AgentForge

Opinionated Claude Code skill for building production AI agents and RAG systems. Modeled after UI/UX Pro Max, but for the agentic AI domain.

You ask:

"Build a RAG over 50k internal wiki pages. Citations required, low latency, Claude."

Vanilla Claude gives you five options. AgentForge first asks for load, latency, cost, then gives one prescriptive path with code, anti-patterns, and a starter eval — drawn from a curated knowledge base instead of generic LLM advice.

---

Install

npx agent-forge install claude-code          # → ./.claude/skills/
npx agent-forge install claude-code --dir ~  # → ~/.claude/skills/ (user-level)
npx agent-forge install cursor               # → .cursor/rules/agent-forge.mdc
npx agent-forge install windsurf             # → .windsurf/rules/agent-forge.md
npx agent-forge install copilot              # → .github/copilot-instructions.md

Or clone and copy:

git clone https://github.com/m0han22/AgentForge.git
cp -r AgentForge/.claude/skills/* ~/.claude/skills/
pip install rank-bm25 pyyaml   # optional, for BM25 search

Then in your editor: "Help me design an agent that reviews pull requests." The skill auto-activates on agentic / RAG keywords.

---

What's inside

Layer	What
7 skills	1 hub (agent-forge) + 6 focused (agent-architectures, agent-rag, agent-tools, agent-memory, agent-evals, agent-deployment)
Knowledge base	10 domains (safety, tools, loop, retrieval, evals, cost, memory, architecture, prompt, deployment) with 200+ inline rules + 30+ pattern docs
Framework profiles	Claude Agent SDK · LangGraph · LangChain · LlamaIndex · OpenAI Agents SDK · Pydantic AI · CrewAI · Deep Agents
Scaffolds	6 runnable starter projects, one per framework
Scripts	search.py (BM25) · recommend.py (synthesize) · eval_harness.py (generate golden set + Ragas)
CLI	cli.js — installs into Claude Code, Cursor, Windsurf, or Copilot

---

Use the scripts directly

SCRIPTS=.claude/skills/agent-forge/scripts

# Synthesized recommendation (search + framework + scaffold pointer)
python3 $SCRIPTS/recommend.py "RAG over 50k wiki pages" --framework llamaindex

# Drill into one domain
python3 $SCRIPTS/search.py "infinite loop budget" --domain loop

# Generate an eval harness for your task
python3 $SCRIPTS/eval_harness.py --task rag   # → ./eval/{golden_set.csv, eval.py, ci_gate.sh}

Try a scaffold

cd scaffolds/claude-agent-sdk
pip install -r requirements.txt
cp .env.example .env   # add ANTHROPIC_API_KEY
python main.py "What is the capital of France?"

---

Design principles

1. One prescriptive answer. Never "here are 5 options".
2. Load / latency / cost first. Asked before recommending — they drive most downstream decisions.
3. Markdown knowledge, not CSV. Patterns are markdown files with YAML frontmatter (richer than tabular).
4. No decision-rules engine. Claude reads matched patterns and reasons directly.
5. Knowledge in files, not prompts. SKILL.md routes; knowledge/ holds the IP.
6. Dual audience. TL;DR up top for experts, deeper explanation below for beginners.

---

Adding a pattern

Drop a markdown file into knowledge/<domain>/ with frontmatter (name, category, when_to_use, frameworks, related, anti_patterns, tags) and a body. The search engine picks it up automatically — no registration step.

cat > .claude/skills/agent-forge/knowledge/retrieval/colbert-late-interaction.md <<'EOF'
---
name: ColBERT Late Interaction
category: retrieval
when_to_use: token-level matching on long technical queries
frameworks: [llamaindex, langchain]
tags: [retrieval, colbert]
---
# ColBERT Late Interaction
**TL;DR:** ...
EOF

---

File layout

AgentForge/
├── cli.js                            # CLI installer
├── package.json                      # npm metadata
├── .claude/skills/                   # 7 skills (hub + 6 focused)
│   └── agent-forge/
│       ├── SKILL.md                  # router + embedded Quick Reference
│       ├── knowledge/                # 10 domains + frameworks + anti-patterns
│       └── scripts/                  # search.py, recommend.py, eval_harness.py
├── scaffolds/                        # 6 runnable starter projects
└── templates/                        # CLI install bundles per platform

---

Roadmap

• Publish to npm so npx agent-forge works without cloning
• Fill out reasoning/ (CoT, ToT, self-critique) and grow other domains past 1 pattern each
• Live dogfooding pass — verify auto-activation on real prompts
• Add scaffolds/deep-agents/ starter

---

Inspiration

Built after studying UI/UX Pro Max (same skill architecture, different domain). Architecture patterns informed by LangChain's Deep Agents, FareedKhan-dev/all-agentic-architectures, and Claude Code itself.

PRs welcome — add a pattern, fix a typo, or open an issue for substantive changes.

---

MIT © 2026 Sai Mohan Kesapragada