Practice

The Workbench: run a project through the eight steps

One project, eight steps, your own words. The workbench gives you the structure and the prompts to paste into whichever tool you use. It never does the thinking for you — that would make it a Dependent's tool.

Planned for release phase 2. The page is complete as a draft; interactive parts run in the browser and do not yet save to an account.

In brief

Work through the steps in order. Your notes save automatically in this browser (nothing is sent anywhere). The prompts in dark boxes are for you to copy into your own AI tool at the moments the framework brings AI in — Step 2 Phase B, and Step 5. Step 1 has no prompt, on purpose.

Project name

Step 1 — Define the End State alone, no AI

Close every other tab. Write in paragraphs, not bullets. Answer these, in your own sentences.

Who, specifically, is this for? Name the humans, not the roles. What does each of them want from it? When it is done and they pick it up — what is the texture? Do they nod, or wince? What should it feel like? Constraints: budget, timeline, scope, quality bar. And the misstep you are trying to avoid.

Test: is this a deliverable list wearing an end state's clothes? If you have written "complete X with Y done", go back and write who it is for.

Step 2 — Map the Flow

Phase A — your list first

Every task from here to the end state, one per line, in sequence. Mark dependencies with → (e.g. "test reports → after migration produces reproducible results"). Do not look anything up yet.

Phase B — AI reviews your flow

Copy this into your conversational tool, with your end state and list pasted in.

I am planning a piece of work. Here is what success looks like, in my words: [PASTE YOUR END STATE] Here is my own list of every task I think it involves, in sequence, with the dependencies I have seen: [PASTE YOUR LIST] Do not rewrite my plan. Review it. Tell me specifically: what is missing; what dependencies or sequencing errors I have not seen; and what tasks someone with deep experience of this kind of work would have included that I haven't. For each gap, explain why it matters, in one or two sentences, so I can decide whether it applies at my scope.

Phase C — brainstorm each gap, then decide

For each gap the tool named: what did it say, what did you ask back, and what did you decide — add as-is, add in a smaller form, or decline with a reason?

Step 3 — Assign Tasks with the Decision Filter

Add each task from your flow. Answer the four questions. The filter suggests Human, AI or Decompose; you decide.

If the verdict is "decompose", split the task into smaller pieces and add each as its own row. Ambiguity always means the task is too coarse.

Step 4 — Select the tool category

For each AI task: text→text is a language model; image→text a vision model; text→image a generator; audio→text a voice tool; data→pattern analytics. Then: how much context, conversation or one-shot, how accurate, how often. The directory has current names per category.

Task → category → tool (and why)

I need to do this task: [DESCRIBE THE TASK]. The input is [text / image / audio / data] and the output should be [text / image / etc.]. It needs [a little / a lot of] context, it is [iterative / one-shot], accuracy matters [a lot / less than range], and I will do it [once / weekly]. Search for current AI tools suited to this, evaluate them against those criteria, and give me a shortlist of three with one line on why each fits.

Step 5 — Prepare and Direct

For each AI task, before you open the tool. Criteria first — once you see polished output, your standards drop.

Full context: what this is for, who it is for, the situation, the constraints, what to avoid. Expected output: format, length, depth, scope. Success criteria — written now, before any output. How will you know it is good?

Two-step prompting: paste the three boxes above in plain language, then add this line.

That is my briefing, in my own words. Given everything I just told you, write me an optimised prompt that captures all of it — the context, the expected output and the success criteria — in the format you would respond to best. Show me the prompt before you do the task.

Step 6 — Execute

What you did yourself; what you directed the tool to do; anything that surprised you.

Step 7 — Evaluate against your criteria

Go back to the criteria box in Step 5. Not "does it sound good" — does it meet each line?

For each AI output: which criteria it met, which it failed, what you changed (input, context, tool).

Step 8 — Integrate and stress-test the whole

Put everything together and look at it as one thing against the Step 1 end state. Then triage what you find.

Describe the problem you found.

Can you fix it without affecting anything else in the plan? Yes — it is local and self-contained No — fixing it changes other tasks, resources, sequence, or what I wanted

What you did about it, and — if structural — which step it traced to and which tasks you redid.

Step 1 — Define the End State alone, no AI

Step 2 — Map the Flow

Phase A — your list first

Phase B — AI reviews your flow

Phase C — brainstorm each gap, then decide

Step 3 — Assign Tasks with the Decision Filter

Step 4 — Select the tool category

Step 5 — Prepare and Direct

Step 6 — Execute

Step 7 — Evaluate against your criteria

Step 8 — Integrate and stress-test the whole

Continue