Breaking any role into AI-ready tasks: a five-step method

Executive overview

Most people try to use AI at the wrong level — the job title or department — instead of the specific activities that make up the role. The fix is task decomposition: breaking roles into discrete activities with clear steps, inputs, outputs, and criteria.

Once decomposed, each activity becomes a unit AI can reliably execute. The five-step method produces a ranked list of activities and a folder-based workspace structure ready to hand off to an AI agent.

The skill that compounds is the ability to break your own work into pieces small enough for AI to follow.

Step 1: List all activities in a role

• Ignore departments and job titles; zoom into what actually happens week to week
• Use the "GoPro test": what would a camera strapped to your chest capture you doing?
• A single role typically yields 20–30 distinct activities

Step 2: Pick three to five and write out the steps

• Prioritise quick wins: simple, repetitive tasks with steps already clear in your head
• Prioritise big time savers: automating these returns hours per week, and the data is readily available
• Defer low-priority or rare tasks (every six months or annually) — the ROI rarely justifies the effort
• For each chosen activity, list every step explicitly; avoid vague words like "realistic"
• Define vague terms for the AI (e.g. "realistic timeline" = every phase has a one-week buffer, total length does not exceed similar past projects)

Using AI to extract steps: paste a structured interview prompt into a fast model (GPT-mini or Claude Haiku). The prompt asks AI to interview you one question at a time — up to 15 questions — and deliver back a numbered step list, inputs/outputs, and any criteria or heuristics. Saves significant time versus writing steps manually.

Step 3: Define inputs and outputs

• State what data the AI will receive (e.g. a CSV export, a draft proposal document)
• State what the AI must return (e.g. an on-track/off-track list, redline edits with notes)
• Be explicit about format and file type
• If you used the AI interview in Step 2, inputs and outputs may already be captured

Step 4: Rank the activities

Score each of the three to five activities on three criteria:

1. Is the data easy to give to the AI?
2. Are the steps clearly written down?
3. How much time does automating this actually save?

The highest-scoring activity is where you start.

Step 5: Structure folders and instructions

• Give each activity its own dedicated folder on your desktop or in the cloud
• A focused AI — one activity, one folder — produces higher-quality outputs than one handling broad context
• Each folder contains three things to start:
  • An instructions file (CLAUDE.md for Claude/Claude Code, agents.md for OpenAI Codex) — this holds the steps, criteria, and rules
  • The input file or data the AI will process
  • A location for the output the AI returns
• As you scale to multiple clients or projects, add subfolders per client; each subfolder maps to one activity
• If an activity (e.g. proposal review) applies across all clients with only content changing, convert it into a reusable skill in Claude or ChatGPT — call the skill from any project folder without duplicating instructions