Choose the work. Choose the gear. Package the expertise. Keep human judgement and exact execution visible.
The finish line
You will leave with one real workflow classified, measured, bounded, and ready for a build — or an intentional decision not to use AI.
Progress is saved in this browser.
The five decisions
Name the business result and baseline
Match the work to the right gear
Decide whether the procedure should become a Skill
Separate AI work from deterministic execution
Place human judgement, review, and accountability
The output is not another list of AI ideas. It is one reviewable decision about work that matters, how it should run, and who remains accountable.
The Longhand model
Each gear owns a larger unit of responsibility. The leader still sets direction, defines the boundary, and reviews the result.
Use when you need a quick answer or idea.
Use when you need a document made.
Use when you need a multi-step task done.
Use when you need a system that runs itself.
Shift up only when the unit of work requires it. Maximum autonomy is not the goal; useful ownership with visible accountability is.
Gear 1
One question or idea
Use it when
you need a quick answer or idea.
Example
“What are the three biggest risks in this plan, based on the attached assumptions?”
Human boundary: A person supplies the context and decides which risk matters now.
Gear 2
One finished deliverable
Use it when
you need a document made.
Example
“Create a board-ready memo from these approved sources and this reporting template.”
Human boundary: A person owns the audience, claims, tone, and approval to share.
Gear 3
One multi-step objective
Use it when
you need a multi-step task done.
Example
“Research the competitors, compare the evidence, and produce a review-ready recommendation.”
Human boundary: A person defines scope, permissions, stop rules, and the acceptance decision.
Gear 4
One recurring outcome
Use it when
you need a system that runs itself.
Example
“Every Monday, gather current signals and produce a linked priority brief.”
Human boundary: A person owns operating policy, exceptions, and changes to what the loop may do.
Step 1 checkpoint
You are done when
You can classify a quick decision question as Chat, a board memo as Assistant, a research-and-draft brief as an Agent, and a weekly priority heartbeat as a Loop — and explain what stays human in each one.
The future of work
General agents bring raw intelligence. Agent Skills add the repeatable procedure that turns a brilliant generalist into a reliable professional teammate.
The Future of Work 2026 · 7 minutesWatch: The Future of Work 2026
youtube.com/watch?v=TXiaDHn1dUk
A useful computing analogy
Domain experts can capture how work is really done. The team library compounds, and a new teammate can inherit working knowledge instead of starting from zero.
Step 2 of 5 · 9 min
Build a library of professional procedures instead of one giant agent.
A general model can reason, but repeatable professional work also needs the procedure, sources, tools, examples, and quality rules a good colleague would learn over time.
Do this
A useful Skill carries the procedure
| Part | What it holds |
|---|---|
| Instructions | The sequence, decisions, and boundaries. |
| Approved sources | The evidence and context it may use. |
| Tools and scripts | The reliable actions it can run. |
| Good examples | The quality pattern worth repeating. |
| Finish and stop rules | What done looks like and when to stop. |
Step 2 checkpoint
You are done when
You can explain why a Skill is reusable procedural knowledge — not a fifth gear — and name one piece of team know-how worth capturing as a Skill.
Step 3 of 5 · 8 min
Use TDAR and the DMP before deciding what to build.
Once you can build an agent for almost anything, capability stops being the constraint. The harder leadership job is choosing work that returns more than it costs.
Do this
TDAR — define the return before the solution
| Decision | Write this |
|---|---|
| Result | The business measure this work should change. |
| Baseline | The number and current cost today. |
| Contribution | What AI takes and what stays human. |
| Re-measure | The date and evidence used to judge return. |
The Decision Making Process
The Big What
Name the decision topic before debating the answer.
Why
Why act — or not act — at all?
What
What specifically must be true when this is done?
How
What steps, resources, and ownership could achieve it?
What If
If we take this path, will it produce the result we intended?
Step 3 checkpoint
You are done when
Your candidate has a measurable result, a baseline, a re-measure date, and a complete DMP that makes the decision to proceed — or not proceed — visible.
10 / 80 / 10
10%
Set the result, context, approved data, permissions, and decision gates.
80%
Research, extract, compare, draft, process, and prepare the reviewable result.
10%
Check the Definition of Done, resolve exceptions, approve, and feed learning back into the Skill.
Put approval at the point where judgement occurs. One sign-off after a long run may reveal that something is wrong without revealing where it went wrong.
Step 4 of 5 · 9 min
Separate interpretation, exact execution, and accountable decisions.
AI is probabilistic. Some work benefits from interpretation; other work must produce the same exact result, account for every record, and be safe to rerun. Strong workflows use both deliberately.
Do this
Remember: A final review is an approval boundary, not a substitute for schema checks, transactions, idempotency, reconciliation, or other deterministic controls.
Three different jobs in one safe system
| AI can help | Deterministic systems own | Humans decide |
|---|---|---|
| Interpret messy headers | Validate schema and types | Approve ambiguous mappings |
| Suggest classifications | Deduplicate with explicit rules | Resolve exceptions |
| Explain rejected rows | Import transactionally | Authorize the real run |
| Draft remediation | Reconcile exact counts | Accept business risk |
Worked boundary · data import
AI can help interpret ambiguity. It should not silently own the transaction, identity rules, or proof that every record landed where it belongs.
Schema, types, required fields, and exact row count.
Propose mappings only for ambiguous fields, with evidence.
Accept, correct, or reject ambiguous mappings and exceptions.
Apply explicit rules without changing production data.
Use an idempotent transaction that is safe to retry.
Account for every imported, rejected, and quarantined row.
input rows = imported + rejected + quarantined
Step 4 checkpoint
You are done when
You can place each part of a workflow in the AI, deterministic, or human lane and explain why a final skim alone cannot prove an exact operation was correct.
Step 5 of 5 · Run it
Do this
AI Work Decision Card
AI WORK DECISION CARD WORKFLOW Name the real job or recurring outcome: TDAR — RESULT FIRST Business result: Current baseline: AI contribution: Human contribution: Re-measure date: DMP Big What — the decision topic: Why — why this matters: What — the exact result and criteria: How — the proposed path: What If — risks, rewards, and decision to proceed: RIGHT GEAR Chat / Assistant / Agents / Loops: Why this gear fits: REUSABLE SKILL Should this become a Skill? Yes / No Instructions, approved sources, tools, examples, and quality rules it needs: WORK BOUNDARIES AI may: Deterministic software must: A person must decide or approve: FINISH LINE Definition of Done: Stop rules and exceptions: Accountable owner:
The overview is complete
Keep the card with the work. Use the audit to identify the next capability gap, then choose the training that helps you shift up without hiding the judgement or control the result requires.
Assess
Rate the evidence you have today and open the exact training for each red or yellow gap.
Open trainingSet up
Choose a home base, set the data boundary, connect work, and save a reusable workflow.
Open trainingBuild
Turn the chosen job, workflow, toolbox, quality criteria, and stop rules into a reusable Skill.
Open training