Chapter 1  ·  Foundation  ·  Exercise 1 of 7  ·  5 min Ch.1 · Foundation · Ex.1 · 5 min

Your two projects. Create two projects.

Build your foundation: two projects, one for exploring, one for executing. Build your working foundation before prompting. Two projects: one for exploration, one for structured output. All exercises build on this.

Learning goal Goal Understand the Human/Machine balance by building both sides of it. Build both working modes as concrete projects.
You'll build Two projects in your AI tool — one set up for human-mode work (exploring, thinking out loud, open questions) and one for machine-mode work (structured tasks, repeatable output, clear constraints). You'll write the instructions for both. Two projects. Collab: open-ended, exploratory. Dispatch: structured, repeatable. Instructions for each.
Outcome A working setup you'll actually use — not a demo. By the end you'll have real persistent instructions in place for both projects and you'll have felt the difference between writing each one. Working persistent instructions for both projects. Usable immediately.
Watch for The moment you realise you're writing very differently for each project. That gap — between the instructions you'd give a collaborator and the instructions you'd give a machine — is the whole course in miniature. Notice: the instructions you write for each project will feel completely different. That gap is the concept.

What you're setting up Context

Persistent context on Claude

What most people do wrong first Common failure patterns

"Help me with work stuff."
Why No role, no context, no format. The model invents all three and usually gets them wrong. You get a response calibrated for nobody. No role · no context · no format. Model guesses all three. Output is calibrated for nobody.
"Hi! I was hoping you could help me when I need it. I'd love it if you were really helpful and friendly :)"
Why This tells the model nothing useful. You're spending tokens on politeness. The model reads it and learns you're friendly, that's not context. It changes nothing about the output. Zero signal. Pleasantries consume tokens and calibrate nothing. Model learns: you're polite. Irrelevant to output.
"I need you to help me with everything I work on. I'm a marketing manager but also sometimes I do strategy and occasionally I help with product decisions. I want responses that are professional but also friendly, not too long but detailed enough to be useful and please always consider..." [continues for 300 more words]
Why Length is not precision. A 300-word blob of stream-of-consciousness can't be applied consistently — the model will weight things arbitrarily. Structure and specificity beat volume every time. Length ≠ precision. Unstructured instructions are applied inconsistently. Structure + specificity > volume.
Part 1 — Collab project Part 1 — Collab
Collab Exploring and thinking out loud

Paste this into a new conversation to build your Collab instructions Prompt — paste into new conversation

Part 2 — Dispatch project Part 2 — Dispatch
Dispatch Structured tasks and repeatable output

Paste this into a new conversation to build your Dispatch instructions Prompt — paste into new conversation

Note on Tines: Workbench has one global instruction set — you can't create two separate projects the way other platforms can. The workaround: set your Dispatch instructions globally (most automation work is structured and benefits from precision) and keep your Collab instructions as a short block of text you paste at the start of explorative conversations. Label the paste clearly: // Collab — exploratory, push back, ask questions.
What a good response looks like Expected output
When you paste either prompt, the model should ask you one focused question, then produce structured instructions based on your answer. If it skips the question and writes generic boilerplate or asks you five questions at once, your setup isn't working yet. Good instructions are specific, scannable and under 15 lines. Model should: ask 1 question → receive answer → produce structured instructions (<15 lines). Red flags: generic boilerplate · 5 questions at once · no questions at all.
Now evaluate what came back Evaluate output

Before you move on, read both sets of instructions the model produced — one for each project. Then answer these questions. You don't need to write them down, but you do need to actually think through them. Read both outputs. Answer each question before continuing.

  1. Which set of instructions is more specific? Could a stranger read them and know exactly how to work with you? Which output is more specific? Would a stranger know how to work from it?
  2. Did the Collab instructions describe a working relationship — tone, collaboration style, when to ask questions? Did the Dispatch instructions describe a task spec — format, length, what to skip? Collab: relationship or spec? Dispatch: spec or relationship? Each should be clearly one.
  3. Is there anything in either set that wouldn't change the output if you removed it? If so, cut it now. Every line should earn its place. Apply the load-bearing test: remove any line that wouldn't change output if absent.
  4. Which project are you more likely to actually use? That's your primary project — you'll work from it in the exercises from Chapter 2 onwards. Which project is your primary? You'll use it from Exercise 3 onwards.

In Exercise 2 you'll run the same real task in both projects and see how the same opening lands differently depending on which one you're in. That contrast is the last thing the two-project setup is here to teach. Exercise 2: same task, both projects, compare output. After that — one primary project.

← All exercises Exercise 2 →