Yun Slice Plan

A tracer bullet is a thin path fired end to end — one ticket you can grab and finish in a single sitting. Slicing a plan into tracer bullets, each declaring its blocking edges, is what turns an airtight spec into executable work.

Steps

1. Gather context

Work from the plan or spec already in the conversation. If the spec was published out of conversation — a fresh agent, a prior session — fetch the spec kind by its feature slug through skills/_shared/artifact-convention.md first. Done when the full plan or spec is in front of you.

2. Explore the codebase, if you haven't

Read enough of the code to slice against its real shape — so slices respect the seams already there. Vocabulary comes from the domain model, not the code: consult it through skills/_shared/domain-convention.md, since the glossary knows which synonyms it rejects and the code only carries whichever name won last. Look for prefactoring that makes the work easier: make the change easy, then make the easy change — any prefactoring becomes the first slice. Done when you can name tickets in the project's ubiquitous language and have decided whether prefactoring is needed.

3. Draft the vertical slices

Break the work into tracer bullet tickets.

• Each slice cuts a narrow but complete path through every layer the change touches (e.g. schema, API, UI, tests) — vertical, end to end
• A completed slice is demoable or verifiable on its own
• Each slice is sized to fit in a single fresh context window

Give each ticket its blocking edges — the tickets that must finish before it can start. A ticket with no blockers can start immediately.

Wide refactors are the exception. A wide refactor — rename a column, retype a shared symbol — has a blast radius that breaks thousands of call sites at once, so no vertical slice lands green. Sequence it expand–contract instead: expand (add the new form beside the old, nothing breaks), migrate the call sites in batches sized by blast radius (each batch a ticket blocked by the expand), then contract (delete the old form, in a ticket blocked by every batch). Keep CI green batch to batch where each batch stands alone; where it can't, the batches share a named integration branch that each of them declares, and all block one final integrate-and-verify ticket that takes that branch as its base — green is promised only there. Naming the branch on the ticket is what makes the batch recognisable to whatever builds it: a skill that builds one ticket at a time reads that line, declines the batch instead of chaining the rest of the plan onto a red one, and reports the branch the batch needs. A batch that promises green downstream and names nowhere to put it is indistinguishable from an ordinary ticket that simply fails.

Done when every slice obeys the rules above (expand–contract batches excepted) and its blocking edges are drawn.

4. Quiz the user

Present the breakdown as a numbered list. Per ticket show Title, Blocked by, and What it delivers — the end-to-end behaviour it makes work. Then ask:

• Is the granularity right — too coarse, too fine?
• Are the blocking edges correct — does each ticket depend only on tickets that genuinely gate it?
• Should any tickets merge or split?

Iterate until the user approves; publish only once they do.

5. Publish the tickets

publish each approved ticket through skills/_shared/artifact-convention.md, in dependency order (blockers first) so each ticket's edges can name real ones. Work the frontier — any ticket whose blockers are all done; for a linear chain that is top to bottom. Fill the template per ticket; the convention decides where it lands and how its edges are drawn.

Publish every ticket open. That is where its lifecycle starts, and the skill that builds it flips it to resolved once its branch is gated and chained — slicing never writes a terminal state. How open is expressed is the convention's call, not this skill's. Done when every ticket the user approved in step 4 has a published artifact, all of them open.

Then tell the user to type yun-build-plan, with the repo, the feature key, and the slug it should prefix each run's branch with: it walks these tickets to completion, cutting each run from the last one's branch, and leaves the final branch to review. That skill is user-invoked, so this one cannot reach it. Slicing stops at published tickets — it never dispatches one, and an expand–contract batch it does not walk at all: say plainly that those tickets are built by hand.

—

What to build: the end-to-end behaviour this ticket makes work, from the user's perspective.

Blocked by: the numbers of the tickets that gate this one, or "None — can start immediately".

• Acceptance criterion 1
• Acceptance criterion 2

The number leads the title, and blocking edges name numbers rather than titles — that number is the join key the convention addresses a ticket by, and it survives whatever id a tracker assigns. Publishing also stamps the ticket open — a Status: line where the convention's floor is a file, the tracker's own open state otherwise; that shape is artifact-convention.md's, which is why the template above does not carry it.

One further line is written on an expand–contract batch and on nothing else:

**Integrates on:** <the named integration branch this batch merges onto>

It is kept out of the template on purpose. A building skill reads that line to decide whether to build the ticket at all, and the template is where a field description gets dutifully filled in — so a template slot here would put Integrates on: None on every ordinary slice and a builder would decline the entire plan. Absence is what marks an ordinary ticket. Never write the line empty, and never write it as None.

Tickets describe behaviour and acceptance criteria. A file path or code snippet goes stale fast — the one exception is a prototype snippet that pins a decision prose can't (a schema, a type shape, a state machine); inline just the decision-rich part and say it came from a prototype.

Attribution

Adapted for this workspace from Matt Pocock's to-tickets (github.com/mattpocock/skills, MIT). Changes: renamed to the yun-<verb>-<noun> convention; publish and fetch route through skills/_shared/artifact-convention.md instead of a setup-bound tracker config, so the skill never names a tracker; Matt's two ticket templates (local vs. issue) collapse into one, since the convention — not the skill — maps a ticket's edges onto the backend; dropped the triage-label vocabulary and the OpenAI agent wiring. The vertical-slice rules, blocking edges, expand–contract exception, and frontier are preserved from the source.