Skip to main content

development-lifecycle

Use when doing frontend/React/TypeScript/UI work. Auto-guides phase: understand -> plan -> grill -> implement (TDD) -> /go (4->4b review->5->5b->6). One skill, full lifecycle. Alias: /work.

Quellinformationen

Repository
redpanda-data/ui-harness
Letzte Quellaktivität
27. Mai 2026 um 14:03
Erkannte Sprache von SKILL.md
Englisch
Sterne
34
Forks
4

Installationsoptionen

Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.

Quelldateien prüfen

Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.

Datei-Explorer
2 Dateien

SKILL.md wird angezeigt

SKILL.md
Quellanweisungen · Schreibgeschützte Vorschau
name
development-lifecycle
description
Use when doing frontend/React/TypeScript/UI work. Auto-guides phase: understand -> plan -> grill -> implement (TDD) -> /go (4->4b review->5->5b->6). One skill, full lifecycle. Alias: /work.
# Development Lifecycle Auto-detect phase, guide process. ## Phases ### 0. Worktree (automatic, silent if already isolated) Feature branch? Skip. On main/master/develop? Spawn worktree before code. [ETHOS: Worktree Isolation] scripts/mux-worktree.sh <type>/<branch-name> Claude invoke silent when phase 1 start on default branch. User never run. ### 1. Understand - Explore | clarify one-at-a-time | new->2-3 approaches+tradeoffs | bug->failing test->root cause - Spawn background agents: alternatives, prior art, edge cases parallel - Mixed patterns area? Refactor to single pattern FIRST before add features - **GATE: no impl code until approach approved.** ### 2. Plan - Every step: exact file paths, exact code, expected output. No placeholders. - Bite-sized tasks (2-5 min each) - UI work: use `/prototype` for 2-3 runnable UI variations, review with user, pick best. See [REFERENCE.md](REFERENCE.md). - 5+ tasks -> stacked PRs (one per logical group) - Complex plan (5+ tasks, multi-stakeholder)? Consider `/ultraplan` - `/codex:rescue` available -> auto-dispatch second opinion ### 2b. Grill **GATE: no impl until plan survive grilling.** - Auto-invoke `/grill-with-docs` | grill until every branch resolved | update CONTEXT.md + ADRs inline - Update plan with changes | get explicit user confirmation - Skip only if: trivial bug fix AND <3 tasks AND no architectural decisions ### 3. Implement (TDD) - RED: failing test first | GREEN: minimal code to pass - **Test deletion guard**: verify test+assertion count not decrease after GREEN. AI may weaken tests -> reject and redo. - REFACTOR while green | no `setTimeout` hacks | run `--detectAsyncLeaks` ### 4-6. Ship -- `/go` Impl done -> run `/go` to ship. Handle all: - **4. Verify** -- types + lint + tests + browser smoke - **4b. Review / Refine** -- self-reviewer + adversarial-reviewer agents (4b->5) - **5. Ship** -- `/simplify` -> `/commit-push-pr` -> code-reviewer agent - **5b. Iterate** -- monitor CI -> `/resolve-pr-feedback` -> AI self-review: up to 3 rounds, early-exit on clean; human review: address ALL (hook-enforced) - **6. Compound** -- codify lessons as `.claude/rules/` See `/go` skill full details. See [REFERENCE.md](REFERENCE.md) phase-specific checklists. ## Phase Selection Full flowchart [REFERENCE.md#phase-flowchart](REFERENCE.md#phase-flowchart). | User says | Phases | |---|---| | "Build a new feature" | 1->2->**2b**->3->**`/go`** | | "Fix this bug" | 1(reproduce)->3(TDD)->**`/go`** | | "Refactor this module" | 1->2->**2b**->3->**`/go`** | | "Write tests for X" | 3 only | | "Ship it" / "Create a PR" | **`/go`** only | | "Quick question" | Just answer | | "Batch these 5 issues" | **Sandcastle** -- parallel agents | | "Work on this overnight" | **Sandcastle** -- AFK delegation | See [REFERENCE.md](REFERENCE.md) detailed checklists + Sandcastle integration.
Auf GitHub ansehen