Stage: Implement
Canonical lifecycle stage: write the code that satisfies the plan — including building it, testing it into existence, and fixing it. Framework-neutral name (was stage-execute, GSD’s term) in standard SDLC vocabulary. Characteristically parallelized and context-isolated to avoid degradation at scale.
Derived projection — evidence is the capabilities that implements: [[stage-implement]].
Implemented by (backlinks)
GSD:
- gsd-execute-phase — wave-based parallel execution (pattern-wave-parallelism).
- gsd-executor — one atomic commit per task → artifact-atomic-commit; sole Edit-capable agent.
- gsd-debugger — evidence-tracking bug investigation (pattern-systematic-debugging).
Matt Pocock — Skills for Real Engineers:
- mp-implement — build the work from a spec/tickets, TDD at seams, then code-review → artifact-atomic-commit (v1.1; MP’s first first-class execute skill).
- mp-tdd — test-first red → green loop (v1.1: refactor moved to mp-code-review) (pattern-test-driven-development).
- mp-prototype — throwaway prototypes (pattern-throwaway-prototype).
- mp-diagnosing-bugs — structured debugging loop (pattern-systematic-debugging).
- mp-resolving-merge-conflicts — resolve merge/rebase conflicts from each side’s primary-source intent (v1.1; pattern-trunk-based-development).
Addy Osmani — Agent Skills (its whole Build phase, plus debugging which it files under Verify):
- addy-incremental-implementation — thin vertical slices → artifact-atomic-commit (pattern-vertical-slice, pattern-feature-flags).
- addy-tdd — red-green-refactor; Prove-It for bugs (pattern-test-driven-development).
- addy-debugging — five-step triage: reproduce, localize, reduce, fix, guard (pattern-systematic-debugging).
- addy-api-design — contract-first interfaces (pattern-contract-first, pattern-deep-modules).
- addy-frontend-ui — production-quality UI + WCAG 2.1 AA accessibility.
- addy-context-engineering — right context at the right time (pattern-context-engineering).
- addy-source-driven-development — cite official docs (pattern-source-grounding).
- addy-doubt-driven-development — in-flight adversarial review (pattern-adversarial-review).
OpenSpec:
- openspec-apply — walk the
tasks.mdchecklist sequentially, resuming from checkpoints.
Spec Kit:
- speckit-implement — execute
tasks.md, with TDD mandated by the constitution (tests fail before code) (pattern-test-driven-development).
BMAD:
- bmad-dev — the Implementation persona (Amelia); absorbed the classic SM + QA roles.
- bmad-dev-story — implement one fully-contexted story test-first in a fresh context → artifact-atomic-commit (pattern-test-driven-development, pattern-context-engineering, pattern-fresh-context-subagents).
- bmad-quick-dev — Quick-Flow fast path for small work (pattern-scale-adaptive-planning).
(The story it executes is prepared upstream by bmad-create-story in stage-plan — the plan→implement bridge.)
Compound Engineering:
- ce-work — execute the guardrailed plan (HOW) inside an isolated worktree, simulating real usage via MCP (pattern-worktree-isolation, pattern-vertical-slice) → artifact-atomic-commit.
- ce-worktree — the git-worktree isolation the work runs in (pattern-worktree-isolation).
- ce-debug — causal-chain + prediction-driven root-cause (pattern-systematic-debugging).
- ce-polish — conversational UX iteration against a live dev server.
gstack (build + debugging + the guardrails/session infrastructure that make autonomous building safe):
- gstack-design-shotgun — explore 4-6 AI mockup variants with taste memory → artifact-design-mockup (pattern-throwaway-prototype).
- gstack-design-html — turn an approved mockup into production Pretext-native HTML/CSS.
- gstack-investigate — Iron-Law root-cause debugging (no fix without investigation; auto-freeze) (pattern-systematic-debugging); debugging cluster with gsd-debugger / mp-diagnosing-bugs / addy-debugging / ce-debug.
- gstack-scrape — browser data-extraction automation (compounds via gstack-skillify).
- gstack-careful · gstack-freeze · gstack-guard · gstack-unfreeze — safety guardrails / directory-scoped edit locks (pattern-edit-guardrails).
- gstack-context-save · gstack-context-restore — save/restore working context + WIP checkpoints (pattern-session-handoff).
- gstack-pair-agent — cross-vendor multi-agent collaboration through a shared browser (novel; no counterpart).
Builder Methods — BM Skills:
- bm-design-system — scaffold a React + Tailwind v4 design system into the codebase (runnable components + a live reference page +
AGENTS.md/CLAUDE.mdguardrails → artifact-design-md); counterpart to addy-frontend-ui (pattern-context-engineering).
Superpowers (the densest implement roster here — execution engines + the TDD/debug/isolation/parallelism disciplines that run inside them):
- sp-subagent-driven-development — fresh implementer subagent per task + two-stage task review + final whole-branch review → artifact-atomic-commit (pattern-fresh-context-subagents, pattern-adversarial-review).
- sp-executing-plans — the parallel-session executor: batch execution with human checkpoints (subagent-free fallback) → artifact-atomic-commit (pattern-test-driven-development).
- sp-test-driven-development — RED-GREEN-REFACTOR; Iron Law “no production code without a failing test first” (pattern-test-driven-development).
- sp-systematic-debugging — four-phase root-cause process; Iron Law “no fixes without investigation”; question the architecture after 3 failed fixes (pattern-systematic-debugging).
- sp-using-git-worktrees — ensure an isolated workspace (native-tool-first, git fallback, clean-baseline check) (pattern-worktree-isolation).
- sp-dispatching-parallel-agents — one focused subagent per independent problem domain, dispatched concurrently (pattern-fresh-context-subagents, pattern-wave-parallelism).
Cross-framework equivalents
The capability-level clusters landing here now span all eight frameworks:
- Debugging: GSD’s gsd-debugger ↔ Matt’s mp-diagnosing-bugs ↔ Addy’s addy-debugging ↔ Compound Engineering’s ce-debug ↔ gstack’s gstack-investigate ↔ Superpowers’ sp-systematic-debugging (
equivalent_to), all applying pattern-systematic-debugging. (Addy categorizes debugging under its Verify phase, but it is the same activity, so it clusters here; CE adds an explicit prediction-confirmation step; gstack and Superpowers share an Iron-Law “no fix without investigation” and a stop-after-3-fixes rule.) - Test-first: Matt’s mp-tdd ↔ Addy’s addy-tdd ↔ GSD’s gsd-plan-phase
--tdd↔ Spec Kit’s speckit-implement (TDD “NON-NEGOTIABLE”) ↔ BMAD’s bmad-dev-story (Amelia is “disciplined in Kent Beck’s TDD”) ↔ Superpowers’ sp-test-driven-development, clustered at pattern-test-driven-development — Spec Kit mandates it via the artifact-constitution; Superpowers matches its non-negotiability through rhetoric (delete-and-restart any code written before its test). - Execute loop (now eight frameworks): OpenSpec’s openspec-apply ↔ Addy’s addy-incremental-implementation ↔ GSD’s gsd-execute-phase ↔ Spec Kit’s speckit-implement ↔ BMAD’s bmad-dev-story ↔ Compound Engineering’s ce-work ↔ Matt Pocock’s mp-implement ↔ Superpowers’ sp-executing-plans / sp-subagent-driven-development — walk the plan/tickets unit-by-unit. They differ in mechanics: GSD parallelizes into pattern-wave-parallelism waves, Addy cuts thin pattern-vertical-slices, OpenSpec runs a plain resumable sequential pass, Spec Kit drives test-first per the constitution, BMAD drives from a single fully-contexted story per fresh context, CE bounds execution with plan guardrails in worktree isolation, MP drives mp-tdd at pre-agreed seams, and Superpowers offers two modes — continuous same-session fresh-subagent-per-task with a two-stage review (sp-subagent-driven-development, clustering with gsd-execute-phase / bmad-dev-story) or checkpointed batch execution (sp-executing-plans).
- Worktree isolation: Compound Engineering’s ce-worktree ↔ Superpowers’ sp-using-git-worktrees — both make an isolated git worktree a first-class execution step (pattern-worktree-isolation).
- Interface design: Addy’s addy-api-design ↔ Matt’s mp-codebase-design — deep modules / stable contracts (Spec Kit authors its
contracts/upstream in speckit-plan).
See Also
- stage-plan — produces the plan executed here.
- stage-validate — confirms the result works; stage-review — confirms it’s good.