Breaking an epic into agent-sized work · Agentic Engineering: the hands-on course

Turn a big feature into thin vertical slices, each with its own brief and check, in an order that lets you learn and ship as you go.

Lesson 3.1 said to split tasks that are too big. This lesson is about how. Splitting badly is easy: you end up with pieces that can’t be tested alone, or that only work once all of them are done.

The skill is to cut an epic into slices that each deliver something working, and to write the breakdown down where every session can find it.

Slice vertically, not by layer

Horizontal layers can only be tested once all of them exist. Vertical slices each cut through every layer and can be verified alone.

The first slice is a walking skeleton: the thinnest path that works end to end, even if it’s ugly. It flushes out the surprises early, and every later slice builds on something that already runs.

The breakdown, step by step

1
Write the epic outcome
One paragraph: who gets what, and how you’ll know it’s done. This is yours to write, not the agent’s.
2
Let the agent draft the breakdown
Read-only. Ask it to explore the code and propose slices. It’s good at spotting the files and dependencies involved.
3
Cut and reshape
Merge slices that can’t be tested alone, split any that fail the sizing questions from lesson 3.1.
4
Order by dependency and risk
Skeleton first, then the riskiest unknowns, then the rest. Mark which slices don’t depend on each other.
5
Save it in the repo
One file with the slices, each with its acceptance criteria and a status. Every session reads it and updates it.

Here is the prompt I use for the draft. The output goes into a file, not just the chat, so it survives resets (Module 2).

Epic: [one paragraph outcome]

Explore the code. Don't change anything except the plan file.
Write docs/plans/csv-export.md with:

- Slices: each one a thin vertical change that works and can be
  tested on its own. Start with the thinnest end-to-end version.
- For each slice: one-sentence goal, files likely touched,
  2-4 acceptance criteria, depends on (slice numbers or none)
- Risks and unknowns: what might change the plan
- Status: a checkbox per slice

Aim for slices a reviewer can check in one sitting.

Play: good slice or not?

🎮 Judge the slice

1 / 7 · Score: 0

Each card is a proposed slice from an epic breakdown. Keep it, or send it back for rework?

Before you move on

✅ Key takeaways

0 / 4 completed

I split epics into vertical slices that each work and can be tested alone.
My first slice is a walking skeleton: the thinnest end-to-end path.
I let the agent draft the breakdown, but I write the outcome and reshape the slices.
The breakdown lives in a plan file in the repo, with criteria and a status per slice.