Putting it together

Workflows, decomposition, and choosing the right tool for each step.

Real work is rarely one prompt. It is a chain: gather, draft, check, refine, deliver — often across more than one tool. The combining competency is about designing that chain deliberately: breaking work into steps, placing human review where it protects the outcome, and picking the right instrument for each link.

Done well, this is where AI stops being a novelty and becomes infrastructure.

1. Stage 1 · E1 FOUNDATIONAL

   Notice the repeat

   The first step is simply noticing: you do this same AI-assisted task every week, and you rebuild the prompt from memory every time. Stop. Save the prompt that worked, along with the context it needs and an example of a good output. That is a workflow, embryonic.

   Rewriting from scratch each week feels flexible but silently loses every refinement you ever made. A saved, evolving prompt gets better forever; a remembered one resets to average.

   Try it this week

   Pick your most frequent AI task. Save the current best prompt, its required inputs, and one good output into a note. Use the note next week instead of your memory.

2. Stage 2 · E2 PROFICIENT

   Decompose when checkpoints matter

   Chain several small steps instead of one big prompt when intermediate outputs need your verification before they feed the next step. Summarize the tickets, check the summary, then draft the report from the checked summary — errors get caught before they compound into the final product.

   But do not decompose reflexively: a task with no meaningful intermediate checkpoint is often better as one well-specified prompt. The reason for stages is verification, not ceremony.

   Try it this week

   Take a recurring multi-part deliverable and sketch it as steps with a defined input and output each. Mark the one step where an error would do the most downstream damage — that is where your review goes.

3. Stage 3 · E3 DISTINGUISHED

   Mix tools and models on purpose

   The distinguished operator matches the instrument to the step: a purpose-built tool for the recurring structured job (transcription, extraction), a general assistant for the judgment work around it, a stronger model for the step where reasoning depth pays, a fast cheap one for volume. (The "Choosing the right model" guide goes deep on this.)

   They document the workflow so it survives contact with other people: prompt, inputs, output format, review checklist, and examples of good and bad outputs. A workflow only you can run is a hobby.

   Try it this week

   For your sketched workflow, annotate each step with the tool or model class you would use and one line of why. If every step says "the same one for everything," challenge each choice.

4. Stage 4 · E4 EXCEPTIONAL

   Design systems with judgment placed exactly right

   At this level the interesting question is where the human sits. Review placed at the end catches errors after they have compounded; review placed after the riskiest step catches them cheap. The exceptional answer can always articulate why the checkpoint is where it is.

   You also know when a single prompt suffices and decomposition is overhead — and when a task should not be a workflow at all because it changes shape every time. Judgment about the meta-question is the tell of mastery.

   Try it this week

   Run a pre-mortem on your best workflow: assume it silently produced a bad output that shipped. Which step failed? Redesign so that specific failure gets caught, and write the reasoning down.

Ready to see where you stand? The free check scores all six competencies in about fifteen minutes.