13 min · Power user
Multi-Step Prompting and Prompt Chaining
Run a large job as a checked chain: questions, source notes you paste, an outline, then a draft, and execute only a plan you have critiqued.
Learning objectives
- Split an overloaded request into a chain with a human check at each handoff.
- Use iterative reasoning so a plan is proposed, critiqued, and only then executed.
- Decide what context to carry, what to drop, and when a fresh chat plus a written brief is cleaner.
01
Why this matters
You can already get a fluent answer. The next skill is a controlled answer on a job that is too large for one prompt. Prompt chaining — making the checked output of one prompt the controlled input of the next — gives you a handoff you can reject. Without that pause, a weak middle step hides inside a polished ending.
02
Concept
Prompt chaining — making the checked output of one prompt the controlled input of the next — keeps each handoff visible. A single overloaded prompt asks Gemini to invent the questions, judge the evidence, design the outline, and write the memo in one pass. If the questions are shallow, the prose can still sound confident. The chain gives you a place to stop. You read the artifact, correct it, and only then paste it forward with a rule about what must not change.
Use a four-step chain for a research-backed memo. Step one asks only for the questions a decision needs, ranked. Step two is not a model invention: you paste source notes you gathered, and the prompt extracts claims, limits, and gaps. Step three builds an outline that maps each section to one question and one source. Step four drafts only from that outline. If the draft reaches for a new fact, the chain has collapsed. Name the allowed input in the prompt.
Iterative reasoning — proposing a plan, critiquing the plan, then executing only the approved plan — stops you from executing a foggy job. First, ask for a plan with assumptions, inputs, and what would make you reject it. Second, ask for a critique that names weak logic and missing facts. Only the third prompt may produce the artifact, and it must follow the approved plan. If the critique flatters the plan, send it back. Approval is your act, not a line the model writes about itself.
Context management — choosing what the next prompt is allowed to see — is how a long job stays sharp. Carry the written brief, the approved plan, the source notes, and the open decisions. Drop rejected drafts and abandoned ideas. When the thread is long or contradictory, a fresh chat is cleaner than another correction. Long-form projects need that written brief outside the chat so a new thread can restart from the page. Paste the brief first, then only the artifact the next step needs.
03
Real-world example
A Saturday-hours memo in four checked steps
A clinic manager needed a board memo on Saturday opening hours and nearly asked for the research, the decision, and the prose in one prompt. She used the four-step chain instead. First she requested only ranked questions and deleted two the board would never vote on. Second she pasted six lines from the duty roster and the reception log, and asked for claims, limits, and gaps. The notes did not prove demand, so the gap stayed open. She refused a guessed number. Third she approved an outline of four headings, each tied to one question and one note. Fourth she asked for a draft that could use only that outline. The recommendation was labeled a staff proposal, not a fact the model had discovered. The board could see what was known.
04
Gemini demonstration
A realistic exchange. Open Gemini beside this page and try a version with your own facts.
You
I need a board memo on whether our clinic should add Saturday hours. Do the research, pick a side, and write it so they can vote this month.
Gemini
I will not research, decide, and draft in one pass. That is an overloaded prompt, and a fluent memo would hide a weak question. Prompt chaining — making the checked output of one prompt the controlled input of the next — starts with an artifact you can reject. These are the questions a Saturday-hours memo must answer, ranked for a board vote. What problem are the extra hours meant to solve. Which evidence would change the vote: wait time, staff cost, or demand. What do your notes already show, and what is missing. What would make the idea fail even if demand is real. I have not recommended anything. Correct this list, then paste source notes you actually hold. I will not invent figures, citations, or a staff view you did not write.
You
Keep questions one, two, and four. Drop nothing else. Here are my notes: Saturday reception logged 40 turnaways last month, two nurses asked not to add shifts, and we have no cost sheet yet. Extract claims, limits, and gaps only.
Gemini
Claims I can support from your notes: the reception log records 40 Saturday turnaways last month, and two nurses asked not to add shifts. Limits: turnaways are not the same thing as a full demand study, and a request from two nurses is not a staffing plan. Gaps: there is no cost sheet, no doctor view, and no comparison with weekday demand. I am not recommending Saturday hours. I am not filling the cost gap with a guess. Next, if you approve, I will outline four sections, each tied to one question and one of these notes, and I will mark the cost section as unknown. Say what to change before that outline starts. The draft waits until you accept the outline.
05
Prompt breakdown
The prompt
Using only the notes I paste below, extract claims, limits, and gaps. A claim must point to a note. Do not add numbers, causes, or a recommendation. End with the single gap that most changes the decision. Notes: [paste].
- It names the allowed input, so the model cannot wander back into a general essay about clinics.
- It defines a claim as something that points to a note, which makes invented support easy to spot.
- It forbids the recommendation, so this step cannot smuggle in the decision before the evidence is stable.
- It asks for one decision-changing gap, which tells you whether the chain is ready for an outline.
06
Common mistakes
Usually
One prompt asks for research, a recommendation, a board memo, and a slide outline.
Better
Four prompts, and you approve the questions, the note extract, and the outline before any draft.
Usually
The next prompt says continue from everything above and figure it out.
Better
Paste only the approved artifact and state what the next step must not change.
Usually
The model writes a plan and a finished memo in the same reply, then calls the plan approved.
Better
Stop after the critique, edit the plan yourself, and only then allow execution.
07
Practice exercise
Chain a memo you actually owe
Done when: You have a saved outline where every section names its question and its note, and the gaps are still visible.
08
Challenge
Approve the plan before the work
Take a decision you face this month, small enough for one sitting but real enough that a wrong assumption would matter. Run iterative reasoning in three prompts. First, ask for a plan with assumptions, inputs, and tests that would make you reject it. Second, ask for a critique with no draft and no praise, then edit the plan yourself and mark what you changed. Third, execute only that approved plan, and tell the model to stop if it needs a fact you did not provide. Keep the critique beside the final artifact so you can see what the pause changed.
Hint: If the critique only compliments the plan, add the line: start with defects, skip praise, and say none found if you cannot name a miss.
09
Key takeaways
- An overloaded prompt hides the step that failed inside fluent prose.
- A research memo chain is questions, pasted notes, outline, then draft.
- Iterative reasoning means you approve the plan before any execution.
- A written brief outside the chat lets a fresh thread restart cleanly.
10