Multi-step workflows

Lessons 2 and 3 were about features where the app runs the loop for you. This one is about the loop you run yourself, which is still the most common way small businesses get real work out of AI. A workflow here means a job you have split into steps, where each step is one AI request and the output of one step feeds the next. No special feature required. Just a plan and a bit of discipline.

The four-step shape

Almost every job that is worth more than one prompt has the same skeleton. Learn it once and you will see it everywhere:

• Gather. Collect the real inputs before anything is written. Prices, the actual services you offer, last year's version, the customer's email, the competitor's page. This step is mostly you, sometimes a research run.
• Draft. Turn the inputs into a first version. This is what most people think of as "using AI," and it is only a quarter of the job.
• Check. Compare the draft against the inputs and the goal. Did a price drift? Did it invent a service? Is the tone off? Ideally this happens in a different chat, for a reason the hard part below explains.
• Deliver. Format, export, send, print. The boring step that makes the work real.

Notice this is the plan, act, check loop from lesson 1, unrolled. The agent features automate it. Here you drive, which means you get to put your own eyes on the most important boundary: between Check and Deliver.

Breaking a job into steps

The skill is noticing when a request is secretly several requests. "Write me a quote for the Hendersons' backyard" sounds like one thing. It is at least four: pull the measurements and the price list, draft the line items, check the math and the wording, and format it on the company template. Ask for all four in one breath and the AI will do the middle two and guess at the rest.

Three tests for whether something should be its own step:

• Does it need different inputs? Drafting needs the price list. Checking needs the draft plus the price list. Different inputs, different step.
• Would you want to look at the result before continuing? If a wrong answer here would waste the next ten minutes, stop and look. That is a step boundary.
• Is it a different kind of thinking? Gathering is searching. Drafting is writing. Checking is comparing. Models do each better when the instruction is about one of them.

Write the steps down before you start, even on a sticky note. Two minutes of planning saves the most common failure, which is a forty-message chat that drifted and now nobody, including the AI, remembers what the goal was.

Handoffs between chats

The move that feels wrong to beginners and turns out to matter most: start a fresh chat for a new step. Copy the output of step one, open a new conversation, and paste it in with the instruction for step two. Three reasons:

• A clean slate. A long chat drags every earlier wrong turn along with it. Recall from AI Basics lesson 6 that everything in the conversation is context the model is reading. Old drafts, abandoned ideas, and your corrections all pull on the next answer.
• Fresh eyes for checking. A chat that wrote a draft will defend it. A new chat that has only the draft and the facts will find the problems.
• The right tool per step. Gather in a research run. Draft in your everyday chat with your custom instructions loaded. Check in a chat with a deliberately cold, picky system prompt. Deliver in your document app.

A handoff is just a paste with a label. "Here is the draft from step 2 and the price list it was built from. Your only job is step 3: find every number in the draft that does not match the list." Label the step and the inputs every time. It keeps you honest too.

What to carry across a handoff, and what to leave behind: carry the output of the last step, the original inputs it depends on, and one sentence of goal. Leave behind the conversation that produced it, your corrections, and the three drafts you rejected. The new chat does not need the history. It needs the facts and the job. If you find yourself pasting a whole conversation into a new chat, you are rebuilding the problem you were trying to escape.

A tip for busy weeks: keep a plain text file per workflow with the step instructions already written. Gather the inputs, then paste step 2, paste step 3, done. The instructions do not change from job to job. Only the inputs do. That file is the beginning of the automation the next module builds.

One long prompt or a chain?

Chaining steps takes more of your time. Sometimes one well-built prompt is the right call. A quick guide:

• One prompt when all the inputs fit in the message, you can judge the whole result at a glance, and a miss costs you a retry, not a customer. Social posts, review replies, rewording, summaries of text you provide.
• A chain when the inputs come from different places, the output has numbers or commitments in it, or you would not want to send it without a second look. Quotes, proposals, price sheets, anything with a dollar sign or a date.
• An agent feature when the chain is one you run every week. Lesson 3's tools and the next module on agents are about turning a chain you have proven by hand into something that runs on its own. Prove it by hand first.

There is a middle path: one prompt that asks for the steps in order and shows its work. "First list the facts you are using, then write the draft, then list any claim in the draft not covered by the facts." It is weaker than a real chain, because the checker is still the drafter, but it is a big step up from "write me a quote" and it takes no extra time.

Try this yourself

Run one real job as a three-chat chain this week. A customer quote email is a good first one. Step 1 is you gathering: the job details and your price list, in a text file. Then:

Chat A (draft):
Step 2 of 4. Using only the job details and price list below, draft a quote email for the customer. Plain, friendly, under 200 words, line items in a simple list with a total. Do not add services that are not in the list.
[paste job details and price list]

Chat B (check), a brand-new conversation:
Step 3 of 4. Below is a quote email and the price list it was built from. Your only job: list every number, service, or promise in the email that does not match the price list or the job details. If everything matches, say so in one line.
[paste the email from Chat A, then the same job details and price list]

You (deliver):
Fix anything Chat B found. Paste into your email app. Send.

Time it. Most people find the chain takes about five minutes longer than a single prompt, and catches something on the first try. Once you have run it three times, you have a workflow worth automating, which is where the Agents module picks up.