Chapter 1  ·  Foundation  ·  Exercise 2 of 7  ·  5 min Ch.1 · Foundation · Ex.2 · 5 min

The opening seed. The opening seed.

The first message sets everything. Get it right from turn one or spend the session correcting it. First message = context + intent + constraints + done criteria. Gets it right from turn one or you spend the session correcting drift.

Learning goal Goal Feel the difference one opening message makes and pick your primary project. Run the same task in both projects. Pick your primary.
You'll do Pick a real task from your actual work this week. Write an opening seed for it in your Collab project, then write one for the same task in your Dispatch project. See what comes back differently. One real task. Two opening seeds — one per project. Compare output.
Outcome A felt understanding of how the same task lands differently depending on how you open it. After this exercise you'll pick your primary project — the one you'll use from Exercise 3 onwards. Your second project stays open for reference. Visceral understanding of opening seed impact. Primary project selected. Two-project phase ends here.
Watch for Which response asked you a question before producing output? Which one just started? Did one feel more useful for this specific task? The answer tells you which mode this task belongs in and probably tells you something about how you currently default. Which project asked before executing? Which just ran? Which output is more useful? That's your mode signal for this task type.

What makes an opening seed Anatomy

Four elements — always in this order

1. Role. Who is speaking and why does it matter for this task? One sentence. Not a title — a calibration. "I'm a product manager preparing a brief for an engineering team" tells the model what level of detail, what vocabulary and what kind of output will be useful. Without it, the model guesses.

1. Role. Who is speaking · why it matters for this task. One sentence. Calibrates vocabulary, depth, output type.

2. Task. Verb first, then object, then scope. "Write a one-page brief" not "I need some help with a brief." The verb sets the blast radius — write, draft, summarise, analyse and refactor all imply different scopes. Choose deliberately.

2. Task. Verb + object + scope. Verb sets blast radius — choose deliberately. Not "I need help with X."

3. Context. Only what the model needs to do this task well. Apply the load-bearing test: if removing a piece of context wouldn't change the output, cut it. Background that's interesting but not task-relevant is noise and noise degrades output.

3. Context. Minimum that makes the task possible at target quality. Load-bearing test: if removing it wouldn't change output, cut it.

4. Done criteria. What does the output look like when it's finished? Format, length, audience, tone and the finish line. Without a finish line, iteration is unbounded and you have no basis for knowing you've arrived.

4. Done criteria. Format · length · audience · tone · finish line. No finish line = unbounded iteration.

Your task for this exercise Task

Pick something real from your work this week — a document to write, a decision to think through, a piece of research, an email that needs care. It should be something you actually need to do, not a made-up example. The exercise only works with real stakes. Use a real work task this week. Document · decision · research · communication. Must be real — exercise requires genuine stakes.

Note on Tines: You only have one Workbench instruction set. For this exercise: run your Collab seed by pasting the // Collab override block at the top of a new conversation before your seed, then run your Dispatch seed in a fresh conversation without it. The contrast will still be visible.
Part 1 — Collab project
Collab

Open a new conversation in your Collab project and paste this, then fill in the brackets with your actual task New conversation · Collab project · fill brackets with real task

Part 2 — Dispatch project
Dispatch

Open a new conversation in your Dispatch project — same task, different opening New conversation · Dispatch project · same task

Compare what came back Evaluate

This is your calibration in action Calibration check

  1. Did one response ask you a question before producing output? Did the other just start? What does that tell you about how each project is calibrated? Which asked before executing? Which ran? What does that signal about calibration?
  2. Which output is more useful for this specific task — not in general, but for this task, right now? Does the task type explain which mode fits it better? Which output is more useful for this task specifically? Does task type predict mode fit?
  3. Where on the dial between the two openings would you actually want to be? Full Collab warmth? Full Dispatch precision? Somewhere between — context-heavy but instruction-clear? Where on the dial? Full Human · full Machine · hybrid (context-rich + instruction-precise)?

From Exercise 3 onwards, you'll work from your primary project. Pick the one that produced more useful output for most of your work or the one whose instructions feel most like how you actually want to work. Your second project isn't deleted; it's a reference point. When a task clearly belongs in the other mode, you'll know and you'll know what to open. From Exercise 3: one primary project. Pick the one with more useful output for your typical work. Second project = reference, not deleted. You'll know when a task needs the other mode.

← Exercise 1 Exercise 3 — coming soon