cashcrown // case study · narzędzia AI / Claude Code skill
From a product brief we generate a coherent DAG of 200-1200+ atomic phases with seed code and gates that an agent works through on its own in a loop — with no way to mark anything "done" without proof.
An agent building a large app autonomously falls apart after a few dozen steps: it loses context, marks a phase "done" without implementing it, thrashes, or quietly stops. What was missing was a structure that holds the plan across hundreds of steps and won't let it fake progress.
Project timeline
Select a node to see its description and data flow.
A mature, battle-hardened internal skill for Claude Code — not a public SaaS. It generates the GENERATOR and a gated playbook, not a finished product; the app itself is built by the /goal loop. A postmortem of a real 668-phase build fed straight back into new gates. The limits are explicit: the TS-compile gates SKIP without node, and the evidence comes from self-audit and synthetic apps, not an external benchmark.