skills.rest
antstanley
antstanleyCOMMUNITY
Explore all skills by this author →
GitHub

spec-builder

Executes a spec-planner plan by dispatching one sub-agent per task behind correctness and completeness gates.

Updated May 4, 2026
Search Tags:#task execution#git worktrees#spec-driven development#dependency graph#code verification#sub-agent orchestration#jj workspaces
One-click install
npx skills add https://github.com/antstanley/skills --skill spec-builder-antstanley
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: spec-builder
Source: https://github.com/antstanley/skills/tree/main/plugins/spec-builder/skills/spec-builder
Command: npx skills add https://github.com/antstanley/skills --skill spec-builder-antstanley

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Turning a written implementation plan into merged, verified code is error-prone: tasks get built out of order, builders grade their own work, and partially-done tasks get marked complete. This Skill executes a spec-planner plan task by task, with each task built by its own sub-agent in an isolated workspace and merged only after independent correctness and completeness verification. ## Core Features & Use Cases - Sub-agent task execution: Dispatches one implementer sub-agent per task into its own jj workspace or git worktree, walking the plan's dependency graph in parallel waves (default max 4 agents) or sequentially. - Dual verification gates: Every task must pass a semi-formal correctness review and a definition-of-done completeness validation — run by an agent other than the builder, combined into one gate by default or split into two. - Kanban plan board: Tasks move between backlog/, in-progress/, blocked/, and done/ subfolders as their status changes, so progress is ls-legible and interrupted builds resume from folder membership. - Use Case: You have a spec-planner plan folder with six dependency-ordered tasks. Ask the agent to build the plan, and it dispatches implementers in waves, gates each task, merges passing work into an integration point, and parks tasks that fail repeated verification instead of forcing a green build. ## Quick Start Build the plan in .specs/plans/2026-05-22-implement_journal_app/ using spec-builder with default parallel execution.

Frequently Asked Questions about spec-builder

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I execute a spec-planner plan with sub-agents?▼

Point the skill at the plan folder, for example "build the plan in .specs/plans/...". It reads the dependency table, dispatches one implementer sub-agent per ready task into an isolated workspace, and gates each task through correctness and completeness verification before merging.

How does spec-builder verify completed tasks?▼

Each task is verified by an agent other than its implementer. By default one combined gate returns a correctness verdict (CORRECT to BUGGY) and a completeness verdict (DONE, PARTIAL, NOT_DONE) over a single reading of the diff; gate_mode: split runs them as two separate agents.

Does spec-builder work in a plain git repo without jj?▼

Yes. Backend detection checks for jj first and falls back to git worktrees, using a spec-builder/integration branch as the integration point and one worktree per task. All orchestration, gating, and status bookkeeping are identical on both backends.

Can spec-builder run without sub-agent dispatch support?▼

Yes, via a documented sequential single-agent fallback that builds tasks one at a time in-context while still running the verification pass. The skill states plainly that this mode ran, since same-session verification is weaker than a separate agent grading the work.

What happens when a task fails its verification gate?▼

The implementer is re-dispatched in the same workspace with the verdict as feedback, bounded by a default of 2 retries. Past that bound the task is parked in blocked/ with a reason line and surfaced to the user rather than forced to pass.

When should I not use spec-builder?▼

Skip it when there is no plan (use spec-planner first, since building straight from a spec is not reviewable) or when the change is a single trivial task, where spinning up a workspace and verification gate costs more than building directly.

Related Skills

Explore More Software Engineering Skills→
OzeroHAXOzeroHAX
32

project-fast-task-blast

Decompose proto-specifications into ordered MVP tasks with acceptance criteria and dependencies.

Community
Advanced
richard-jh-mccraerichard-jh-mccrae

code-review

Reviews code changes since a reference point against standards and spec requirements.

Community
Intermediate
sourcelabbgsourcelabbg

hlr-55-certificate-transparency

Enforce EUDI Certificate Transparency requirements for access certificates and SCT verification.

Official
Basic
mrjuguymrjuguy

spawn-child-agent

Spawn independent Claude Code agent processes for isolated sub-tasks.

Community
Advanced