Field-tested resources for
/building AI agents

Organize any agent's project into a filesystem-first layout — one agent = one folder, every capability a file in a conventional place. Scaffolds new agents or reorganizes messy ones.

A predictable structure keeps an agent findable and safe to modify as it grows; scattered files are where reliability and handoffs break down.

Faster onboarding and safer changes — less rework and lower maintenance cost as your agent fleet scales.

Gate sequence that picks an agent's build pattern and forces a "patterns NOT used" list — over-engineering dies in planning, before a line of code is written.

Most agent cost and fragility is decided before any code, when complexity gets chosen by default. Naming the pattern up front stops that.

Simpler agents shipped faster and cheaper; fewer wasted build cycles on complexity nobody needed.

Pick the right GTM tools for a play — filtered by customer-lifecycle stage and job category, ranked by real usage across 48 workflows, recalled live from tools.json.

Choosing tools by gut wastes budget and time; grounding picks in real usage data beats guessing from training data.

Right-sized tooling per play, less redundant SaaS spend, and faster GTM execution.

Choose the infrastructure that runs an agent — execution, LLM routing, evals, observability, state — per layer, sized to the agent's tier, with the layers NOT provisioned listed.

Under-provision and the agent breaks; over-provision and you burn budget. A tier-sized layer map avoids both.

Production-reliable agents without over-building — controlled infra cost and a faster path to launch.

A ~130-line local prompt router before OpenRouter — classify a prompt in <1ms, pick the model, hand the ID to your call. Cheap models for the workflow miles, frontier for thinking.

Routing every prompt to a frontier model is slow and expensive; sending cheap work to cheap models is where LLM cost is won or lost.

Cuts LLM spend materially with no quality loss on the work that matters — margin on every agent you run.

Draw and refine real diagrams on a live Excalidraw canvas in the browser — architecture, flows, sequences, Mermaid conversion — then export .excalidraw + PNG. Brand-first.

Architecture only aligns a team once it's drawn, and the diagram is usually the bottleneck. Letting the agent draw and export removes it.

Faster architecture alignment and clearer client comms — on-brand visuals without pulling in a designer.