Training
Certification Leadership Frameworks Agentic for Business
Community Keynotes Retreat Blog Book A Consultation
Protocol 09 of 18 · Track, Building · What we ship

Workflow design

Designing repeatable agent-friendly workflows that turn intent into output. The shape of a good workflow, the failure modes of a bad one.

Protocol 08 Design systems and Claude design All 18 Protocol 10 Communications, SMS, email, Resend
Why this matters

The pain it
solves

Most people, given an AI, ask it to do whole jobs. Write me a marketing plan. Run my hiring process. The AI does a mediocre version of the whole job, the person is unimpressed, and the conclusion is AI is overhyped.

The real unlock is workflow design: breaking a job into steps, identifying which steps are AI-friendly, then orchestrating them. Same AI. Better output. Because the work was decomposed first.

The teaching

What this
actually is

The five-question shape of every workflow

Every workflow has the same shape. Done well, this turns vague jobs into shippable processes. Done badly, you get a Zapier flow nobody understands.

  • What is the work, in one sentence?
    If you cannot say it in one sentence, you have two workflows, not one. Split.
  • What are the steps, three to eight of them?
    Fewer than three and it does not need a workflow. More than eight and it is a multi-workflow process.
  • Who does each step today?
    Be honest. Person, agent, tool, or no one. The no-one steps are usually the bottleneck.
  • Which steps can AI do?
    Autonomously, as a draft for human review, or never. The honest answer per step matters. Wishful answers compound into bad workflows.
  • What is the workflow, with the AI/human routing made explicit?
    Write it down. The artefact is the workflow file. From then on, the team and the agents read the same document.

AI-friendly vs human-only

Most people, given an AI, ask it to do whole jobs. Write me a marketing plan. Run my hiring process. The AI does a mediocre version of the whole job, the person is unimpressed, and the conclusion is AI is overhyped.

The real unlock is workflow design: breaking a job into steps, identifying which steps are AI-friendly, then orchestrating them. Same AI. Better output. Because the work was decomposed first.

AI is excellent at: research, drafting, transformation, classification, summarisation. AI is poor at: novel judgment, taste calls, emotional intelligence, negotiation. Knowing the line saves hours.

The routing distinction that matters most

Step where AI scores and human spot-checks is different from step where AI suggests and human decides. The first is automated with sampling. The second is collaborative with handoff.

Naming the routing kills 80% of workflow friction. Most failed workflows mix the two: an AI step pretends to be autonomous when it actually needs a human in the loop, or a collaborative step pretends to be automated and produces garbage.

Try it yourself 10 minutes

Decompose one real workflow in ten minutes

Pick a task you do regularly that you wish were faster. Lead intake, weekly report, content review, code review, customer email triage. Open a markdown file called workflows/[task].md.

  1. Step 01
    Write the one-sentence purpose

    Top of the file: "This workflow takes X and produces Y." If you cannot, split into two workflows.

  2. Step 02
    List the steps, three to eight

    Each step is one verb plus what gets touched. "Read inbound email. Extract company name. Look up size from Crunchbase. Write a one-line summary. Post to Slack."

  3. Step 03
    Tag each step with who does it today

    [me], [agent: writer], [tool: zapier], [no one]. Be honest. The gaps become visible.

  4. Step 04
    Tag each step with AI-readiness

    [auto], [draft + review], [never]. "Look up size from Crunchbase" -> auto. "Write a one-line summary" -> draft + review. "Decide if we pursue" -> never.

  5. Step 05
    Build the first auto step

    Pick the highest-value [auto] step. Ask Claude to build it. Test it once. Wire it into your daily flow.

Outcome

One workflow decomposed and written down. At least one automated step shipped. The five-question drill is now a habit you can run on any task without prompting.

Official resources

Straight from
the source

What you walk out with

By the end of this
protocol

At the retreat

You learn it by
doing it

You take a task you do regularly, run it through the five questions in ten minutes, and walk out with a working workflow file plus one automated step.

Connects to

Other protocols this
compounds with

← Previous, Protocol 08 Next, Protocol 10 →