Tasks. Frame and Run.
- set a task for a chat in four parts
- check an answer as deeply as a mistake would cost
- break a big task into steps and start a new chat in time
Time shrinks: what took a day now takes minutes
- Understand a topic: minutes instead of days
- Write a document: minutes instead of hours
- Gather and compare options: a couple of minutes instead of a day
The course goes step by step: how the model (the program that writes the chat's answers) works, how to set a task, how to check, how to run big tasks. One example all the way: a book exchange, the announcement and the event plan.
Not a wizard: a probability machine
- Text is cut into small pieces called tokens.
- Inside there are no words or meanings, only odds: which piece follows which.
- Each time it adds the most likely continuation, like phone autocomplete with trillions of examples.
It is not a reference book or a smart search. What you put in decides what comes out.
One question, different answers. It is not a fault
- A new chat is a blank sheet: it does not remember the last task and starts from zero.
- Memory (if on) keeps general facts about you, not the content of earlier chats.
- A different chat, file or wording gives a different result.
Context is a table. What is not on the table does not exist for the model.
You know the book exchange is on Saturday. If you did not write it, for the model there is no Saturday exchange.
One word, different meanings: keep your own glossary
| Word | You mean | The model may read it as |
|---|---|---|
| “Exchange” | a meeting where people swap books | a currency or data exchange |
| “Card” | a line about one book in a list | a bank card or a playing card |
| “Event” | the book exchange itself | an event in software |
Goal — Context — Format — Constraints
- Goal: what you want and why.
- Context: everything that must be on the table: who, when, files, from whose point of view.
- Format: what the result looks like: length, tone, structure.
- Constraints: what not to do and what is out of scope.
A weak request and a strong one: four parts
“Write an announcement”
You get a generic template that you will have to rewrite.
- Goalinvite 20 people to the book exchange
- Contextneighbours and colleagues; Saturday at 12:00; bring one to three books you are happy to give away
- Formatfive lines
- Constraintsno exclamation marks
Add one line: “Ask me clarifying questions” and the model will ask instead of guessing.
Why the model rushes and flatters
- It rushes: a fast, confident answer beats a question back.
- The cost: the answer looks fine but misses the request, and time goes on redoing it.
- It flatters: “great question” is not an independent verdict on your idea.
A short chat setup (text to paste)
A prompt is the text you give a chat. Paste it once into the chat settings section where you can say how to work with you. From then on it applies in every chat. The text stays in English; the chat will still answer in your language. Try it: give any task and the chat asks questions first.
Have it retell the task before it works
“Make a plan for the event”
The model silently writes a plan for the wrong event.
“Before you start, tell me how you understood the task and what you are going to do”
Model: “I understood: I write an announcement.” You: “No, I need a plan on a page.” Caught in ten seconds.
Ten seconds of checking save half an hour of work off target.
Precise words, dialogue, options with reasons
| Move | What to say | What it gives |
|---|---|---|
| Precision | “Five lines, no exclamation marks” instead of “make it good” | “Good” is not a format: the model guesses |
| Dialogue | “Shorten it”, “add time and place”, “drop the second paragraph” | The first answer is a rough version; edits are normal |
| Options | “Give 2–3 options, pros and cons of each, and your recommendation” | You grasp the choice instead of guessing |
workit: our file of ready-made working rules for any chat
One plain text file: open it, copy it and paste it at the start of a chat, and the chat works by the rules
- A goal, and how you will know it is done
- Stays on goal: the chat does not drift
- Questions with options and reasons
- A map of the work: what is done and what is next
What not to paste and what to check in your chat settings
- Passwords for services and sites, text-message codes
- Card and document numbers
- Other people's personal data and private material you may not share
- Names become “Person A”, sums become “about N”, remove addresses and numbers
- Read the text yourself before you send it
- For sensitive things, use a private chat without history
What to check in the settings. In the data settings, find three things: whether the service uses your conversations to keep training its model, and how to switch that off; how long deleted chats are kept; whether temporary (incognito) chats without history are offered. Every service has its own rules and they change, so read them for the one you use.
Hallucination: confident invention
- The model does not feel the line between knowing and guessing: the tone is the same.
- The invention looks the way you expect: numbers, quotes, links.
Source: AP news report, 7 October 2025
The cost of a mistake sets how deep you check
| Depth | For | What it adds |
|---|---|---|
| Small | Everyday tasks: a sketch of the announcement | A sceptic's stance: assume mistakes exist, list them, fix them |
| Medium | Work documents: the event plan on a page | Re-read from the top, check against what was agreed, an honest “needs checking” mark |
| Full | Responsible material: the announcement everyone will see | A list of every claim, a mark of how important each problem found is (course 2 uses three: SERIOUS, MISSED, MINOR), up to three rounds of checking |
Did it read the whole file?
With a long file the model often reads the start and the end and skips the middle. Upload the file and paste this prompt. Instead of “are you sure?” ask: “What is this based on? Give a source” and “Name 2–3 weak points”.
Politeness does not work: ask for checkable actions
“Try hard, be careful, don't make mistakes”
The model says “sure” and writes just as confidently.
- Quote“Quote verbatim the lines your conclusion rests on”
- Source“A source for every figure; without one write ‘not confirmed’”
- Weak points“Name 2–3 places where this will break”
Ask for actions that cannot be faked, not for effort.
A big task: plan first
- Plan: “Do not do it yet. Write the plan: 5–7 steps, order, result.”
- Edit: you read the plan as the person who knows the situation.
- Steps: do, show, check, go on. A mistake in step one does not reach step seven.
- A short brief: goal; what is in; what is out; how we will check the result.
A research chat for fresh data
Fetch fresh data in a separate chat where internet search is turned on (switch it on in the chat settings or with a button by the input box; if there is none, this service cannot search) and bring only the short digest back to the main chat. Replace the [brackets] with your topic.
A tired chat is not healed, it is moved
Signs: it answers something else, blends old and new, returns to what you cancelled, repeats itself. Do not coax it. Ask for a hand-over summary (snapshot), open a new chat, paste it there.
Your own task: the whole cycle, step by step
- Task: take your own, a real one.
- Formula: Goal, Context, Format, Constraints, plus “Ask me clarifying questions”.
- Test = test: have it retell the task, correct it.
- Work: a big one by plan and in steps.
- Check: depth by the cost of a mistake; chat tired, move it.
All your tools
Ten prompts
Chat setup, test = test, options, checks, hand-over summary, research.
Download all the prompts as one archive (.zip: opens with a double click)The task formula
Goal — Context — Format — Constraints.
Test = test
Have it retell the task before it works.
Checks
Depth of checking follows the cost of a mistake.
Next, course 2: we build a skill (a saved set of instructions you can reuse) that writes the announcement from the same details every time: date, place, theme, what to bring.
One agent prepares, another checks; what matters goes to you
We plan the demo version for November 2026 and the working version, with licences on sale, from 1 January 2027.
Go to the SANKORD site and subscribe there to the project’s updates.