1. Plan first, answer second

For any sizable task, add one line: 'Before writing, lay out your approach in three bullets and wait for my OK.' The plan costs twenty words and is checkable at a glance; the finished piece costs eight hundred and isn't, you'd have to read it all to discover the frame was wrong.

Even without the pause, 'first outline the approach, then execute it' improves multi-constraint work, because the model commits to a structure before it starts producing sentences. You are buying a checkpoint at the moment errors are cheapest to fix.

2. Decompose the big deliverables

Ask for a finished ten-section report in one prompt and you usually get uniform shallowness, effort spread thin across everything. Decompose instead: 'First, propose the outline.' Then: 'Draft section 1 only.' Then section 2, using what stands. Assemble at the end.

The gain isn't only depth. It's the seams. Reviewing between steps means corrections happen while they're small: you fix the outline in one line instead of fixing the finished report over an evening. This is Course 1's stacked verbs grown into a working method.

3. Self-checks and surfaced assumptions

After a draft, don't ask 'is this good?'. Ask for a check against named criteria: 'Check the draft: under 500 words? every claim sourced or labeled an assumption? pricing never mentioned? Answer pass/fail for each, then fix the failures.' Vague review invites a shrug; explicit criteria force a real pass.

The companion move: 'List the assumptions you made.' Every answer to an underspecified task contains silent guesses, audience, currency, timezone, what counts as done. Surfacing them turns invisible errors into a checklist you can correct in one message.

4. When step-by-step helps, and when it's noise

Reasoning requests pay on tasks with intermediate results: analysis, anything arithmetic-flavored, multi-constraint balancing, tricky edits where order matters. If a wrong middle step would sink the answer, ask to see the steps.

On simple lookups and mechanical conversions they just add cost: a slower, longer answer to a question that needed one line, and more text for you to wade through. Ceremony should match difficulty. A good prompt is not maximally rigorous everywhere, it is rigorous where the task can actually go wrong.