Agent skills for VS Code, Google Antigravity, and Claude Code that turn a vague idea into a confirmed spec and a chunked, executable implementation plan — in as few as 3 exchanges.
/ame → single compiled interview → .ame/spec.md
/ema → layer-based plan → .ame/plan.md → chunk-by-chunk execution
The biggest source of rework in software is starting to code before the requirements are actually understood — by the developer or the AI. These two skills enforce a spec-before-code discipline without slowing you down.
/ame closes the knowledge gap in one compiled interview message — all relevant questions at once, answered once, written to spec once. No 6-round back-and-forth.
/ema reads that spec, identifies which architectural layers the project spans, generates a dependency-ordered plan, and asks for confirmation before touching anything.
- An AI coding IDE that supports agent/agentic mode with file-write capability (see Install section)
- Optional: context7 MCP server for live library documentation during the interview
The skills use the open agent skills standard — a SKILL.md file in a named folder. The same files work across VS Code, Antigravity, and any other IDE that supports the standard. Claude Code uses a different installation method documented below.
Copy the skill files to your VS Code agent skills folder:
Windows:
%USERPROFILE%\.agents\skills\ame\SKILL.md ← from skills/ame/SKILL.md
%USERPROFILE%\.agents\skills\ema\SKILL.md ← from skills/ema/SKILL.md
macOS / Linux:
~/.agents/skills/ame/SKILL.md
~/.agents/skills/ema/SKILL.md
Then restart VS Code. Both skills require Agent mode — Ask mode cannot write files.
Copy the skill files to Antigravity's global skills folder:
Windows:
%USERPROFILE%\.gemini\antigravity\skills\ame\SKILL.md
%USERPROFILE%\.gemini\antigravity\skills\ema\SKILL.md
macOS / Linux:
~/.gemini/antigravity/skills/ame/SKILL.md
~/.gemini/antigravity/skills/ema/SKILL.md
Antigravity's agent is always in agentic mode — no extra configuration required. Skills are auto-discovered and invoked by description matching.
Claude Code uses a custom slash commands directory instead of the skills standard.
Global (all projects):
~/.claude/commands/ame.md ← copy contents of skills/ame/SKILL.md
~/.claude/commands/ema.md ← copy contents of skills/ema/SKILL.md
Project-level only:
.claude/commands/ame.md
.claude/commands/ema.md
After copying, /ame and /ema will appear as slash commands in Claude Code. Claude Code runs in agentic mode by default, so file writing works without extra configuration.
npx skills add github:CSKishan/ame-skillYou can also install skills at the project level — useful for team repos where everyone should have access:
<project-root>/.agents/skills/ame/SKILL.md
<project-root>/.agents/skills/ema/SKILL.md
# VS Code (Agent mode) / Google Antigravity / Claude Code:
/ame I want to build a patient medication tracker for a hospital tablet app
# Answer the interview. AME writes .ame/spec.md.
# When done:
/ema
# EMA reads the spec, writes .ame/plan.md, and shows a chunk summary.
# Reply "yes" to execute chunk by chunk, or "run all" to execute everything at once.
All three IDEs support agentic file writing, so both skills work identically regardless of which IDE you use.
| Scope | Requests |
|---|---|
micro (single file / script) |
2 — questions, then confirm |
small (<5 files, no auth/compliance) |
2–3 — questions, answer, confirm |
full (auth, integrations, compliance) |
2–3 — questions, answer, confirm |
The old sequential model used 7+ requests for full scope. The batch interview model compresses this to 2–3 regardless of scope.
| Action | Requests |
|---|---|
| Generate plan | 1 |
| Execute each chunk (with per-chunk confirmation) | 1 per chunk |
| Execute all chunks ("run all") | 1 for all + 1 final summary |
AME automatically estimates scope from your opening description:
| Signal | Scope | Dimensions asked |
|---|---|---|
| Single file, script, formatter, quick fix | micro |
Edge cases only (3 questions) |
| <5 files, no external systems, no auth | small |
Tech stack + Features + Edge cases |
| Multiple systems, auth, integrations, compliance, or unknown | full |
All 5 dimensions |
You can correct the scope estimate before answering.
| Dimension | Covers |
|---|---|
| [A] Tech Stack | Language, runtime, framework, platform, team |
| [B] Features & Architecture | Core features, actors, data flow, patterns |
| [C] Security & Compliance | Auth, data sensitivity, regulations, trust boundaries (full scope only) |
| [D] Quality & Operations | Accessibility, performance, testing, CI/CD (full scope only) |
| [E] Edge Cases & Gaps | Failure modes, degraded state, assumptions, open questions |
EMA chunks the work by architectural layer dependency:
| Layer | Standard | Embedded / Firmware | Data Pipeline |
|---|---|---|---|
| 0 — Foundation | Models, auth, config | BSP, HAL, MCU init | Source connectors, schema |
| 1 — Core Logic | Business logic, APIs, services | RTOS tasks, drivers, protocols | Transform, enrichment, validation |
| 2 — Delivery / IO | UI, integrations, wiring | Application, host comms | Sinks, serving layer |
Layer 2 is absent for headless backends, pure CLI tools, standalone firmware, and data pipelines where sinks are core logic.
AME writes and maintains this file. It lives in your project root and is the contract between AME and EMA.
Commit it. It documents every decision made before coding began — invaluable for onboarding, audits, and revisiting requirements.
A blank template lives at .ame/spec-template.md.
EMA writes this file. It contains:
- Plan risks (LOW-confidence spec dimensions)
- Project summary and stack
- Per-chunk breakdown: files, implementation steps, security/accessibility checklists, validation gate, rollback instructions
- Post-execution checklist
Commit this alongside your code for traceability.
| Trigger | Behaviour |
|---|---|
/ame with a description |
Starts interview immediately from your description |
/ame without a description |
AME asks for a one-sentence intent |
"done" / "plan it" / "proceed" during /ame |
Finalises spec and shows handoff message |
/ema during /ame |
AME finalises spec, EMA takes over immediately |
"review" during /ema |
Shows full .ame/plan.md before execution |
"run all" during /ema |
Executes all chunks back-to-back; pauses only on validation failure or when a gate item requires manual confirmation |
"adjust {description}" during /ema |
Updates plan before executing |
AME uses the context7 MCP server to fetch live library documentation when you name a specific framework in your opening description. Context7 works wherever MCP is supported — VS Code, Antigravity, and Claude Code all support MCP servers.
- Fires once, before the interview questions are compiled
- Capped at 3 call-pairs per session
- If context7 is unavailable, AME logs it and continues — the interview is never blocked
VS Code — add to your VS Code MCP configuration:
"context7": {
"command": "npx",
"args": ["-y", "@upstash/context7-mcp@latest"],
"type": "stdio"
}Antigravity — add to ~/.gemini/antigravity/mcp_servers.json (or via Antigravity's MCP settings panel).
Claude Code — add to ~/.claude/claude_desktop_config.json under mcpServers:
"context7": {
"command": "npx",
"args": ["-y", "@upstash/context7-mcp@latest"]
}| IDE | Install method | Agent / file-write mode |
|---|---|---|
| VS Code (GitHub Copilot) | ~/.agents/skills/ global or .agents/skills/ per-project |
Agent mode (must be enabled) |
| Google Antigravity | ~/.gemini/antigravity/skills/ global or .agents/skills/ per-project |
Always agentic |
| Claude Code | ~/.claude/commands/ global or .claude/commands/ per-project |
Always agentic |
Any IDE that supports the open agent skills standard and can write files to the workspace will work.
Built to work across:
- Web and mobile applications (React, Vue, Flutter, React Native, etc.)
- Backend services and REST / GraphQL APIs
- Embedded firmware and RTOS-based systems (STM32, ESP32, FreeRTOS, Zephyr)
- Data pipelines and ETL systems
- CLI tools and developer utilities
- Industrial, medical, and safety-critical software
Yes, but EMA will flag the missing spec as a plan risk and note that accuracy is low. For best results, always run /ame first.
Skip any that don't apply. AME marks unanswered fields as [NOT PROVIDED] and flags them as assumptions. You can correct anything significant in the Section [6] confirmation step.
Yes. EMA reads whatever is there. Preserve the section headers ([A]–[E]) so EMA can parse them correctly.
Correct it before answering the interview. AME accepts any correction and recompiles the right set of questions.
Fully. Context7 is optional. AME logs when it's unavailable and continues the interview without it.
Pull requests are welcome. Please keep skill changes backward-compatible with the .ame/spec-template.md format — EMA's parser depends on consistent section headers ([A]–[E] and [6]).
When changing the spec format:
- Update
spec-template.md - Update AME's Step 2 spec write template
- Update EMA's Step 1 spec read references
MIT