AI Tools Academy
Cross-tool mastery 0/8

Phase 6 · Cross-tool mastery

Prompt chaining across tools

Walkthrough · 12 minLast checked against the live product: 13 July 2026

30-second recall from earlier lessons
You anchor a Gem to your Fernway sales sheet, and it reports the South region as weakest. But the sheet has rows misspelled 'Sotuh'. What's the risk?
A Deep Research report comes back with clear sections, a confident conclusion and a list of citations. You need one of its figures for a client proposal. What should you do?

By the end, you'll be able to…

  • Break a large task into a chain where each step's output feeds the next
  • Run a chain across more than one tool, choosing each tool for what it's best at
  • Spot where a chain breaks and add a human check at the point that matters

Why it matters

Big tasks fail when you ask for everything in one giant prompt. The tool makes every decision at once, out of sight, and one wrong turn ruins the whole thing. Chaining breaks the job into stages and passes each answer to the next, so you steer as you go and catch mistakes before they compound. Done across tools, it lets you use each one for what it does best: research in one, structured rewriting in another, slides in a third.

What chaining actually is

Prompt chaining is a plain idea with a technical name: break a bigger job into smaller steps, where the answer from one step becomes the starting material for the next. The input is what you give the tool; the output is what it hands back; in a chain, one step's output is the next step's input. You already do this in ordinary work: to write a report you gather facts, pick the important ones, sketch a structure, then draft. Each stage builds on the last.

The reason to bother is control. A single enormous prompt asks the model to decide everything simultaneously, and if it drifts early you often can't see where. A chain gives you a checkpoint between every stage, a place to read, fix, and only then move on. That pause is not overhead; it is the whole point.

Chaining gets more powerful still when the steps don't all happen in one tool. Each tool you've learned has a genuine strength: Gemini and ChatGPT can run Deep Research across many web sources; Claude is strong at careful, structured rewriting and holds a long document steadily; Copilot and Gemini turn structured text into slides inside your Office or Google world. A cross-tool chain lets you use each for what it's best at instead of forcing one tool to do a job it's mediocre at.

A worked pipeline

Fernway's Priya needs a short board pack on the customer-feedback problem: a researched briefing, a clean one-page summary, and a handful of slides. Here is the same job as a three-tool chain, carrying the real Fernway material through it.

Step 1: Research (Gemini or ChatGPT, Deep Research). Start broad and let a research-capable tool gather sources. The output is a long, referenced report, far more than the board needs but rich raw material.

Step 1: gather the raw materialGemini
Produce a researched briefing on best practice for handling customer feedback and complaints in a small UK business: acknowledgement timescales, tracking methods, and what drives 'we never heard back' complaints. Cite each source. Aim for depth over brevity. I'll cut it down afterwards.

Why this works: Deep Research is the right tool for breadth: it reads many sources and cites them. Asking for sources by name is what makes the next tool's job, and your fact-checking, possible.

Before anything moves forward, you read this. Step 1 is where facts enter the chain, so it is where checking matters most: open a couple of the cited sources and confirm they say what the report claims.

Step 2: Structured rewrite (Claude). Now hand the vetted report, plus Fernway's own context, to a tool that rewrites carefully. The output is a tight one-pager shaped for a specific reader.

Step 2: turn research into a one-page summaryClaude
Below is a researched briefing on customer-feedback best practice, and Fernway's own project brief for fixing this. Rewrite them together into a one-page summary for a UK board: the problem in two sentences, three recommendations tied to Fernway's situation, and the two success measures from the brief. Plain UK English, no jargon. Don't add any statistic that isn't in the briefing. [paste Step 1 output] [paste the project brief]

Why this works: Claude holds the long report and the project brief together and rewrites to a fixed structure. Feeding it the real Fernway brief grounds the generic research in this company's actual situation.

Open in Claude ↗

Read the one-pager against the brief. If a recommendation drifted from Fernway's real scope, fix it here, before it becomes a slide.

Step 3: Slides (Copilot or Gemini). Finally, turn the approved one-pager into a slide outline in the tool that lives where you'll present from.

Step 3: one-pager into a slide outlineCopilot
Turn this one-page summary into a six-slide outline for a board update: a title slide, the problem, three recommendation slides (one each), and a closing slide with the two success measures and next step. Keep each slide to a heading and three or four short bullpoints. [paste Step 2 output]

Why this works: Copilot builds inside Microsoft 365 where the deck will live. Giving it finished, approved text means it only has to structure and format (the job it's actually good at) rather than invent content.

Notice the flow: breadth in step 1, judgement and shaping in step 2, presentation in step 3. No single tool does all three well, and you steered at every handover.

Where chains break

Chaining is powerful, not magic, and it has one real weakness plus a few practical failure points.

  • Errors compound. A small wrong fact in step 1 gets built on in step 2 and is baked in and hard to spot by step 3. This is the biggest risk, and it is why the fact-check belongs at step 1, the earliest point, where a fix is cheapest.
  • Context gets dropped at the handover. Each new tool only knows what you paste in. Forget to include the brief, and the rewrite invents context. Between tools especially, you must carry the material forward deliberately. Nothing is remembered across apps.
  • Formatting mangles on the copy-paste. Tables, headings and references can degrade moving between tools. Check the pasted input looks right before you run the next step.
  • Too many tiny steps. Eight steps for a small job is more work than one good prompt. Match the number of stages to the size of the task.

The fix for all of these is the same discipline: read each output before feeding it onward. A chain is only as honest as the human check between its links.

Try it now

Common mistakes

  • Chaining on autopilot. Copying each output straight into the next step unread defeats the purpose. The checking between steps is where the value lives.
  • Fact-checking at the end. By the final step a wrong fact is disguised as a polished bullet. Check where facts enter, usually step 1, not where they exit.
  • Assuming the next tool remembers. It doesn't. Between apps there is no shared memory; paste in every piece of context the next step needs.
  • Over-trusting the fluent final artefact. A slick deck at the end of a chain feels authoritative precisely because it's had three passes. But three passes polish the wording, not the truth. A confident error that survived every handover is more dangerous, not less, because it now looks finished. Read the final output as critically as the first draft.

Keeping current

The chaining method is durable (break the job up, check between links, use each tool for its strength) and will outlast any specific feature. What changes is what each tool can take as input and how well it hands off to the next. For the current state of the research and rewriting features you'll chain, check each tool's official updates, such as OpenAI's ChatGPT release notes and Anthropic's Claude release notes. Accurate as of 13 July 2026.