AI-Driven Development
Practices, protocols, and skills for AI-driven software development
An umbrella collection of 24 skills + 1 hook that work across Claude Code, Codex CLI, OpenCode, Cursor, Gemini CLI, Antigravity, and any other agent that supports the Agent Skills standard.
Related Projects
- Vibe Stack — an opinionated AI-native development stack with one verified default per layer, covering web apps, agent apps, deployment, persistence, error tracking, and observability.
Installation
Via Skills CLI (recommended — works in 12+ agents):
# Install all 24 skills
npx skills add CodeAlive-AI/ai-driven-development
# Or pick a single skill
npx skills add CodeAlive-AI/ai-driven-development --skill prompt-engineering
Via Claude Code plugin marketplace:
/plugin marketplace add CodeAlive-AI/ai-driven-development
/plugin install ai-driven-development@ai-driven-development
# Restart Claude Code for changes to take effect
The Bash safety hook (bash-guard) ships separately — see hooks/balanced-safety-hooks/ for its install one-liner.
What's Included
Agent reflection & meta (7)
Meta-skills that let AI coding agents configure themselves. No more editing config files — just tell your agent what you need.
| Skill | What It Does |
|---|---|
| mcp-management | Install and manage MCP servers across 10+ coding agents (Claude Code, Cursor, VS Code, Gemini CLI, Codex, Goose, Copilot CLI, OpenCode, Zed) — single command, handles JSON/YAML/TOML config differences |
| hooks-management | Manage hooks and automation for Claude Code, Codex CLI, and OpenCode |
| settings-management | Configure settings for Claude Code (JSON), Codex CLI (TOML), and OpenCode (JSON/JSONC) |
| skills-management | Organise, discover, share, and train skills for coding agents — SkillOpt-style optimisation loop with bounded edits, held-out validation gate, rejected-edit buffer, epoch-wise slow/meta update, plus trigger / transfer / diff / edit-log tooling |
| plugins-management | Package and publish plugins for Claude Code and OpenCode (validate, scaffold, submit to Anthropic's directory) |
| agentic-readiness | Audit and improve repositories for Codex/Codex App, Claude Code, and OpenCode, including AGENTS.md/CLAUDE.md discovery and routing |
| installing-cli-tools | Install, upgrade, authenticate, verify, and roll back developer CLIs while keeping tokens and API keys out of chat, logs, shell history, startup files, and repositories |
Engineering practices (8)
Disciplined approaches that improve how agents and humans collaborate on code.
| Skill | What It Does |
|---|---|
| code-that-fits-in-your-head | Keep agent-authored systems understandable and maintainable by resisting accidental complexity, Big Balls of Mud, and technical-debt accumulation. Supports large changes when architecture, acceptance criteria, verification, and recovery are explicit; uses structural heuristics without universal formatting or task-size limits |
| prompt-engineering | Universal prompt-engineering toolkit. 14 reference docs covering Claude / GPT / Gemini family-specific guidance, prompting techniques, evaluation & red-teaming, common failure modes (hallucinations, security, structure, debt) |
| fpf-problem-solving | First Principles Framework — a transdisciplinary "operating system for thought". Decompose cross-domain problems, audit conclusions, and reason rigorously about complex systems. Based on Anatoly Levenchuk's FPF |
| bug-fix-protocol | 8-step bug-fix protocol. Treats every production bug as two failures (the code defect + the testing system that allowed it through) and enforces a step-8 audit that closes the gap |
| plan-mode | Planning gate for any project change. Investigate context → analyse the task → surface ambiguities, contradictions, risks, and blockers → ask focused questions → produce an evidence-based, step-by-step plan → implement only after explicit approval, then validate. Compact always-on SKILL.md + full PROTOCOL.md |
| investigating-repository-history | Reconstruct historical intent before risky edits. Local git blame/log + GitHub PR/review evidence (via gh, never direct API), with squash/rebase/cherry-pick/rename/revert anomaly handling, decision-atom extraction, and confidence-scored output — produces a cited history note instead of guessing |
| refactoring-csharp | One-shot Roslyn contract for renaming C# symbols from file + 1-based line + oldName, with dry-run preview, safe file moves for supported named types, and explicit error codes |
| repo-activity-summary | Summarize what a repository has been doing from local git log — technologies, work types, churn hotspots, contributor patterns, velocity, and health signals. Stdlib-only Python, no network or API keys. Benchmarked at −42% cost / −37% wall clock versus the same agent answering the same question unaided, at equal factual accuracy; flags its own low-confidence classification instead of asserting it |
Research & docs (3)
| Skill | What It Does |
|---|---|
| semantic-scholar-deep | Deep research over the Semantic Scholar Graph API — backward references, recommendations, batch lookup (up to 500 IDs), multi-hop citation-graph BFS, snippet search. Ships with an optional paired subagent for token-isolated literature reviews |
| fetch-url-as-markdown | URL → clean Markdown via local trafilatura (real-browser UA, anti-stub guards, structured exit codes), with Exa MCP as a fallback for JS-rendered or anti-bot pages. Drop-in replacement for built-in WebFetch |
| ubiquitous-language | Domain thesaurus manager — DDD naming consistency, thesaurus generation, naming audit, git-history mining for naming ambiguities |
Multi-agent orchestration (1)
| Skill | What It Does |
|---|---|
| agents-consilium | Multi-agent review (read-only ask / multi-depth code) and single-agent delegate (full YOLO, steerable or detached, with wait / watch / list to block on, follow, or rediscover a run) via scripts/consilium. Stateful Grok 4.5 workers handle repository research; Codex Sol is an explicit second opinion for difficult spec verification or optimization planning. Codex, Claude Code, OpenCode, native Grok Build, Gemini (review-only). Independent opinions, structured specialists, super/ultra judge pipelines |
macOS & Windows system health (2)
| Skill | What It Does |
|---|---|
| maintaining-macos-health | Disk cleanup + dev-machine optimisation + proactive health alerting. Triage flow for kernel panic / watchdog timeout / vm-compressor-space-shortage / Jetsam events. Tiered cleanup playbook (zero-risk → discuss-first), Mole-style safety guards, and a noise-resistant LaunchAgent alerter (3 CRITICAL-only triggers, hysteresis, calibration window). Apple Silicon focus |
| maintaining-windows-health | Windows 11 port of the macOS health skill. Triage flow for BSOD / Kernel-Power 41 / crash dumps / commit-memory pressure. Ten-tier cleanup playbook (Storage Sense → discuss-first) using native Microsoft tooling (cleanmgr / DISM / pnputil / vssadmin / wevtutil / powercfg), a drift-protected HTML cleanup UI with an NTFS-hardened apply validator, and a Task Scheduler + BurntToast alerter (interactive-session, hysteresis, calibration). Heavy AI/Docker/WSL focus |
Niche utilities (2)
| Skill | What It Does |
|---|---|
| clipboard | Copy text to macOS clipboard with optional rich formatting (HTML + plain text). Slack/LinkedIn-aware — uses HTML rich text instead of mrkdwn, and warns about <table> not rendering in chat targets |
| windows-qa-engineer | Manual QA copilot for Windows 11 desktop apps (WinForms / WPF / UWP). Uses Microsoft UFO (UIA / Win32) + FastMCP mount() composition + MCP protocol |
Hooks (1)
Standalone agent-safety tooling that doesn't fit the skill format — typically because it has to live inside an agent's hook protocol or run as a compiled binary.
| Hook | What It Does | Install |
|---|---|---|
| balanced-safety-hooks | bash-guard — Claude Code PreToolUse:Bash safety hook in Go. Real Bash AST parsing via mvdan.cc/sh (heredocs, single-quoted prose, sudo/env/xargs/bash -c/eval/ssh/pipe-to-shell). Catastrophic-path matrix with safe-path carve-outs. Default rule set uses ask, not deny — agents trivially bypass deny by rephrasing. Covers rm/unlink/shred, ORM migrations, infra (kubectl/gcloud/helm/docker/terraform/git push -f), PaaS CLIs (railway/fly/heroku), DB clients (psql/redis-cli), and cloud control-plane API mutations | curl -fsSL https://raw.githubusercontent.com/CodeAlive-AI/ai-driven-development/main/hooks/balanced-safety-hooks/install-prebuilt.sh | sh |
Working principles
This collection is built on four principles that show up across all skills:
- Cross-agent first. Skills follow the Agent Skills open standard so they work in Claude Code, Cursor, Codex, Gemini, OpenCode, Antigravity, Copilot CLI — anywhere
npx skills addreaches. - Low false-positive friction. Tools that interrupt your workflow (the bash-guard hook, agent prompts) are tuned for high signal: only ask when the action is genuinely destructive, not when there's a fuzzy keyword match.
- Disciplined defaults. Skills like
bug-fix-protocolandfpf-problem-solvingcodify practices that pay off when the agent goes off the happy path — multi-step protocols rather than one-shot prompts. - Token-cheap. Skill descriptions are kept tight so they don't dominate context. The
semantic-scholar-deepskill goes further with an optional paired subagent for token-isolated research.
Repository structure
ai-driven-development/
├── .claude-plugin/
│ ├── marketplace.json ← lists this repo as a single plugin (source: "./")
│ └── plugin.json ← umbrella plugin manifest
├── skills/ ← canonical agent-skills layout
│ ├── agents-consilium/
│ ├── bug-fix-protocol/
│ ├── clipboard/
│ ├── code-that-fits-in-your-head/
│ ├── fetch-url-as-markdown/
│ ├── fpf-problem-solving/
│ ├── hooks-management/
│ ├── installing-cli-tools/
│ ├── investigating-repository-history/
│ ├── repo-activity-summary/
│ ├── maintaining-macos-health/
│ ├── maintaining-windows-health/
│ ├── mcp-management/
│ ├── agentic-readiness/
│ ├── plan-mode/
│ ├── plugins-management/
│ ├── prompt-engineering/
│ ├── refactoring-csharp/
│ ├── semantic-scholar-deep/
│ ├── settings-management/
│ ├── skills-management/
│ ├── ubiquitous-language/
│ └── windows-qa-engineer/
├── hooks/
│ └── balanced-safety-hooks/ ← bash-guard (Go, pre-built binaries via release)
├── README.md (+ ru/zh/pt-BR translations)
├── AGENTS.md ← canonical repository instructions
├── CLAUDE.md ← imports AGENTS.md for Claude Code
└── LICENSE
The same skills/ directory is read by both npx skills add (the cross-agent CLI) and Claude Code's plugin install — no symlinks, no duplication. The .claude-plugin/marketplace.json declares this repo itself as a single plugin via source: "./".
Contributing
Each skill lives in skills/<skill-name>/ with a SKILL.md (agent-facing contract) and an optional README.md (human-facing entry). Most skills also have references/, scripts/, or assets/ directories for supporting material.
When adding a skill or materially changing one, refresh:
- The skill's own
SKILL.mdandREADME.md - The relevant section table above
.claude-plugin/plugin.jsonand.claude-plugin/marketplace.jsonversion (semver — minor bump for new skills, major for breaking interface changes)
See AGENTS.md for the full development guide.
Heritage
This collection is the consolidation of nine previously-separate CodeAlive-AI repositories:
agents-reflection-skills(the 7 meta-skills, the marketplace base)prompt-engineering-skillfpf-problem-solving-skillclipboard-skillclaude-repo-explorer-skillwindows-qa-engineer-skillai-driven-development(the original — bug-fix protocol)awesome-agent-skills(5 skills + the bash-guard hook)
The 7 meta-skills' commit history lives on in this repo (consolidated via git mv). The other source repos retain their own histories at their original URLs and will be archived with pointer READMEs shortly.
License
MIT