Apply.

Take out your Stage 1 leverage map. The top five tasks on it are your build list. Each one becomes a single documented workflow.

Why document instead of winging it in chat? Three reasons:

• Consistency. An attorney who improvises a demand-letter prompt gets a different letter every time. A documented chain gets the same structure, tone, and checks on every run.
• Speed. The thinking is done once. After that, you execute. Most of the time savings in this stage come from not re-deciding how to do the task.
• Proof. You cannot measure an improvised process. You can measure a workflow. Stage 4 needs that measurement.

Scope each workflow to one task with one output. "Handle client email" is not a workflow. "Draft the weekly status email to active clients" is. If a task on your map is too broad, split it and pick the highest-volume piece first.

Build one workflow at a time. Get the first one running before you start the second. Five half-built workflows save nothing.

Every workflow has the same skeleton. Five parts, in order:

1. Trigger. What starts the workflow. A date (first Monday of the month), an event (a grant notice posts, a client file arrives), or a request (the principal asks for the parent newsletter). If you can't name the trigger, you'll forget to run it.
2. Input prep. What you gather before the first prompt. Files, numbers, prior examples, constraints. This is also where confidentiality lives: decide now what never goes into a prompt, client names, student records, privileged material, anything covered by your professional rules.
3. Prompt chain. The 2-4 prompts that do the work, in sequence, with each output feeding the next.
4. Verification. The named checks you run on the output before it leaves your hands.
5. Output standard. What done looks like: format, length, quality bar, where it goes.

Example: a military spouse running a part-time resume service. Trigger: client intake form submitted. Input prep: client's old resume plus the target job posting, contact details stripped. Chain: extract job requirements, map client experience to them, draft the resume. Verification: confirm every date and title against the intake form. Output standard: one page, client's voice, delivered as a Word file within 48 hours.

One prompt rarely does a whole task well. Chains of two to four small prompts beat one giant prompt, because you can inspect and correct the work between steps.

Start from the prompt library for your profession on this site. It holds tested starting prompts for the common tasks in your field. Copy the closest match, then adapt it: your templates, your standards, your reader.

A typical chain has three jobs:

1. Extract or analyze. Pull the facts out of the source material. A nonprofit grant writer's step 1: extract the funder's priorities, eligibility rules, and word limits from the RFP.
2. Draft. Produce the working document. Step 2: draft the statement of need using those priorities plus the org's program data.
3. Tighten and format. Enforce length, tone, and structure. Step 3: cut the draft to 500 words and match the funder's required headings.

Write each prompt into the template, not into your memory. The test of a good chain: you could hand the page to a competent colleague and they could run it without asking you anything.

Keep the chain short. If a step's output is reliably good, leave it. If a step needs heavy editing every run, rewrite that prompt. Don't add more steps.

AI gets things wrong. It invents citations, mangles numbers, misreads dates, and states wrong facts with full confidence. The error rate is low enough to lull you and high enough to hurt you. So verification is a written step in every workflow, with named checks, not a general intention to "review the output."

Good checks are specific and fast:

• An attorney: every cited case gets opened and confirmed to exist and to say what the draft claims. No exceptions.
• A bookkeeper: every number in the client summary ties back to the source P&L.
• A teacher building lesson plans: every fact in the handout checked against the textbook or a primary source.
• A veteran in transition writing tailored cover letters: every claim about their own record matched to their actual dates and roles, because AI will happily upgrade your job title.

Two rules. First: never send output you didn't check. Your name is on it; the AI's isn't. Second: verification time counts. Log it in your time tracking. A workflow that drafts in 5 minutes but needs 40 minutes of checking may still be worth it, but only the honest math tells you.

A workflow on paper proves nothing. Run each one on real work for two weeks and log every run in the Before/After Time Log.

The sequence:

1. Baseline first. Time the task the old way at least twice before you switch. Real runs, not guesses. Without a baseline you have no proof later.
2. Run and log. Time every workflow run: prep, prompting, verification, rework. All of it.
3. Expect a slow start. Run 1 is often slower than your baseline. You're debugging prompts. Runs 3-5 are where the curve shows up.
4. Refine the weak step. After each run, note what broke and change one thing in the template. One change per run, so you know what worked.
5. Be willing to kill one. If a workflow saves nothing after two honest weeks, that's a finding, not a failure. Some tasks resist automation. Swap in the next task from your Stage 1 map.

When all five workflows have two weeks of logged runs, total the hours saved per month. That number, your personal, measured proof, is what you carry into Stage 4.