You'll Want To Use This If...
Solves the manual context-shuttling problem by giving Claude access to logs, prompts, user behavior, infra, and ads so it can turn live product signal into validated fixes and priorities.
Arsenal is most useful when Claude needs to reason across product, prompts, infra, and live user signal instead of just the local codebase.
- You are rapid prototyping with live users: product, prompt, and infrastructure changes all need to move in the same loop.
- User signal should inform dev priorities: onboarding behavior, traces, feedback, and ad performance need to flow directly into what gets built next.
- Claude needs real product context: the answer depends on logs, LangFuse, databases, prompts, code, and operational systems, not just repo files.
- Validation is the bottleneck: you want AI to propose and implement changes, but only inside a workflow with checks, diffs, and review gates.
- Your team is AI-native: this fits small teams using Claude Code heavily, with infrastructure and product systems defined in code and accessible to the agent.
Problem This Solves
- Skill adherence: Hash gates force Claude to follow the right skill for every task, eliminating skipped or freelanced workflows.
- Context shuttle: Arsenal gives Claude live access to your product's world — logs, DB, prompts, ads, user behavior — so you get answers, not suggestions.
- Production visibility: Instead of pasting logs into Claude and getting generic restatements, you can ask what is happening in production right now and get trace-level answers tied to real systems.
- Feedback loop closure: User behavior, prompt changes, validation, and product priorities stay in one loop so live signal can drive the next fix or decision.
Before, I was a vehicle to shuttle context from the rest of the world to the language model to the developer. I underestimated the amount of overhead that created. — Sam Odio
3-Minute Demo
What makes the before/after unique is not just better prompting. Arsenal removes the manual context shuttle, so Claude can move from alert or user signal to trace, diagnosis, risk assessment, prompt or code change, and validation in one loop.
"Why did this error happen?" → pastes error log
"One of your LLM calls hit the max completion tokens limit." (just restated the error)
Pinpointed the exact Feb 26 13:17 UTC trace in LangFuse, showed the prompt had 11K tokens vs. a 512 completion limit, gave a risk assessment, proposed a fix, then implemented it — all in one turn.
"50 new users onboarded this week. Zero converted to couples."
"The 0% conversion rate isn't a funnel problem to fix — it's the market telling you what the product is."
Claude identified the core strategic problem, proposed two solutions, picked the winner from the data, wrote a spec, and updated all 150 LangFuse prompt templates — live pivot of the company in one session.
Also: generated 273 Facebook ad variants (images + copy) in ~15 minutes.
Getting Started
Add the installable Arsenal bundle from the
AgentCribs resources page
to your project repo as ./arsenal, then run:
# From your project root ./arsenal/install.sh # Optional: start Docker services (semantic search) cd arsenal && docker-compose up -d
What install does:
- Symlinks
arsenal/dot-claude/→.claude/in your project - Symlinks
arsenal/pre-commit-scripts/andAGENTS.md - Installs Node.js dependencies (Playwright, etc.)
- Sets up
.envfrom.env.example
Required env vars:
OPENAI_API_KEY=... LANGFUSE_PUBLIC_KEY=... LANGFUSE_SECRET_KEY=... LANGFUSE_HOST=...
After install, Claude Code automatically gets:
- A bootstrap hook that fires every session: reads
getting-startedskill before any work begins - Skill router that injects skill-checking instructions on every user message
- Manager review gate that blocks responses until a checklist token is included
Bundle Details
For install details, workflow inventory, and the full bundle contents, see the bundle README.
Community Reactions — Feb 25 Session
- Questions focused on: how validation pipeline was built, whether it's a standalone product, how the 40K CLAUDE.md is organized
- "The validation pipeline concept is a product by itself"
- Strong interest in Arsenal repo access (Sam offered to share privately)
Similar Tools
| Tool | Relationship |
|---|---|
| Superpowers (Jesse Vincent) | Arsenal was directly inspired by this. "Assume it's Superpowers." |
| OpenClaw | 4 community members using it; comparable agent orchestration layer |
| LangFuse | Arsenal wraps LangFuse for prompt debugging/iteration |
| Linear | Arsenal has Linear integration commands (/create-linear-ticket, /linear-agent) |
| HumanLayer | Mentioned in tool survey — similar human-in-the-loop concept |