| name | new-feature |
| description | Slice a whole capability into ordered releasable increments (increment backlog in increments/<stem>.md). Delegate each slice to skills/new-increment — do not implement slices yourself. Default step mode stops after each increment for feedback; automatic mode (explicit opt-in) continues until the backlog is done. Use for an entire feature or epic (e.g. full FizzBuzz), not for implementing one increment. |
| allowed-tools | Read, Edit, MultiEdit, Bash, Grep, Glob |
New feature (increment planning + orchestration)
Mission
Plan a whole feature as ordered, releasable increments. Do not implement slices in this skill — each open line is delivered by skills/new-increment (or the new-increment subagent) in a separate invocation.
First step: create increments/<feature-stem>.md (see references/increment-backlog.md).
Feature complete when every in-scope increment is [x] on that backlog — each line completed via new-increment, not by implementing in this session.
Orchestration modes
| Mode | Activation | Behavior |
|---|
| Step (default) | Assumed unless the user explicitly opts into automatic | After each increment → full post-increment review → stop for user feedback |
| Automatic | Explicit only — e.g. “modalità automatica”, “automatic mode”, “implement the whole feature”, “run until the backlog is done” | After each increment → light review (default) → if no blocking gaps and open [ ] remain, hand off the next line; if all [x], stop (feature complete) |
Invariants (both modes):
new-increment always implements one line, runs full verification (all green), commits, and stops. It never starts the next backlog line.
- This skill decides whether to continue or pause — only continues in automatic mode, and only when the review did not report blocking gaps.
- After every successful increment, run a post-increment review via
skills/refactoring / refactoring subagent before stopping or continuing.
Review depth:
| Depth | When | What |
|---|
| Full | Step mode (default); or user asked for “full post-increment review” / “review with applies” | Explain, suggest, optional tiny in-surface applies |
| Light | Automatic mode default | Explain + suggest-only — no applies (keeps long backlogs affordable) |
If the user does not name automatic mode, stay in step.
Workflow
Shared delivery rules: docs/delivery-process.md (orchestration — §1, §10).
- Clarify capability and whole-feature definition of done.
- Resolve orchestration mode (step vs automatic) from the user request; state it in the return payload.
- If feasibility unknown →
skills/spike on spike/… branch; promote before slicing.
- Write ordered increment lines in
increments/<stem>.md (all [ ] initially). Optionally note expected test layers per line (see test-strategy-selection.md §2) — e.g. “mutation likely”, “ATDD at API”, “property-based”.
- Loop (one open line at a time):
- Hand off the next open line to
new-increment (subagent or new chat). Do not implement it here.
- Wait for return payload: backlog
[x], verification all green, commits.
- Hand off post-increment review to
refactoring with feature stem, backlog line, commit SHAs, and review depth (full vs light).
- Incorporate review summary (full mode: applied tiny refactors stay committed; light mode: suggestions only).
- If review reports blocking gaps (missing tests / strategy holes for this slice) → stop; do not continue automatic — present gaps to the user.
- If step mode → stop; present review + next open line; wait for user.
- If automatic mode and open
[ ] remain → go to step 5.1.
- If automatic mode and all
[x] → stop (feature complete).
Optional: create test-lists/<feature-stem>.md with empty ## <increment-slug> headings as a skeleton — no behavior lines or [x] until new-increment runs.
See examples/fizzbuzz-increments.md.
Post-increment review (mandatory when orchestrating)
After each committed increment, invoke skills/refactoring in post-increment review mode (prefer refactoring subagent). Pass review depth (full vs light).
- Scope: only the commits / files from that increment (plus any tiny follow-up the review itself applies in full depth).
- Goals: explain what changed; flag test/strategy gaps; suggest small, targeted improvements. Full depth may apply tiny mechanical refactors only when clear leftover debt remains after the increment’s own TDD REFACTOR — not a second stylistic pass by default.
- Forbidden: rewrite the feature, expand to future increments, large redesigns, or continuing automatic when gaps are blocking.
Pending: mark review pending only when the user must run /refactoring manually or the refactoring agent is missing until ./scripts/install-cursor.sh and/or ./scripts/install-claude.sh is re-run after pull.
Details: skills/refactoring/references/post-increment-review.md.
What this skill may edit
| Allowed | Forbidden |
|---|
increments/<stem>.md | src/, test/, tests/ |
Optional skeleton test-lists/<feature-stem>.md (headings only) | acceptance-examples/ (created per increment in new-increment) |
| Docs clarifying scope | Production or test code |
Orchestration may invoke subagents that edit code; this skill’s own edits stay planning-only.
Composition
| Need | Skill / agent |
|---|
| One slice delivery | skills/new-increment / new-increment |
| Post-increment review | skills/refactoring / refactoring (post-increment review mode) |
| Untested area | skills/legacy-testing (before or per increment) |
| Green prep | skills/refactoring (dedicated session) |
| Unknown tech | skills/spike |
This skill does not run ATDD/TDD or mark increments [x] from implementation work.
Anti-patterns
- One backlog line for the entire feature
- Implementing code or tests in the new-feature session
- Marking multiple
[ ] → [x] without a new-increment pass each
- Skipping
new-increment (no scoped TDD per slice)
- Assuming automatic mode without explicit user opt-in
- Continuing to the next increment in step mode without user go-ahead
- Skipping post-increment review between increments
- Continuing automatic after a review that reported blocking gaps
- Asking
new-increment to “do the rest of the backlog”
- Running full apply-capable review on every automatic increment when light depth would suffice (cost blow-up)
- No on-disk
increments/ file
- Creating many per-increment markdown files during planning
See references/anti-patterns.md.
Additional resources