How You Build With Agents

How You Build With an Agent

You do not need to get it right on the first try. You need a repeatable build loop.

The seven-step loop

  1. Describe the manual workflow.
  2. Identify constants vs. variables.
  3. Build one step.
  4. Test with real examples.
  5. Refine the instructions.
  6. Compose multiple steps.
  7. Name and reuse.

Step 1: Describe what you currently do

Start in plain language. Tell the agent:

  • What you do manually (the full process, start to finish).
  • Where inputs come from (email, a website, a spreadsheet, a conversation).
  • What a good outcome looks like (be specific — "a one-page brief" not "something helpful").
  • What usually goes wrong (what you spend time fixing).

You are not writing code. You are explaining your process to a capable new team member.

Step 2: Identify constants and variables

Ask the agent (or yourself): "If I did this for a different client / project / week / product, what would stay the same and what would change?"

The constants become the instructions. The variables become the inputs.

Step 3: Build one step first

Do not build the entire workflow at once. Pick one step that is frequent and low-risk.

Good first steps:

  • Summarize these meeting notes into five bullets with action items.
  • Extract deadlines and owners from this email thread.
  • Reformat this data into my standard weekly report template.
  • Flag items in this list that match these priority criteria.

Step 4: Test with real data

Use actual inputs from your work, not made-up examples. Real data exposes edge cases that hypotheticals miss.

Ask:

  • Is anything important missing from the output?
  • Is anything irrelevant being surfaced?
  • Is the format something I can act on immediately?
  • Would this have made a real decision easier this week?

Step 5: Refine the instructions

If quality is weak, do not abandon the approach. Tighten your instructions:

  • Make priority criteria more specific.
  • Add exclusion rules ("ignore anything related to…").
  • Specify the exact output format and length.
  • Add examples of good and bad output.
  • Clarify edge cases.

Better instructions beat better prompting tricks every time.

Step 6: Compose steps into a flow

Once one step is reliable, chain 2–3 together:

  1. Gather updates from three sources.
  2. Filter for urgency and my priority criteria.
  3. Summarize into a decision-ready brief.

Each step's output feeds the next step's input. Now you have a workflow.

Step 7: Name it and run it

Give it a name: Morning Briefing. Client Prep Pack. Weekly Risk Scan.

Document when to run it, what inputs it needs, and where the output goes.

You are no longer doing ad hoc tasks. You are operating a system.