Agent skill
specs-dev
Plan-first development workflow with review gates. Use when implementing features, refactoring, or any task requiring structured planning, iterative implementation with reviews, and clean commits. Triggers on requests like "implement feature X", "plan and build", "spec-driven development", or when user wants disciplined, reviewed code changes.
Install this agent skill to your Project
npx add-skill https://github.com/vaayne/agent-kit/tree/main/skills/specs-dev
SKILL.md
Specs-Dev Workflow
A disciplined, review-gated development workflow: plan first, implement in phases, review between phases.
Workflow
| Phase | Purpose | Gate |
|---|---|---|
| 1. Discovery | Understand requirements | User approves summary |
| 2. Planning | Create reviewed plan | User approves plan |
| 3. Implementation | Phase-by-phase coding | Each phase reviewed |
Phase 1: Discovery
Goal: Shared understanding before planning.
- State your understanding of the request
- Ask clarifying questions (goals, constraints, success criteria, out-of-scope)
- Iterate until clear
- Present final requirements summary
Gate: "Do I understand correctly? Should I proceed to planning?" — Wait for approval.
Phase 2: Planning
Goal: Create a comprehensive, reviewed implementation plan.
- Draft plan using
references/templates/plan.md - Review with reviewer subagent (see
references/agents/reviewer.md) — max 3 rounds - Integrate feedback, iterate until approved
- If still not approved after 3 rounds — summarize unresolved concerns and ask user whether to revise scope, continue anyway, or stop
- Resolve all Open Questions — convert remaining unknowns into explicit assumptions before proceeding
- Present plan to user — wait for approval
- Create session folder:
.agents/sessions/{YYYY-MM-DD}-{feature-name}/ - Save
plan.md,tasks.md, andhandoff.md(initialized fromreferences/templates/handoff.md) to session folder
Plan quality check:
- Every requirement from Phase 1 addressed
- Tasks are actionable and logically ordered
- Testing strategy specified
- Risks captured
- No unresolved Open Questions (converted to assumptions or removed)
Phase 3: Implementation
Goal: Implement the plan phase-by-phase with reviews between phases.
Read
references/loop.mdbefore starting. It defines the phase loop.
Summary of the loop:
For each phase in the plan:
1. Spawn worker subagent → implements all tasks, commits each one
2. Worker writes phase summary to handoff.md
3. Spawn reviewer subagent → reads handoff.md, reviews the phase
4. If changes needed → fix in main agent context
5. Continue to next phase
After the last phase: run full test suite, update session docs, confirm with user.
Subagents:
- Worker:
references/agents/worker.md - Reviewer:
references/agents/reviewer.md - Handoff protocol:
references/handoff.md
Session Structure
.agents/sessions/{YYYY-MM-DD}-{feature-name}/
├── plan.md # Strategic plan
├── tasks.md # Task checklist
└── handoff.md # Phase handoff log (appended each phase)
References
| File | When to Read |
|---|---|
references/loop.md |
Start of Phase 3 |
references/handoff.md |
Start of Phase 3 |
references/agents/reviewer.md |
Plan/phase reviews |
references/agents/worker.md |
Phase implementation |
references/templates/ |
Phase 2 (plan creation) |
Recommended Agent Skills
Expand your agent's capabilities with these related and highly-rated skills.
pi-delegate
Delegate tasks to Pi subagents with preset roles (oracle, worker, reviewer, ui-engineer, librarian) or ad-hoc prompts. Use for fresh context, model specialization, second opinions, code review, isolated execution, or parallel task delegation.
mcp-jetbrains-ide
Control JetBrains IDE (IntelliJ, WebStorm, PyCharm) via MCP. Use when manipulating IDE files, running configurations, searching code, or performing refactoring. Triggers on "open in IDE", "run configuration", "refactor code", "IDE search", "JetBrains".
mcp-context7-docs
Query up-to-date documentation and code examples for any programming library or framework. Use when looking up API docs, finding code examples, or checking library usage. Triggers on "how to use [library]", "docs for [package]", "show me examples of [framework]", "Context7 lookup".
web-fetch
Fetch and extract clean content from URLs using Jina Reader API. Use when users need to read webpage content, extract article text, or fetch URL content for analysis. Triggers on "fetch this page", "read this URL", "extract content from", "get the content of", "what does this page say".
tmux
Control interactive terminal sessions via tmux. Use when tasks need persistent REPLs, parallel CLI agents, or any process requiring a TTY that simple shell execution cannot handle.
mcp-grep-code
Search real-world code examples from over a million public GitHub repositories. Use when finding code patterns, implementation examples, or how libraries are used in practice. Triggers on "find code examples", "how is [library] used", "search GitHub code", "grep.app search", "code pattern".
Didn't find tool you were looking for?