Give it the thinking, not just the request

Lesson 01.1

A prompt is a brief, not a spell

The goal is not to find a magical sentence. It is to transfer enough context for the model to work usefully—and enough criteria for you to judge its work.

For consequential tasks, define six things: the goal, context, evidence, constraints, expected output and what to do when information is missing. A good brief makes success and failure observable.

Start with the decision you need to make rather than the role you want the AI to pretend to have. The model can be remarkably capable and still lack your organisational context.

• Goal — the decision, answer or artifact you need
• Context — audience, background and what has already happened
• Evidence — sources it may use, and facts it must not invent
• Constraints — boundaries, standards and exclusions
• Output — format, depth and definition of a good result
• Uncertainty — what to flag, verify, ask or leave unresolved

Lesson 01.2

Dictation is a thinking tool

You do not need to finish thinking before talking to AI. Sometimes the most useful input is the thinking you have not yet edited into a neat paragraph.

A text box encourages premature editing. You start typing, delete half the thought, replace the words, and quietly discard the contradictions or half-formed ideas. Those discarded fragments are often where the real problem lives.

Dictation changes the sequence: think aloud first, structure second, execute third. Speak the messy version: what you are trying to solve, what bothers you, what you have already tried, what you are unsure about and even the moments when you change your mind.

The objective is not a beautiful transcript. It is to capture more of your reasoning before self-editing strips it away. Let AI organise the material, but keep ownership of the meaning.

Dictation is especially useful for complex prompts, skill design, requirements gathering, root-cause analysis, article outlines and difficult decisions. It is less suitable for sharing private or regulated information with a tool that has not been approved for it.

• 1. Speak freely for a few minutes. Do not correct every sentence.
• 2. Include doubts, contradictions, priorities and examples—not only your final conclusion.
• 3. Ask AI to organise what you said into goals, known facts, assumptions, open questions and proposed next steps.
• 4. Require it to mark gaps rather than silently supplying missing information.
• 5. Review the structure. Correct what it misunderstood and restore anything important it dropped.
• 6. Only then ask for the deliverable: a prompt, skill specification, project plan or decision memo.

I'm going to think aloud. The following may be messy, repetitive, contradictory or unfinished.

Your first job is NOT to solve the problem.

1. Capture my actual goal and the context.
2. Separate facts, assumptions, preferences, constraints and unanswered questions.
3. Identify contradictions or important trade-offs.
4. Preserve useful examples and ideas even if they sound unfinished.
5. Do not invent facts or quietly fill gaps.
6. Give me a concise structured brief and list what needs my confirmation.

After I review and correct the brief, we can use it to create the final output.

My thinking:
[Dictate here]

Lesson 01.3

Correct the difference, not just the output

“Not quite right” tells the model very little. Explain what failed and why. For example: “You treated an inference as a confirmed fact,” or “This structure makes readers wait too long to understand the decision.”

Show it a good example and a bad example, then explain the difference. Examples demonstrate tone, evidence, structure and exclusions more precisely than adjectives such as “professional” or “engaging.”

If the same correction keeps recurring, promote it from an ad-hoc chat comment into a reusable instruction, skill or validation check. A correction in one conversation does not train the underlying model or guarantee the behaviour will recur.