L09. Prompts, tokens and the limits of a conversation

You will refine a prompt and create project-brief.txt, a short brief for a new conversation. Allow about 45 minutes. Prepare one available chat, about-me.txt, project-list.txt and Notepad. The same example introduces tokens, context and an illustrative cost calculation.

In class

The teacher compares a vague request with a precise one, then starts a fresh conversation using a short brief. Together, check that the projects, constraints and next task survive without the old history.

What your assistant needs to know

“Make a good plan” does not say what to plan or how much time is available. Olena adds three projects, one first action each, and no payment or publication. The answer can now be checked against those conditions.

State the task, supply the information and explain what you will check in the answer. A few sentences often suffice.

• Open the conversation where you made the plan. Read the saved project-list.txt.
• Paste the plan after the prompt below and send it.

Refine this plan for someone beginning to learn.
Keep three projects and one action for each.
Each action takes up to 20 minutes and has a visible result.
Use up to 12 words per item. Do not add projects.
Plan:

• Check the number of items and project names. Choose understandable actions. “Write three tasks in Notepad” is more useful than “Explore the productivity concept”.

> [RESULT] You have three short actions you can start now: one for each project.

Why length matters

A model processes text in pieces called tokens. A token may be a word, part of a word or punctuation. Ukrainian and English phrases with the same meaning need not contain the same number of tokens. The exact division depends on the model; “one word equals one token” is not a universal rule. [F219]

Context is the material available when answering: instructions, supplied text, some conversation history and tool results. This working space is limited. In long conversations, a service may shorten or summarise older messages. Being able to scroll back to the beginning does not mean the model currently receives every line. [F220]

[[VIS:f5_prompt_context]]

Save a short brief

In Notepad, create Documents\AI-practice\project-brief.txt:

Author: Olena. Interests: books, walking, organising the home.
Projects: personal website; My Day; a 30–45-second video.
Status: initial notes exist. Nothing is published yet.
Next task: outline the personal page.
Condition: do not invent contacts or completed achievements.
Unknown: public contact details have not been chosen.

In a fresh conversation, paste the brief and ask: “Name the three projects and the next task.” Expect the personal website, My Day and the video, with outlining the page as the next task. Update the status after completing work on your own project.

> [FIX] The assistant returns an old name or forgets a constraint. Supply the current brief, state the correction and request a revision of the specific item. “We already discussed this” cannot restore missing information.

How tokens relate to cost

A chat follows its plan’s conditions. API costs may be calculated from input and output tokens and separate tool charges; this is a different billing route. [F203][F221]

For this arithmetic example, not a current service price, use $1 per million input tokens and $5 per million output tokens. A request with 3000 input and 1000 output tokens would cost:

3000 / 1 000 000 × $1 = $0.003
1000 / 1 000 000 × $5 = $0.005
Total: $0.008. One hundred such requests: $0.80.

Input can include earlier conversation material, so counting only the user’s latest sentence is insufficient.

Home practice

• Copy project-brief.txt to AI-practice\project-brief-home.txt. Add only one result you actually obtained to its status line, such as a saved introduction. Do not call the whole website finished when you only have its text.
• Paste this brief into a fresh conversation and ask, “What is done, and what is the next task?” Compare the answer with the copy. It should retain the new status and unknown information. Preserve the original file.

Self-test

• Do 100 words always equal 100 tokens?
• In a fresh conversation, the assistant does not know the website plan. What should you paste to continue?
• What do 100 requests at $0.008 each cost?

Words you will need

• Token: a piece of data processed by a model.
• Context: material available for the current answer.
• Input and output: information supplied to the model and the answer it produces.