New: Boardroom MCP Engine!

Ready to put this into action?

Get the complete AI Integration PlaybookPractical AI implementation guide — prompt engineering, workflow automation, and ROI frameworks.

Article 015 · Part 2

Break a Big Job into Small, Checkable Steps

Make progress visible by giving each stage something concrete to produce.

By Randy Salars · Published

On this page
  1. Define the finished package
  2. Name outputs rather than vague activities
  3. Notice what has to happen first
  4. Make the numbers inspectable
  5. Assign ownership where a person must act
  6. Save progress at meaningful checkpoints
  7. Reassemble the parts and check the joins
  8. Students: plan around evidence of learning
  9. Your exercise

Make progress visible by giving each stage something concrete to produce.

“Create our fundraiser” sounds like one request. It contains several jobs: decide what will happen, estimate costs, choose a place, organize volunteers, prepare messages, and confirm who is responsible.

An assistant may respond with a confident ten-page plan. Yet the room is not booked, the prices are estimates, and two tasks depend on a volunteer who has never agreed to help.

The missing ingredient is a sequence of outputs that can be checked as the project develops.

You already know this idea from ordinary life. You do not finish a garden by writing “do gardening” on a calendar. You choose the bed, prepare the soil, obtain plants, and check what each stage needs before moving on.

Define the finished package

For this lesson, imagine a fictional community book fair that will raise money for a reading group. The immediate assignment is to prepare an organizer's planning package. It does not include holding the fair, spending money, booking the room, or sending announcements.

The package will contain:

  • A one-page event brief.
  • A comparison of two possible rooms.
  • A draft expense budget.
  • A volunteer schedule with unfilled roles visible.
  • A draft invitation.
  • A final list of decisions required before launch.

This scope lets the assistant do substantial work while keeping the distinction between preparation and real-world completion clear.

For another project, the authorized scope might include sending approved messages or creating calendar events. State that explicitly. A project plan should match the authority actually given.

Name outputs rather than vague activities

“Research venues” can continue indefinitely. “Compare two available rooms on cost, capacity, access information, and booking conditions” has a visible endpoint.

“Work on the budget” is vague. “Produce an expense table with confirmed quotes separated from estimates” is checkable.

Here is a six-stage plan for the fictional fair:

StageOutputDepends onCompletion check
1Event briefOrganizer's goals and constraintsPurpose, audience, date options, and spending limit are explicit
2Room comparisonBrief and venue informationSources and unresolved venue questions are shown
3Draft budgetRoom choice or separate room scenariosArithmetic is correct; estimates are labeled
4Volunteer scheduleEvent tasks, room timing, supplied availabilityEvery task has an owner or an explicit vacancy
5Draft invitationConfirmed public event factsNo unconfirmed details appear as promises
6Organizer's packageOutputs 1–5Dates, costs, roles, and status agree across documents

Each stage produces something you can open and inspect. That makes it easier to see whether the project is moving forward.

Notice what has to happen first

A dependency is simply something a task needs before it can be finished.

The invitation needs a confirmed date and place before it is ready to send. Its wording can be drafted earlier with clearly labeled missing fields, but it is still a draft.

The room comparison and a preliminary volunteer task list may proceed independently once the event brief exists. You do not need a chosen room to recognize that someone must welcome visitors. You do need room timing to finalize that person's shift.

This distinction keeps work moving without pretending uncertainty is resolved.

Ask:

Identify which outputs can be drafted now, which require a decision, and which can proceed independently. Do not mark a dependent task complete merely because a draft exists.

Make the numbers inspectable

Suppose the fictional spending limit is $300. The selected room has a recorded quote of $80. Other amounts are planning estimates:

ExpenseAmountBasis
Room$80Supplied quote
Printing$35Estimate
Signs$20Estimate
Supplies$45Estimate
Contingency allowance$30Planning allowance
Total$210Sum of listed amounts
Unallocated portion of $300 limit$90$300 minus $210

The table does not prove that these are all the expenses. Ask whether the event brief implies missing items, and check the scope with the organizer.

It also does not show money raised. Gross receipts, expenses, and the remaining proceeds are different quantities. A fundraiser can meet its spending limit and still raise less than hoped.

Have the assistant show the arithmetic, then verify it with a calculator or spreadsheet. A neat table is a presentation, not an independent calculation check.

Assign ownership where a person must act

An assistant can draft a venue question. Someone must obtain an answer from the venue unless an authorized tool can actually do so.

A useful plan names the role responsible: organizer confirms booking conditions, volunteer coordinator supplies availability, treasurer checks expenses. If names are not known, leave the role unassigned. Do not turn a suggested person into a confirmed owner.

For each external action, distinguish its status: proposed, ready for review, authorized, or completed with evidence. This is particularly useful when a project mixes drafts with actions in connected applications.

The technical distinction between a predefined AI workflow and a more autonomous agent is explored in Anthropic's engineering guidance. Here, you are building a human-directed sequence whose outputs you can inspect. Anthropic: Building effective agents.

Save progress at meaningful checkpoints

After confirming the room, save the updated brief and its supporting information. After reviewing the budget, save that version too.

A checkpoint should capture a meaningful decision or completed output. You do not need a separate archive for every reworded sentence.

Useful filenames might be book-fair-brief-v2.md and book-fair-budget-v1.csv. Add a short status note explaining what is confirmed and what remains open. Avoid calling a file “final” merely because the assistant finished generating it.

If a later step fails, you can resume from the last sound checkpoint. You should not need to reconstruct settled facts from a long chat.

Reassemble the parts and check the joins

Separate outputs can disagree even when each looks reasonable. The budget might use Room A while the invitation names Room B. The schedule might end after the rental period. The invitation might promise a service with no assigned volunteer.

At the end, compare the documents against a shared fact list. Focus on details that appear in more than one place: date, location, money, capacity, responsibilities, and public promises.

This final check is where the project becomes a coherent package rather than a folder of plausible drafts.

Students: plan around evidence of learning

For a permitted class project, six outputs might be an assignment checklist, a research question, a source list, an outline, your own draft, and a revision record.

Give each a completion check. “Five sources found” is weaker than “Each source is relevant, opened, and recorded with the claim it supports.”

If AI helps organize the project, keep your own reading, calculations, and argument visible. A polished presentation does not show whether you understand its contents. Add a short practice explanation without the slides or assistant.

Your exercise

Choose a project you can define in one paragraph. Create six outputs, list their dependencies, and assign a completion check to each. Mark who supplies facts and who makes decisions.

Then remove one unnecessary step or combine two that do not benefit from separate review. The purpose is clear progress, not a larger planning document.

Completion check: Every stage produces something checkable, dependent work is sequenced realistically, and the final package agrees with itself.

Get the AI Dispatch

Weekly insights on ai & technology — delivered to your inbox. No spam, unsubscribe any time.

Want to choose specific topics? Customize your interests