Agent Automation Recommender

Agent Automation Recommender

Analyze codebase patterns to recommend tailored AI Agent automations across 6 universal extensibility dimensions, customized for Claude Code, Antigravity (Gemini), Codex, or generic Agent environments.

This skill is read-only. It analyzes the codebase and outputs structured recommendations. It does NOT create or modify any files. Users implement the recommendations themselves or ask the agent separately to help build them.

Decision Principles

Use the smallest effective automation, backed by repository evidence.

1. Prefer existing mechanisms: inspect package scripts, CI, tests, configuration, and native Agent features before proposing a new skill, MCP server, plugin, or dependency.
2. Route by enforcement need:
   • stable intent, rationale, or convention -> Rules & Knowledge;
   • mechanically checkable invariant -> tests or CI;
   • fast lifecycle feedback -> Hooks & Guardrails;
   • repeated multi-step workflow -> Skills & Tools;
   • live external data -> MCP Servers;
   • independent focused work -> Subagents & Roles;
   • a bundle -> Plugins & Packages only when several required capabilities share setup.
3. Recommend less: return at most 5 total items, ranked by value, effort, and risk. Omit dimensions with no concrete problem or evidence.
4. Keep memory exceptional: recommend persistent memory only for information that cannot be recovered from code or maintained documentation, and state an audit or expiry plan. Prefer curated project rules for stable constraints.
5. State operational impact: label each item as [read-only], [local write], [external write], or [requires confirmation]. Recommendations must not imply installation, production access, or external mutation without explicit confirmation.

Output Guidelines

• Recommendation budget: Recommend at most 5 items in total, normally no more than 1 per dimension. Omit a dimension when the codebase provides no evidence of a concrete need.
• Agent Context Adaptation: Tailor file paths, syntax, and configuration snippets directly to the target Agent environment (Claude Code, Antigravity, Codex).
• Evidence first: Use web search only when available and when repository evidence leaves a current-fact gap; identify the source and mark uncertainty when it remains.
• Verification loop: Every recommendation must include the expected outcome, implementation cost or maintenance burden, risk or side effect, a minimal verification check, and a disable/rollback path.
• Interactive Follow-up: Inform users they can request additional recommendations or assistance implementing any specific item.

6 Universal Extensibility Dimensions

Workflow

Phase 1: Codebase & Agent Context Analysis

1. Detect Target Agent Environment:

# Check for agent signature files
ls -d .claude .gemini AGENTS.md CLAUDE.md .codexrules 2>/dev/null

• If .claude/ or CLAUDE.md exists -> Target: Claude Code
• If .gemini/ exists -> Target: Antigravity (Gemini)
• If AGENTS.md or .codexrules exists -> Target: Codex
• If explicit user prompt specifies an agent (e.g. "recommend for Antigravity") -> Target: Specified Agent
• Otherwise -> Target: Universal Multi-Agent

2. Detect Codebase Tech Stack & Signals:

# Detect project type and package manifests
ls -la package.json pyproject.toml Cargo.toml go.mod pom.xml 2>/dev/null
cat package.json 2>/dev/null | head -50

# Check key libraries and frameworks
cat package.json 2>/dev/null | grep -E '"(react|vue|angular|next|express|fastapi|django|prisma|supabase|convex|stripe)"'

# Inspect project directory structure & tests
ls -la src/ app/ lib/ tests/ components/ pages/ api/ 2>/dev/null

Use read-only, platform-appropriate commands. On Windows, use PowerShell equivalents. Confirm a signal by checking actual scripts, CI, and configuration; a marker file alone is not proof that a tool is in use.

Phase 2: Generate Recommendations Across 6 Dimensions

Read only the reference guides relevant to dimensions supported by concrete evidence; do not load every reference by default:

• references/agent-mappings.md for platform-specific syntax & paths
• references/mcp-servers.md for MCP server patterns
• references/skills-reference.md for Skills & Custom Tools
• references/hooks-patterns.md for Hooks & Guardrails
• references/subagent-templates.md for Subagents & Roles
• references/rules-and-knowledge.md for Rules & Knowledge Items
• references/plugins-reference.md for Plugins & Extension Bundles

Phase 3: Output Recommendations Report

Format the final report clearly with code blocks tailored to the detected Target Agent:

Every proposed item must state: Evidence, Why this is the smallest fit, Impact / effort / risk, Verification, and Disable / rollback. Start with a short decision summary naming the top choices, existing mechanisms reused, and options deliberately skipped.

## AI Agent Automation Recommendations

I've analyzed your codebase and identified the top automations across all 6 dimensions.

### 🔍 Codebase & Agent Profile
- **Target Agent Context**: [Claude Code / Antigravity / Codex / Universal]
- **Language / Runtime**: [detected language/runtime]
- **Framework & Libraries**: [detected framework & key packages]

---

### 🔌 1. MCP Servers
#### [MCP Name]
- **Evidence**: [Observed dependency, script, workflow, or explicit need]
- **Why**: [Specific reason based on dependencies]
- **Why this is the smallest fit**: [Why existing tools or a simpler layer are insufficient]
- **Impact / effort / risk**: [[read-only] / [local write] / [external write] / [requires confirmation]; maintenance and security notes]
- **Verification**: [Minimal check that demonstrates the expected outcome]
- **Disable / rollback**: [How to remove or turn it off]
- **Config Syntax**: [Target Agent specific installation / config snippet]

---

### 🎯 2. Skills & Custom Tools
#### [Skill Name]
- **Why**: [Specific workflow value]
- **Location**: [Target Agent path, e.g. .claude/skills/[name]/SKILL.md or .gemini/config/plugins/...]
- **Structure / Syntax**:
```yaml
---
name: [skill-name]
description: [description]
---

⚡ 3. Hooks & Guardrails

[Hook Name]

• Why: [Reasoning based on Prettier, ESLint, pytest, or sensitive files]
• Target File: [e.g. .claude/settings.json, RULE[...] or .git/hooks/pre-commit]

🤖 4. Subagents & Roles

[Subagent Name]

• Why: [Parallel analysis or review needs]
• Location: [e.g. .claude/agents/[name].md or .gemini/config/plugins/.../agents/]

📜 5. Rules & Project Knowledge

[Rule / KI Name]

• Why: [Architecture, security, or domain context needs]
• Target Location: [CLAUDE.md / .gemini/ (RULE/KI) / AGENTS.md]

📦 6. Plugins & Packages

[Plugin Name]

• Why: [Workflow bundle value]
• Installation: [Command or path instructions]

Next Steps:

• Ask for additional recommendations in any category (e.g. "show more MCP servers").
• Ask for direct assistance setting up or creating any of the recommended items above!