← ClaudeAtlas

do-sdlclisted

Use when supervising a full SDLC pipeline run in a local Claude Code session. Loops the /sdlc router — dispatching each stage to a subagent on the stage-appropriate model (opus/sonnet) — until merge, a blocking guard, or the iteration cap. Triggered by 'do-sdlc', 'run the full pipeline', 'ship this issue end to end', 'supervise the sdlc'.
tomcounsell/ai · ★ 14 · AI & Automation · score 73
Install: claude install-skill tomcounsell/ai
# do-sdlc — Local Pipeline Supervisor This skill is the **local stand-in for the bridge PM session**. `/sdlc` is a single-stage router by contract: it dispatches ONE sub-skill and returns, expecting a PM session to re-invoke it. In a local Claude Code session there is no PM loop — this skill IS that loop. You are the supervisor, not the worker. You assess, dispatch, and track. The stage subagents do all the work. ## Hard Rules 1. **NEVER write code, run tests, or create plans directly** — every stage executes inside a stage subagent that invokes the stage's `/do-*` skill. 2. **NEVER decide dispatch yourself** — `sdlc-tool next-skill` is the only source of dispatch decisions. It encodes all guards (G1–G7) and dispatch rows. Do not second-guess it, reorder stages, or skip it "because the next stage is obvious". 3. **NEVER continue past a `blocked` decision** — surface the reason to the human and stop. Guards block for a reason. 4. **ALWAYS pass `model:` per the Stage→Model table** when spawning a stage subagent. Never rely on the inherited default. 5. **ALWAYS record the dispatch before spawning the subagent** — this preserves the G4 oscillation signal even if the subagent crashes. ## Stage→Model Dispatch Table Mirrors the PM persona's table (`config/personas/project-manager.md`) — the local equivalent of `valor-session create --model`. | Stage | Skill | `model:` | Rationale | |-------|-------|----------|-----------| | ISSUE | /do-issue | sonnet | Structured writing | |