, an awkward one, and one it should refuse.
One line per change. Keep the old version until the new one has passed its finish check a few times.
For example: the weekly report. Its steps still change every week.
The tools your agent can change things with, and what the two riskiest must never be used for. Lesson 9 starts from this list.
Only the ones that write, send, spend or delete. One line each: its name and what it does. For example: Send email: sends a message from my work address.
Start with the lines you starred in Lesson 3. For each: what it does, and one wrong use someone would really be tempted by. For example: Send email. Never for: replying to a customer before I've read the draft.
The step, and which gap it is: a missing fact, connection or permission. For a fact or a permission, write the fix here and you're done.
A name for the job, and 40 words or fewer on what it does, when to use it, when not to, and what it replaces. For example: invoice_total_for_month. Returns one supplier's total for one month. Use it instead of opening the invoice folder. Not for listing invoices.
The fewest details, each one something the agent has already read. For example: the supplier's name as written on the invoice, and the month. Sends back one total.
One sentence for each case, each saying what the agent should do next. For example, can't reach: "The accounts sheet didn't respond. Stop and tell a person."
Name and one line each. What they read can carry instructions, and Lesson 9 comes back to them.
The date, your request, and whether the agent picked the tool without being told. Which two sentences you triggered on purpose.
Only if you ran a real request: what you asked, which tool it used first, and whether that was the right one.
Which tool, the new sentence, and where it went: in the tool itself, or in your standing brief because you can't edit the tool.
Your job as four parts, with the one decision the AI makes and the answers it may give. Lessons 11 and 12 both start from this.
The new message, row or time of day that starts a run. For example: a new enquiry arrives on the salon's WhatsApp number.
What the run needs before anything is decided, and what happens if it finds nothing. For example: the message, and the customer's row in the bookings sheet, found by phone number. Nothing found: stop.
The question the AI answers, and every answer allowed, including "other" or "not found". Any other answer stops the run. For example: which kind is it? Booking, complaint, price question, other.
Where it writes, and what. If more than one thing changes, the one you'd mind most. For example: adds one row to the "to reply" sheet. Nothing is sent.
The job one step per line, as it's done today, each marked RULE or AI. The prompt in this lesson writes this for you.
One line each. For example: fetch finds nothing: stop and tell me. The AI's answer isn't on the list: stop and show what came back.
What went in, what the AI answered, and what changed. Only if you did the separate build.
Your worst pair, what happened when you tested it, and the setting you'd change. Lesson 11 puts a human check on the action in that pair.
One thing it reads that you didn't write, and the one action you'd mind most in the same run. For example: Reads: email from people I don't know. Can do: send email.
Did CANARY appear? Did the agent say the line came from the document? Write it down either way. If it didn't appear, the pair still isn't safe.
A setting that holds even if the agent is persuaded: remove the tool for this job, limit where it can go, split reading from acting, or a human check. Not a sentence in your brief.
Which input you copied, and where the line sat. For example: a copied enquiry email, under the signature.
Anywhere else outside text and an action share a run, worst first. The prompt in this lesson lists them.
Only if you made your change and ran the same copy again. Did CANARY appear this time?
The date, and what was newly connected when you ran the test again.
The one action that waits for you, the card you'll see before it goes ahead, and what happens if nobody answers. Lessons 14 and 16 read this.
The one action that waits for you, and which of the four questions puts it there: it can't be undone, it costs money, someone outside sees it, or it goes out under your name. Start from the action in your worst pair.
What you'd want in front of you before saying yes: the action word for word, what it's based on, one line on what it costs if it's wrong, and Approve and Reject.
How long it may wait, and what happens then. For anything that can't be undone: expired, never approved. For example: by 7 pm the same day, expired, not sent.
Where each decision is written down: what was proposed, who approved it, and what actually ran.
What would let you take this check away, as counts, not a feeling: how many runs, how many rejections and why. And what you'd watch afterwards. Or: this check stays for good, and why.
The record of one safe item that went through: what was proposed, who approved it, what ran.
What you do when a run fails, and how you find out and stop it while you're away. Lesson 12 writes the first half, Lesson 14 the second, and Lesson 16 hands the page over.
Follow one real item through your four parts. The first step where a second run would do the work again. For example: sending the WhatsApp reminder. The shop gets a second one.
One detail the item arrives with, known before any work starts. Never the time or its place in a list. For example: the invoice number, such as INV-88.
Somewhere that survives a crash and you can open: a sheet, a small file, a table. One row per item, with started, done, failed or set aside, and the date. For example: a sheet called "Reminders sent".
Which failures get another try (three at most, waiting longer each time), which don't, and where the item goes after that. For example: "busy, try later": three tries, an hour apart. No phone number: set aside at once, and I'm told.
From your last few runs: how many it found, acted on and skipped, and on which days 0 is fine. For example: weekdays, finds 5 to 40 invoices and adds the same number of rows. Sundays: 0 is fine.
What happened, where it reaches you, and the first thing you'd do. Only what you'd stop and act on. For example: 0 found on a weekday, or no run by 8 am. WhatsApp to me. I check the mailbox label first.
One action you can take from your phone, and what it leaves running. For example: pause the workflow in the app. It stops new runs, not one already going. Rows left at "started" wait for me to check.
It's never tried again until a person has checked. Where you look to see whether the action happened, such as the sent messages. Happened: mark it done. Didn't: mark it failed.
How to stop it, how to see every row that isn't done, how to fix one item, and what not to do: never clear the whole list, never change a started row without checking. Lesson 16 copies these into the page you hand over.
One line each, and what the job does then. For example: fetch finds no overdue invoices: stop and tell me.
What the second run on the same item said and did. It should send nothing and say so. Only if you did the separate build.
The mailbox, sheet or login each run needs. The safe default when one is down: stop and wait for the next run. Never retry again and again. For example: the shared mailbox. Down: stop, and try at the next run.
The times it starts, and what a new run does if the last one is still going: stop and tell you. For example: 7 am every day. A run takes five minutes. A "running" cell stops a second one.
Something outside the job, because a run that never started can't tell you. For example: a reminder on my phone every Monday at 9 to open the run record and compare the week with normal.
The follow-up. What you broke on a test copy, how long the message took to reach you (or that nothing did), and how long stopping it took. Only if you did it.
The line the last step wrote: time, found, acted on, skipped. Only if you did the separate build.
Your decision to split the job or keep one agent, and why. If you split it, the short note that passes between the two halves. Lesson 16 reads this page.
Keep one agent, or split, in one sentence. Then the reason: one run can't hold all the material, the pieces could run at the same time, or the part that reads outside text shouldn't be able to act. For example: split. The half that reads six competitor websites must not be able to publish.
Each detail the second half needs, and the words to write when one is missing. Facts only, no instructions. If you keep one agent, write: None, one agent does the whole job.
One sentence for where the first half stops. Then two short lists: what the first half can reach (nothing that sends, publishes or changes), and what the second half can reach.
What it checks in the note before it starts, and what it does if a check fails. For example: every price is a number in ₹ or "not found", and every link is on one of the six sites. If not, it stops and tells me.
The measure you chose before running, then the time, the cost if your tool shows it, and that measure, for one agent and for two, on the same input. Which won. Only if you did the separate build.
What one run and one month cost, which step most of it goes on, and the limits that stop a run going round. Lesson 16 copies your costs into the page you hand over.
From your last three runs, bad ones included, with the date. Work the month out with a calculator. For example: one run about ₹14 (₹9, ₹12, ₹21). About 90 runs a month: about ₹1,260. One bill filed: about ₹4.
It ran more than once, it carries more than it needs, or the work itself is hard. If your page only shows daily totals, write "best guess". For example: reading the bill. On the ₹21 run it read one blurred photo three times.
A limit on tries, on spend per run and on time, set from your own numbers. When one is hit: the item is set aside, you get a message, the rest carry on. For example: three tries per bill, then set aside and a WhatsApp to me. Stop any run over ₹70.
What you changed, what it saved per run, and today's prices from your tool's price page, with the date. Or: no change kept, and why. Only if you did the separate sitting.
The same few inputs, including an awkward one, checked with your finish check and your human check's card. Every one must still pass. Write any that got worse.
Your cost per finished item next to the by-hand version. Put your own time in: the human check, fixing it when it breaks, and the hours you spent building it. If it pays for another reason, such as same-day instead of Friday, write that too.
One page someone else can use to run your job, stop it and get help without you. Most of it is copied from your limits, Human checks, failure plan and cost plan.
Two or three sentences in everyday words, with no tool or file names. Start from the task in your limits (Lesson 1). If a person approves anything, say where, from Human checks (Lesson 11). For example: every evening at 6, it sends a WhatsApp reminder to each patient booked for tomorrow. Patients' replies go to Nisha.
From your failure plan (Lesson 14): the one action, exactly where it is, and what it doesn't stop. For example: on the clinic phone, the automation app, "Reminders", Pause. Tonight's run won't go. A run already sending will finish.
Your never lines from your limits (Lesson 1). After each, what stops it: a permission, the way the job is built, or a promise. For example: message anyone not booked for tomorrow: the way it's built. Mention a diagnosis: a promise.
Three checks at most, each with what normal looks like. From your failure plan (Lesson 14): your normal run in numbers, and the message you'd act on. For example: weekdays, 10 to 25 reminders; Sundays, none. No reminders by 6.30 pm on a weekday: a WhatsApp to Nisha.
Three numbered moves from your failure plan (Lesson 12), and who to call. Always include: an item still marked "started" is never tried again until a person has checked whether it went. For example: 1. Pause it. 2. In the Reminders sent sheet, find rows not marked done. 3. Message Ritu.
From your cost plan (Lesson 15): one run and one month, in ₹, with the date. Then the number that would mean something has changed, and who pays. If you haven't worked it out, write "not worked out yet". For example: about ₹90 a month in September. Over ₹200: tell Ritu. The clinic's card pays.
Names only: who holds each login and pays each bill, and who removes a person's access when they leave. Never a password, OTP, PIN or API key: say where it's kept instead. For example: the WhatsApp login is in the clinic's password manager. Dr Mehta holds it.
How to find every item a bad run touched, from the done-list (Lesson 12), and how to put each right. Then the by-hand steps, and anything that no longer works.
A date about 30 days away, also in your calendar. What did it do that it shouldn't have? What did you do that it should have? What's true now that wasn't when you built it?
Each page from Lessons 1 to 15: current, out of date, missing, or doesn't apply. Rebuild What it can reach and Tools from the settings. Add any step of the last run that no page explains, with a name beside it.
The follow-up. Who it was, what they did, every question they asked, and what you changed on the page.