Custom prompts, skills, and workflows for AI coding agents. Provides consistent developer workflows across multiple coding assistants.
| Workflow | Description |
|---|---|
| refactor | Analyze codebase for refactoring opportunities, prioritize by severity/effort |
| review | Review uncommitted changes for bugs, regressions, and improvements |
| pr-reviewer | Fetch PR comments, summarize issues, address them, update PR |
| create-pr | Create PR with auto-generated title and description |
| deslop | Analyze code for quality issues using established software engineering principles |
This repository includes a sophisticated agent architecture for OpenCode using GPT-5.3-Codex as the orchestrator and Fireworks Kimi K2.5 Turbo as specialized subagents.
┌─────────────────────────────────────────────────────────────┐
│ Codex53-Kimi Orchestrator (GPT-5.3-Codex) │
│ • Plans and sequences work │
│ • Makes routing decisions │
│ • Delegates to specialized subagents │
└──────────────────────┬────────────────────────────────────────┘
│
┌───────────┬──────────────┬──────────────────┬──────────────┐
│ │ │ │ │
┌──▼────────┐ ┌▼───────────┐ ┌▼────────────────┐ ┌▼────────────┐
│ kimi- │ │ kimi- │ │ github- │ │ docs- │
│ general │ │ explore │ │ librarian │ │ research │
│ execution │ │ local find │ │ remote GitHub │ │ official │
└───────────┘ └────────────┘ └─────────────────┘ └─────────────┘
│ │ │ │
┌──▼────────┐ ┌────▼─────┐ ┌────────▼──────┐ ┌─────────▼──────┐
│walkthrough│ │ review │ │ deslop │ │ pr-reviewer │
│ diagrams │ │ local QA │ │ code quality │ │ PR fixes │
└───────────┘ └──────────┘ └───────────────┘ └────────────────┘
│ │
┌─────▼─────┐ ┌────▼────┐
│ create-pr │ │ oracle │
│ workflow │ │ deep │
└───────────┘ │ reasoning│
└─────────┘
The orchestrator uses keyword-based deterministic routing:
| Trigger | Keywords | Agent Selected |
|---|---|---|
| PR creation | "create PR", "pull request" | create-pr |
| PR feedback | "PR comment", "address PR" | pr-reviewer |
| Code audit | "audit", "code quality" | deslop |
| Local review | "review changes", "uncommitted" | review |
| Remote repo research | GitHub URL, owner/repo, "reference implementation" |
github-librarian |
| Official docs research | "official docs", "migration guide", "API docs" | docs-research |
| Local walkthrough | "walk me through", "diagram", "architecture" | walkthrough |
| Local discovery | "find", "search", "explore" | kimi-explore |
| Implementation | "implement", "fix", "refactor" | kimi-general |
For implementation/debugging/refactoring tasks, the orchestrator follows:
- PHASE 0 (optional): docs-research / github-librarian → Gather official docs or upstream reference code first
- PHASE 1: spec-compiler → Compile Execution Contract (scope, risks, success criteria)
- PHASE 2: kimi-general → Execute implementation based on contract
- PHASE 3: quick-validator → Run validation tests/checks
- PHASE 4 (optional): change-auditor → Deep audit for high-risk areas
~/.config/opencode/
├── opencode.json # Main configuration (see example)
├── bin/ # Helper scripts (e.g. opencode-gh-librarian)
├── agent/ # Primary agent definitions
│ ├── codex53-kimi.md # Orchestrator (routing logic)
│ ├── kimi-general.md # Execution worker
│ ├── kimi-explore.md # Local read-only discovery
│ ├── github-librarian.md # Remote GitHub research
│ ├── docs-research.md # Official docs + API research
│ ├── walkthrough.md # Architecture walkthroughs + diagrams
│ └── oracle.md # Deep reasoning (GPT-5.4)
└── commands/ # Workflow prompts
├── review.md
├── deslop.md
├── pr-reviewer.md
├── create-pr.md
├── spec-compiler.md
├── quick-validator.md
└── change-auditor.md
NEVER commit your opencode.json with real API keys to version control.
- Use environment variables:
OPENCODE_FIREWORKS_API_KEY - Or keep the config file in a secure location with
apiKey: "YOUR_KEY_HERE" - The example file includes placeholder warnings to help prevent accidental commits
A local reference setup uses:
- Fireworks Kimi K2.5 Turbo for build mode and subagents
- GPT-5.3-Codex for the orchestrator (plan mode)
- Specialized subagents for local discovery, remote GitHub research, docs research, architecture walkthroughs, and execution
See prompts/opencode/opencode.json.example for the full configuration structure.
| Agent | Type | Prompt Location |
|---|---|---|
| Claude Code | CLI | ~/.claude/commands/ |
| Codex | CLI | ~/.codex/skills/ |
| OpenCode | CLI | ~/.config/opencode/commands/ |
| Antigravity | Editor | ~/.antigravity/prompts/ or ~/Library/.../Antigravity/User/prompts/ |
| VSCode Copilot | Editor | ~/.config/Code/User/prompts/ or ~/Library/.../Code/User/prompts/ |
| Copilot CLI | CLI | ~/.copilot/agents/ |
| Amp | CLI/Editor | ~/.config/agents/skills/ |
| Gemini CLI | CLI | ~/.gemini/GEMINI.md |
| Kilo Code | CLI | ~/.kilocode/prompts/ |
| Cursor | Editor | ~/.cursor/commands/ |
| Cline | Editor | ~/Documents/Cline/Rules/ |
| Roo Code | Editor | ~/.roo/commands/ |
| Windsurf | Editor | ~/.codeium/windsurf/memories/global_rules.md |
git clone https://github.com/YOUR_USERNAME/agent-tools.git
cd agent-tools
./scripts/install.sh./scripts/install.sh claude codex ampAvailable options: claude, codex, opencode, antigravity, vscode, copilot-cli, amp, gemini, kilocode, cursor, cline, roocode, windsurf
Claude Code
mkdir -p ~/.claude/commands
cp prompts/claude/*.md ~/.claude/commands/Codex
mkdir -p ~/.codex/skills
cp prompts/codex/*.md ~/.codex/skills/OpenCode (Codex53-Kimi Architecture)
# Create directories
mkdir -p ~/.config/opencode/commands
mkdir -p ~/.config/opencode/agent
mkdir -p ~/.config/opencode/bin
# Copy workflow prompts to commands/
for f in prompts/opencode/review.md prompts/opencode/deslop.md \
prompts/opencode/pr-reviewer.md prompts/opencode/create-pr.md \
prompts/opencode/spec-compiler.md prompts/opencode/quick-validator.md \
prompts/opencode/change-auditor.md; do
[ -f "$f" ] && cp "$f" ~/.config/opencode/commands/
done
# Copy agent definitions to agent/
for f in prompts/opencode/codex53-kimi.md prompts/opencode/kimi-general.md \
prompts/opencode/kimi-explore.md prompts/opencode/github-librarian.md \
prompts/opencode/docs-research.md prompts/opencode/walkthrough.md \
prompts/opencode/oracle.md; do
[ -f "$f" ] && cp "$f" ~/.config/opencode/agent/
done
# Copy helper scripts used by subagents
cp prompts/opencode/bin/* ~/.config/opencode/bin/
chmod +x ~/.config/opencode/bin/*
# Copy and edit config (⚠️ NEVER commit with real API key)
cp prompts/opencode/opencode.json.example ~/.config/opencode/opencode.json~/.config/opencode/opencode.json and replace YOUR_FIREWORKS_API_KEY_HERE with your actual API key. Do not commit this file.
Note: github-librarian requires gh to be installed and authenticated. docs-research works best when websearch is available, which OpenCode enables when using the OpenCode provider or when OPENCODE_ENABLE_EXA=1 is set.
See Opencode Codex53-Kimi Setup for architecture details.
Antigravity
mkdir -p ~/Library/Application\ Support/Antigravity/User/prompts
cp prompts/antigravity/*.md ~/Library/Application\ Support/Antigravity/User/prompts/VSCode Copilot
# VSCode Insiders
mkdir -p ~/Library/Application\ Support/Code\ -\ Insiders/User/prompts
cp prompts/vscode-copilot/*.md ~/Library/Application\ Support/Code\ -\ Insiders/User/prompts/
# VSCode Regular
mkdir -p ~/Library/Application\ Support/Code/User/prompts
cp prompts/vscode-copilot/*.md ~/Library/Application\ Support/Code/User/prompts/Copilot CLI
mkdir -p ~/.copilot/agents
cp prompts/copilot-cli/*.md ~/.copilot/agents/
# Rename files to .agent.md format
for f in ~/.copilot/agents/*.md; do mv "$f" "${f%.md}.agent.md"; doneAmp
mkdir -p ~/.config/agents/skills
cp -r prompts/amp/* ~/.config/agents/skills/Gemini CLI
cp prompts/gemini/GEMINI.md ~/.gemini/Kilo Code
mkdir -p ~/.kilocode/prompts
cp prompts/kilocode/*.md ~/.kilocode/prompts/Cursor
mkdir -p ~/.cursor/commands
cp prompts/cursor/*.md ~/.cursor/commands/Cline
mkdir -p ~/Documents/Cline/Rules
cp prompts/cline/*.md ~/Documents/Cline/Rules/Roo Code
mkdir -p ~/.roo/commands
cp prompts/roocode/*.md ~/.roo/commands/Windsurf
mkdir -p ~/.codeium/windsurf/memories
cp prompts/windsurf/global_rules.md ~/.codeium/windsurf/memories/Claude Code
mkdir -p ~/.claude/commands
cp prompts/claude/*.md ~/.claude/commands/Codex
mkdir -p ~/.codex/skills
cp prompts/codex/*.md ~/.codex/skills/OpenCode (Codex53-Kimi Architecture)
# Create directories
mkdir -p ~/.config/opencode/commands
mkdir -p ~/.config/opencode/agent
mkdir -p ~/.config/opencode/bin
# Copy workflow prompts to commands/
for f in prompts/opencode/review.md prompts/opencode/deslop.md \
prompts/opencode/pr-reviewer.md prompts/opencode/create-pr.md \
prompts/opencode/spec-compiler.md prompts/opencode/quick-validator.md \
prompts/opencode/change-auditor.md; do
[ -f "$f" ] && cp "$f" ~/.config/opencode/commands/
done
# Copy agent definitions to agent/
for f in prompts/opencode/codex53-kimi.md prompts/opencode/kimi-general.md \
prompts/opencode/kimi-explore.md prompts/opencode/github-librarian.md \
prompts/opencode/docs-research.md prompts/opencode/walkthrough.md \
prompts/opencode/oracle.md; do
[ -f "$f" ] && cp "$f" ~/.config/opencode/agent/
done
# Copy helper scripts used by subagents
cp prompts/opencode/bin/* ~/.config/opencode/bin/
chmod +x ~/.config/opencode/bin/*
# Copy and edit config (⚠️ NEVER commit with real API key)
cp prompts/opencode/opencode.json.example ~/.config/opencode/opencode.json~/.config/opencode/opencode.json and replace YOUR_FIREWORKS_API_KEY_HERE with your actual API key. Do not commit this file.
Note: github-librarian requires gh to be installed and authenticated. docs-research works best when websearch is available, which OpenCode enables when using the OpenCode provider or when OPENCODE_ENABLE_EXA=1 is set.
See Opencode Codex53-Kimi Setup for architecture details.
Antigravity
mkdir -p ~/.antigravity/prompts
cp prompts/antigravity/*.md ~/.antigravity/prompts/VSCode Copilot
# VSCode Insiders
mkdir -p ~/.config/Code\ -\ Insiders/User/prompts
cp prompts/vscode-copilot/*.md ~/.config/Code\ -\ Insiders/User/prompts/
# VSCode Regular
mkdir -p ~/.config/Code/User/prompts
cp prompts/vscode-copilot/*.md ~/.config/Code/User/prompts/Copilot CLI
mkdir -p ~/.copilot/agents
cp prompts/copilot-cli/*.md ~/.copilot/agents/
for f in ~/.copilot/agents/*.md; do mv "$f" "${f%.md}.agent.md"; doneAmp
mkdir -p ~/.config/agents/skills
cp -r prompts/amp/* ~/.config/agents/skills/Gemini CLI
cp prompts/gemini/GEMINI.md ~/.gemini/Kilo Code
mkdir -p ~/.kilocode/prompts
cp prompts/kilocode/*.md ~/.kilocode/prompts/Cursor
mkdir -p ~/.cursor/commands
cp prompts/cursor/*.md ~/.cursor/commands/Cline
mkdir -p ~/Documents/Cline/Rules
cp prompts/cline/*.md ~/Documents/Cline/Rules/Roo Code
mkdir -p ~/.roo/commands
cp prompts/roocode/*.md ~/.roo/commands/Windsurf
mkdir -p ~/.codeium/windsurf/memories
cp prompts/windsurf/global_rules.md ~/.codeium/windsurf/memories/Analyzes your codebase for refactoring opportunities:
- Gathers context: High-churn files, TODO/FIXME markers, large files
- Detects issues: Code duplication, long functions, god classes, dead code, deep nesting
- Assesses severity: 🔴 High | 🟠 Medium | 🟡 Low
- Calculates priority:
Severity × (1/Effort) - Reports quick wins: High severity + low effort items first
Reviews uncommitted changes:
- Gets changes:
git diff HEAD, staged changes - Checks for issues: Logic errors, edge cases, type safety, security
- Checks for regressions: Breaking API changes, removed functionality
- Optimizes: Finds redundant code, duplicate logic, performance issues
- Fixes simple issues automatically (≤5 straightforward fixes)
Addresses PR review feedback:
- Fetches comments from GitHub PR via GraphQL
- Identifies reviewers: Distinguishes bots vs humans
- Summarizes issues in a table grouped by severity
- Addresses issues: Human feedback takes priority over bots
- Commits and pushes with summary of changes
Creates a PR from current changes:
- Creates feature branch if on main/master
- Generates conventional commit message
- Creates PR with auto-generated title and description
- Reports result with PR URL and summary
Analyzes code for quality issues using established software engineering principles:
- Reads target code: Files, directories, or entire codebase
- Cross-references principles: KISS, YAGNI, SOLID, DRY, Separation of Concerns, and 30+ more
- Reports violations with before/after code examples
- Groups by severity: 🔴 Critical | 🟠 Warning | 🟡 Suggestion
- Offers automatic fixes for simple issues
Based on deslop by Theta Tech AI.
- Git for version control operations
- GitHub CLI (
gh) for PR operations andgithub-librarian - OpenCode
webfetchfordocs-research - OpenCode
websearchfor bestdocs-researchdiscovery results when URLs are not provided - The respective coding agent installed and configured
In github.com/cli/cli, find where auth token resolution happensUse official docs to verify how SvelteKit remote functions work before changing our implementationWalk me through how authentication works in this repo and include a Mermaid diagramCompare our caching layer with owner/repo's implementation before proposing changesShow me recent commits affecting src/auth.ts in owner/repo
- Fork this repository
- Add or modify prompts in
prompts/<agent>/ - Test with the target agent
- Submit a pull request
MIT