AI coding was working pretty well
six months ago.

Why does the industry act like
you need the newest model
for every problem?

Stop chasing the model.

Engineer around the one you have.

Forge

A factory in a CLI

YAML in. Production code out.

Go service · tests · Terraform · K8s/KRO · CI · docs

Plain English in.

The architect asks
until it's unambiguous.

Six stages. Six agents.

Architect → Implementer → Infra → CI/CD → Quality → Docs

Direct Anthropic API.

NOT the Claude Code subscription.

1. The CLI is the control plane.

Your laptop IS the orchestrator.

2. Schema-validated handoffs.

The CLI ships its own contracts.

3. Self-healing is its own agent.

Recovery is agentic too.

Self-healing in action

architect · FIXING · output-fixer engaged

4. Patterns baked in.

Killed pattern-judge. Used examples.

Fewer moving parts.

5. Wiring is testable.

--dry-run

Scaffolding has its own test suite.

6. Token usage is visible.

Per stage. Per chunk. Real time.

No platform required.

• Control plane
• Contracts
• Recovery

The factory is itself a system.

Same discipline as production.

Use expensive intelligence
to build contracts.

Use cheap intelligence
to honor them.

Forge: built with Opus (Claude Max, 2.5h).
Runs on direct API. Downshifts to Haiku.

Subsidies are ending.

Copilot → usage-based

June 1, 2026

Token-metered. Monthly credits. Overage billed.

Token budget = invoice line item.

The invoice arrives in 30 days.

What this means

• Monitor your usage.
• Cost per unit of work is a metric.
• The factory is the survival strategy.

sniffly: top for tokens

Be constrained by your budget

Not by hourly rate limits.