| name | overseer-to-issues |
| description | Break a PRD into independently-grabbable Issue files written into the PRD's folder under the Overseer root, using tracer-bullet vertical slices. Triage is folded in — each Issue is born routed (ready-for-agent / ready-for-human). Use when the user wants to convert a PRD or plan into Issues. |
To Issues
Break a plan into independently-grabbable Issues using vertical slices (tracer bullets), and write them as files into the PRD's folder under the Overseer root.
The Overseer model (read this first)
An Issue is a markdown file inside a PRD directory. It belongs to that PRD by virtue of where it lives — there is no "parent" reference, the folder is the parent. Issues are read by the Overseer TUI as kanban cards, and by Overseer's dispatcher to spawn implementor agents — so the frontmatter contract below must be emitted exactly. This skill is self-contained — everything you need is below.
Resolve the root and the PRD folder
Read the root from ~/.config/overseer/config.toml (root = "~/...", expand a leading ~). The Issues go into a PRD folder under that root — a directory containing a prd.md. Identify the target PRD (from the argument, the conversation, or by asking). Never assume the cwd is the root; you are usually standing in a code repo.
File contract — match it exactly
The dispatcher parses these fields; emit them verbatim or work won't dispatch.
Filename: NNN-slug.md (e.g. 001-auth.md, 002-password-reset.md).
NNN is a zero-padded 3-digit sort key. Scan the PRD folder for existing NNN-*.md, take the numeric max (0 if none), and assign max+1, max+2, … across the batch. Never reuse or backfill gaps — only ever append above the current max. Order is incidental/positional, not priority.slug is kebab-case from the Issue title.
Frontmatter:
---
title: Password reset flow
status: ready-for-agent
repo: /Users/you/code/app
blocked_by:
- 001-auth.md
slice: 2-api
---
status — folded-in triage (see below). AFK slice → ready-for-agent; HITL slice → ready-for-human. There is no backlog/needs-triage limbo; an Issue is born routed.repo — the local absolute toplevel path of the code repo where this Issue's work happens (git rev-parse --show-toplevel). Required for AFK Issues — the dispatcher skips an Issue with a missing/invalid repo. See "Sourcing repo" below.blocked_by — a YAML list of full sibling filenames (001-auth.md), NNN- prefix included. The dispatcher reads this verbatim to gate dispatch. Omit the field entirely when there are no blockers. Blocking lives here in frontmatter, not in the body.slice — an optional N-name stack slice (e.g. 2-api): the number is the slice's chain position, the name the one-line concern. Omit it entirely by default — an absent slice: is today's single-PR behaviour. Emit it only when the "Slicing judgment" step below decides to slice, and then on every Issue. Authored here and only here; the board only reads it (CONTEXT.md → Slice, ADR 0024).
Body: ## What to build + ## Acceptance criteria (template below). No "Parent" section — the folder is the parent.
Process
1. Gather context
Work from whatever is in the conversation (typically a /overseer-grill-with-docs + /overseer-to-prd session). If the user passes a PRD reference (folder name or path), read its prd.md.
2. Explore the codebase (optional)
If you haven't already, explore the relevant code repo(s) to ground the slices. Use each repo's domain glossary (CONTEXT.md) for titles and descriptions, and respect ADRs in the areas you touch.
3. Draft vertical slices
Break the plan into tracer bullet Issues. Each is a thin vertical slice cutting through ALL integration layers end-to-end, NOT a horizontal slice of one layer.
Slices are AFK or HITL. HITL slices require human interaction (an architectural decision, a design review); AFK slices can be implemented and merged without it. Prefer AFK where possible.
- Each slice delivers a narrow but COMPLETE path through every layer (schema, API, UI, tests)
- A completed slice is demoable or verifiable on its own
- Prefer many thin slices over few thick ones
4. Quiz the user
Present the proposed breakdown as a numbered list. For each slice show:
- Title
- Routing: AFK (
ready-for-agent) / HITL (ready-for-human)
- Repo: which code repo the work happens in
- Blocked by: which other slices (by their planned filename) must complete first
- User stories covered (if the source has them)
Ask:
- Does the granularity feel right? (too coarse / too fine)
- Are the dependency relationships correct?
- Should any slices be merged or split further?
- Are the AFK/HITL routings correct? (this is the triage decision — it sets each Issue's
status)
- Is each slice's
repo correct?
Iterate until the user approves.
Sourcing repo (triage detail)
A feature may span several code repos — typically the session's working directories (the dir you launched in plus any /add-dir'd). A good /overseer-grill-with-docs + /overseer-to-prd session will have established which repo each slice targets; read that from context. For each candidate repo, the repo value is its git rev-parse --show-toplevel. When only one repo is in play, stamp them all the same. When a slice's repo is genuinely ambiguous, ask.
Slicing judgment (the slice: field)
Word warning. A vertical slice above is the tracer-bullet authoring unit (one Issue). A Slice here is something else: a named, ordered group of Issues that becomes one PR in a stack (CONTEXT.md → Slice). This step decides whether to group the Issues you just drafted into a stack of PRs instead of one.
This is the default-off exception: the common case is one PR per PRD, so most PRDs emit no slice: field at all. Only consider slicing when both gates pass — fail either and you emit no slice: fields:
- Too big — the predicted whole-PRD diff would not be comfortably reviewable as one PR. Judge from a rough size estimate (anchors: ~400 lines starts to hurt; ~600+ lines or ~10+ files is worth splitting) plus a cognitive-load sniff test ("reviewable in one sitting?") that can override the anchors either way: don't split a 1000-line mechanical rename; do split a gnarly 350-line PR of three unrelated concerns.
- Clear split — you can group the Issues into 2–4 contiguous, one-line-nameable concerns (e.g.
schema, api, ui), in chain order, where the grouping respects the no-forward-dependency invariant: every Issue in slice N may be blocked_by: only Issues in slice ≤ N. A forward edge makes the stack unbuildable.
When both gates pass, hand your candidate grouping to the pure slice planner (src/dispatch/slicePlanner.ts → planSlices) rather than numbering by hand — it verifies the invariant, applies the soft cap of ~3–4 slices (merging adjacent nameable groups to fit), and falls back to a single PR (no slice: fields) on any failure:
- Pass it
{ tooBig: true, groups, blockedBy }. groups is your ordered list of { name, issues, mergeable? } (bottom-of-stack first; mark a group mergeable: false when folding it into a neighbour would make it un-nameable). blockedBy is the blocked_by graph keyed by Issue filename.
- If it returns
{ sliced: false }, emit no slice: field — that is the single-PR fallback, literally today's behaviour.
- If it returns
{ sliced: true, assignments }, stamp each Issue's slice: from assignments[issueFilename] (e.g. slice: 2-api) when you write the files below.
Slicing is orthogonal to dispatch — every Issue still dispatches in parallel off the one feature branch (ADR 0024); the slice: field is read only later, at the human's Open PR keypress, to materialize the stack.
5. Write the Issue files
Write each approved slice to <root>/<prd>/NNN-slug.md using the contract above. Process slices in dependency order (blockers first) so you know each blocker's final filename before writing the Issues that reference it in blocked_by.
## What to build
A concise description of this vertical slice. Describe the end-to-end behavior, not layer-by-layer implementation.
Acceptance criteria
Do NOT modify the prd.md. A PRD has no stored status — its board column is derived from these Issues automatically.
6. Hand off to the board
The Issues now exist as a live board, not as work to start here. This skill is the terminal station of the authoring chain — the chain converges on the board, where dispatch is tracked (the Reactor, liveness, kill). Do not offer to dispatch agents, run the Issues, or start implementation from this conversation. Igniting the work is a board action the user takes, not something this skill does — offering it here bypasses every operational affordance the board exists to provide. Close by pointing the user at the board:
Next: open Overseer, select <PRD>, press d.
(Substitute <PRD> with the PRD's display title.)
