Delegate it: research & recommend
Hand it your first real job
The skill this whole course teaches is not 'using an AI' — it is delegation. And delegation starts at the bottom of the ladder, where the agent does the legwork and you keep every decision. Give it a job you would otherwise do yourself in a browser.
- 1 Pick a real question from your work that needs looking-into (a tool to choose, a topic to get up to speed on, a comparison to make).
- 2 Level 1 — inform. Ask Hermes:
Look intoRead what it returns.and just report what you find. Do not recommend anything yet. - 3 Level 2 — recommend. Now ask:
Give me three options for, each with pros and cons, and tell me which you would pick and why. - 4 Notice the shift: at level 1 you decide everything; at level 2 it commits to a recommendation you can accept or overrule.
- 5 Pressure-test it: ask
How would you check that this is actually correct?— a good delegate tells you how to verify its own output.
You got a useful report AND a reasoned recommendation — and you, not the agent, made the final call.
Why your instructions decide the quality
When delegation disappoints, it is almost always the brief, not the brain. The same task framed well or badly produces very different results.
A good delegation has four parts: ROLE (who the agent is and what it knows), TASK (the specific job), CONSTRAINTS (what it must or must not do, the boundaries), and OUTPUT (the shape of the answer you want). Levels 1–2 of the delegation ladder — 'look into it and report' and 'recommend with pros and cons' — are the safest rungs because you still make every decision; the agent cannot do anything irreversible. The one question that gates everything above level 2: can you verify the output is right? If you cannot check it, you cannot safely trust it, no matter how confident it sounds.
- ?Take your level-2 prompt. Which of the four parts (role, task, constraints, output) was weakest — and how did that show in the answer?
- ?For your question, how would you actually verify the recommendation — what would you look at to know it is right?
- ?Which is more dangerous: an agent that says 'I don't know', or one that gives a confident wrong answer? How do your constraints push it toward the safer one?
- ?When is 'just report' (level 1) the right choice even though the agent could recommend?
Rewrite a vague ask into a real brief
You will take a lazy one-liner and turn it into a delegation with all four parts — then watch the quality jump.
Write one strong delegation prompt (role + task + constraints + output) for a recurring kind of question in your work, run it, and compare it against the lazy version.
- 1 Write the lazy version first, e.g.
find me some papers on X. Run it. Note what is wrong with the result. - 2 Now write the full version: ROLE ('You are a research assistant in
'), TASK ('find and rank…'), CONSTRAINTS ('only peer-reviewed since 2023; say so if you are unsure'), OUTPUT ('a numbered list, each with one-line why-it-matters'). - 3 Run the full version and compare side by side.
- 4 Add one verification instruction to the prompt, e.g.
for each item, tell me how I could confirm it independently. - 5 Save the good prompt somewhere — you will reuse it (and in lesson 5, teach it to Hermes as a skill).
Your before/after prompts and outputs, plus one sentence on what the four-part brief fixed.
✎Delegation is a skill you already use with people. What makes a brief you would give a new colleague good — and how much of that transfers, unchanged, to briefing an agent?