---
name: next-feature
description: "Prompt-first feature frontier prioritizer — answers 'which feature should we work on now?' by deriving importance/urgency from corpus, git, and authority-doc evidence, and emits rank-distorting tree defects as proposals /sp:dev-feature-change consumes. Triggers: find next, which feature, feature ranking, frontier priority, what should I work on."
license: Apache-2.0
version: 1.0.0
metadata:
  author: spur
  platforms: "claude-code,codex,openclaw,opencode,antigravity,pi"
  category: analysis-core
  interactions:
    - audit
    - recommendation
  pipeline_steps:
    - sync-check
    - gate
    - derive
    - rank
    - report
    - dispatch
see_also:
  - sp:next-router
  - sp:spur-cli
  - sp:conflict-finding
  - sp:spur-dev
---

# sp:next-feature — Feature Frontier Prioritizer

A prompt-first prioritizer that answers **"which feature should we work on now?"** — the question
`sp:next-router` deliberately does not answer (`routing-table.md` §0 step 1c: target omitted → not
v1). It derives importance and urgency from evidence already in the corpus, ranks the actionable
frontier in **tiers with per-candidate evidence**, and emits rank-distorting tree defects as
**proposals** `/sp:dev-feature-change` consumes.

**Honesty contract:** prompt-first. The model applies the rubric; existing deterministic tools
(`spur feature|task … --json`, `git`, `rg`) gather facts. No TypeScript analyzer, no numeric scores,
no new spur verbs, no feature frontmatter fields. The `priority` field is never used as an ordering —
in this corpus it is 76% one value (0493 measurement).

**Propose, never apply.** This skill performs no `spur feature move` and writes nothing under
`docs/features/**`. The only path from a structure proposal to a changed tree is
`/sp:dev-feature-change` (dry-run → confirm → apply). Ranking runs are read-only; the sole exception is
`--task`, which after an **operator confirm** (interactive, or auto-accepted under `--auto`)
dispatches `/sp:dev-plan` and `/sp:dev-refineall` — commands that write `docs/tasks*/` through their
own gates. This skill still creates no tasks itself.

## When to Use

- "Which feature next?" / "what should I work on?" / "rank the frontier" / "find next"
- Portfolio hygiene: which features are stale-done, blocked, or ill-specified.

**Do NOT use for:**

- Advancing an already-chosen task or feature — that is `/sp:dev-next` (`sp:next-router`).
- Applying tree changes — that is `/sp:dev-feature-change` (feature F31).
- Task-level ordering inside a feature — next-router's TABLE A owns that.

## Protocol

Run the steps in order. Each step's depth lives in its reference; this file is the spine.

0. **Sync-first precondition.** `spur feature sync --all --dry-run --json`. Feature `status` is
   bookkeeping, not ground truth (0493 measured 96% drift on this tree). If any proposal would
   change a frontier feature's status, the report **leads with the drift summary** ("sync first")
   and the ranking is computed over the *post-sync* status view. Details:
   [references/signal-derivation.md](references/signal-derivation.md) §0.
1. **Assemble the candidate set.** `spur feature list --json`; drop `done`/`cancelled`, drop
   structural containers (has children, no own open tasks — defect D1), drop the map/feature that
   owns this investigation if present. The `group` tag is **not** a reliable container marker (D3).
2. **Gate on actionability.** Read the frontier predicate **at runtime** from
   `plugins/sp/skills/next-router/references/routing-table.md` row **B3** — this skill intentionally
   does not copy it. A feature with zero open, unblocked child tasks is **gated, not ranked**;
   record the reason (all tasks terminal / blocked on X / no tasks). See
   [references/signal-derivation.md](references/signal-derivation.md) §1.
3. **Derive the four surviving signals** per gated survivor — AC coverage, churn exposure, dogfood
   proximity, authority pull — with the exact commands in
   [references/signal-derivation.md](references/signal-derivation.md) §2. A signal that comes back
   degenerate on the current frontier is reported as rejected-with-spread, never silently dropped.
4. **Tier, don't score.** Place each candidate per
   [references/ranking-rubric.md](references/ranking-rubric.md): ordinal tiers, explicit
   tie-breaks, evidence per candidate. No candidate carries a number derived from estimates the
   corpus does not hold.
5. **Defect pass.** Check the tree for rank-distorting defects D1–D4 per
   [references/proposal-contract.md](references/proposal-contract.md). Each emitted proposal
   conforms to the `docs/plans/feature-tree-restructure-map.md` schema and clears the evidence bar
   (`false_positive_check` mandatory). **Silence is a valid outcome.**
6. **Report + handoff.** Ranked frontier table + gated list + proposals, per
   [references/handoff-routing.md](references/handoff-routing.md). Without `--task` the report stops
   at the ranking; advancing a chosen feature is `/sp:dev-next`'s job.
7. **`--task` only — confirm, then dispatch the planning half.** Offer the rank-1 candidate (or the
   id passed as `--task <feature-id>` — which may name a **gated** feature, since T2/T3/T4 are tiers
   the rubric assigns to the gated list and only T1 comes from the ranked frontier). **Without
   `--auto`:** take an explicit operator confirmation (accept the offer, name another candidate, or
   decline). **With `--auto`:** auto-accept the offered target (rank-1 for bare `--task`, or the
   explicit id) — the flag is operator pre-consent to take the ranking's recommendation — then
   proceed without a HITL pause. Route on the tier step 4 already assigned: T3 with zero tasks →
   `/sp:dev-plan --feature <id>` then `/sp:dev-refineall --feature <id> --auto --depth ready`; T1 →
   refineall only; T3 with invalid AC, T2, and T4 stop with their reason. **This skill creates no
   tasks itself** — the dispatched commands own decomposition and its schema gate. Forward `--auto`
   into the dispatched children. Full contract:
   [references/handoff-routing.md](references/handoff-routing.md).

## Anti-patterns — do not do these

- Ranking by the `priority` frontmatter field (degenerate; fakes a signal).
- Ranking a feature whose actionability gate fails. Gate first, rank second.
- Emitting a numeric score (WSJF/RICE arithmetic) from absent value/effort estimates.
- Copying the B3 predicate into this skill. Cite it; read it at runtime.
- Any `spur feature move`, or writing proposals anywhere `docs/features/**` — feature-change owns apply.
- Decomposing a feature here, or calling `spur task create` / `spur task batch-create` under `--task`.
  Dispatch `/sp:dev-plan`; it owns decomposition and the batch-create schema gate. Equally:
  dispatching under `--auto` without `--task` (there is no confirm to skip), auto-accepting a target
  other than the offer / named `--task <id>`, or decomposing a T1 feature that already has a live
  task frontier.
- Padding the defect list with tidiness findings that move no rank.
- Re-proposing F31's rejected merges (B∪H, J∪K body-merge) or reading
  `## Applied mapping` as current state — letters are recycled; resolve against live features.

## References

| File | Owns |
| --- | --- |
| [references/signal-derivation.md](references/signal-derivation.md) | Sync precondition, B3 runtime citation, per-signal derivation commands, degenerate-spread rejection |
| [references/ranking-rubric.md](references/ranking-rubric.md) | Tier definitions, tie-breaks, evidence-per-candidate output contract |
| [references/proposal-contract.md](references/proposal-contract.md) | D1–D4 defect set, evidence bar, mapping-schema conformance, silence |
| [references/handoff-routing.md](references/handoff-routing.md) | feature-change handoff, next-router seam, the `--task` tier→hop routing table and its confirm contract |

Grounding: tickets 0493 (measured signals), 0494 (reuse ledger), 0495 (defect contract) under
feature H12.