Your work
What you wrote in the practice worksheets. Saved in this browser only, so export a backup before you switch devices. Signing in syncs which lessons you have completed, not this text.
Working With AI
Your tasksNot started
One task from your week, and how you'll check it. Your first request comes from these answers.
About youNot started
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".
More details (optional)
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.
Your requestsNot started
One request you really send, rewritten in five short lines. Lesson 4 starts 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.
More details (optional)
Only if you have one. Keep its vagueness; it shows what changed.
Your background notesNot started
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.
More details (optional)
Names, where things stand, decisions already made. Put a date at the top.
Your checksNot started
The checks you'll actually run on AI's output, written so someone else could run them too.
What to look at, what counts as passing, when to stop. Written so a stranger could run it.
What to look at, what counts as passing, when to stop. Written so a stranger could run it.
What to look at, what counts as passing, when to stop. Written so a stranger could run it.
Saved promptNot started
One request worth reusing, written down with the inputs that change and the tests it has passed.
Fixed part, [slots] in brackets, a rules block, an output format.
v1, today's date: first working version. Add a line for every change, with the failure that caused it.
Typical, short or messy, edge case. Paste them or link to them.
Your memory notesNot started
What AI should remember about you between sessions, and the rule for when it gets deleted.
What comes out, when you check (a date, not a frequency), how you remove it.
One row per topic note: file, what's in it, load when.
Its name, then one fact per line.
Its name, then one fact per line.
Your standardsNot started
The rules your own work already follows, pulled out so you can check AI's work against them.
One rule per line, each with a short quoted example under it. No adjectives on their own.
What must be present for the work to be acceptable at all. Examples under each.
Six to ten exact strings and structural habits, quoted.
Which were flattery, which were misreadings, what it missed.
Extraction workflowNot started
The exact rules for pulling data out of documents, so two people get the same answer.
One line per field: name | type | what counts | what does not count.
Currency, date format, number grouping, document languages, which identifiers stay as text.
When NOT STATED, UNCLEAR and ILLEGIBLE each apply.
A spreadsheet, code, or you. Never inside a sentence.
Three random rows read back to the source, every time.
Research workflowNot started
Your rule for which sources count and when a claim is verified, not just found.
Claim, source, date, position, status.
How old a source can be and still count, and what to do with undated pages.
Verified, partly supported, unverified, contradicted.
Which rows must be opened before the work is used: the ones your conclusion rests on.
What it may do aloneNot started
The line between what AI may do without asking and what needs your say-so first.
Conversation, fixed automation, or supervised agent.
Read, plan, act, check, report, with a real noun in each.
The named step it sits before, and which test it fails: irreversible, spends money, carries your name.
Mark each item PERMISSION (enforced outside the chat) or INSTRUCTION (a line it could read past).
Your reviewNot started
A short look back at what worked, what broke, and the one thing you're changing today.
Working With AI Agents
Your limitsNot started
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.
More details (optional)
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.
Where it runsNot started
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.
More details (optional)
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.
What it can reachNot started
Exactly what it can read, write, and reach on the network, so nothing surprises you later.
Write the working version you would reuse or hand to someone else.
Standing briefNot started
The standing instructions it reads on every run: what this is, how work gets done here, what's off-limits.
Write the working version you would reuse or hand to someone else.
Context planNot started
What loads on every run, what gets fetched, and what you're deciding to leave out.
What loads on every run, and its rough size.
What the agent goes and gets, and what happens if it does not arrive.
Its own steps and results, which only grow.
What does not enter, and what you accept losing.
Moves you reuseNot started
One routine worth naming and reusing, with the steps a stranger could follow.
Findable in three seconds, six weeks from now. A verb and a noun from your work.
The situation this move is for, not a wish.
What it needs, where each comes from, what happens if one is missing.
The sequence a capable stranger could follow. Mark judgement steps.
Conditions on the output a stranger could apply.
Typical, awkward, should-refuse: what happened on each.
ToolsNot started
What each tool does, what it costs, and the one thing you'd never use it for.
One entry per tool: purpose, takes/returns, writes?, speed/cost.
On at least two tools, naming a plausible wrong use.
The same three requests, before and after, and which tool was picked first each time.
WorkflowsNot started
One automated run end to end: what starts it, what it fetches, and where it touches the real world.
What starts a run, and the real folder, sheet or mailbox it names.
What this run needs, and what happens if it comes back empty.
The rules, and the one AI node: what it decides and its allowed set of answers.
The one place the workflow changes the world.
Outside-text drillNot started
A planted test to check whether outside text can make your agent do something it shouldn't.
Every untrusted source, every sink, a line for each pair in one run. Mark the worst.
The harmless marker, and where you planted it.
Whether the marker appeared, and whether the run's own summary would have shown it.
What you changed, and where it is enforced, not in the brief.
The second run, and whether it is a property or a promise.
Human checksNot started
The steps that stop and wait for you, and what happens if nobody answers in time.
The step, and the property that forced it: irreversible, expensive, public, or carries your name.
The action in full, the evidence, the cost of being wrong, and the reject control.
What happens when nobody answers.
Proposed, approved by, what happened.
The evidence that would let you retire this gate, and what you would watch afterwards.
Failure planNot started
What you do when a run fails partway through, so nothing runs twice or gets lost.
From the item's own data, computable before the work starts. Why it is stable.
Where it lives, its columns, and the four states: claimed, done, failed, poison.
Which failures retry, capped at three with growing waits. What happens past the cap.
A missing response is an unknown outcome, not a failure.
How to stop it, how to see the state, how to fix one item, what not to do.
Items found, acted on, skipped, queued for the gate, from real runs.
Condition, channel, and the action you would take, one row per alert.
The one action that stops it. What it does not stop. Timed on a live system.
What it needs to reach, and what it does when one is down: stop and wait, do the part it can, or fail loudly.
How the job splitsNot started
Why this job is split across more than one agent, and what happens when one gets a wrong answer.
Context pressure, parallel time, or separation of privilege. Or: it isn't.
Fan-out and gather, pipeline handoff, or supervisor with workers.
Named fields, which are required, the value for unknown.
Time, cost and a counted quality measure, for both, on the same input. The winner.
What happens when one worker returns a confident, wrong answer.
Cost planNot started
What this actually costs to run, and the one change that made it cheaper without breaking it.
Per run, per month, per unit of work done, each with its date and run count.
Units, share of run, median time, times-run, sorted by share.
Before and after, or 'no change improves cost safely'.
All ten inputs, against the same gate, including anything that got worse.
What each does when it fires.
HandoverNot started
Everything the next person needs to keep this running: how to stop it, who to call, how to undo a bad run.
One paragraph, in business language. No node, model or path names.
The exact sequence, near the top.
From Your limits, each line marked permission, structure, or promise.
Not the secrets themselves, only who holds and transfers them.
Three things, each with what normal looks like.
The first three moves from the runbook, and who to call.
How you undo a bad run, and the by-hand procedure that still has to work.