Peter's Builder Playbook: Models Are Roles, Feedback Is the System

Peter Steinberger did not publish a new long-form thesis this week, so it is better not to manufacture one. His recent public practices around GPT-5.6, Claude Fable, Codex, and Claude Code are still worth revisiting because they form a coherent operating model: models are roles, while the harness, graph, feedback, and recovery path turn those roles into dependable output.

The first lesson is to stop searching for a universal winner. Peter uses Fable in design exploration, Codex as an implementation workhorse, Claude CLI for direct execution in other contexts, and separate agents for review. This is not model loyalty. It is task routing based on where a model lowers the cost of iteration and verification.

Design needs high-bandwidth communication. Natural language can state rules but is slow at aligning visual hierarchy, density, and spatial relationships. Producing a visible design first turns taste into an artifact that both a person and an implementation agent can reference. The value is not that a model can draw; it is that ambiguity becomes inspectable.

Graphs serve the same purpose for system behavior. A long prompt about a complex feature tends to lose state transitions, branches, and failure paths. A diagram of components, events, data flow, and dependencies converts a wish into structured constraints. The graph is not decorative documentation. It is an execution interface.

Implementation then needs endurance rather than a beautiful one-shot answer. An agent such as Codex is useful when it can continue reading a repository, making a small change, running checks, and repairing the edges. That persistence becomes quality only when the harness returns real test and review feedback to the loop.

A second model is valuable because it introduces a different set of assumptions. The agent that wrote a solution is likely to continue rationalizing its own structure. A separate Claude or Codex review can focus on interfaces, failure cases, naming, or user experience. Multiple agents should not vote on the same vague prompt; they should own distinct risks.

The same rule makes parallel QA useful. One lane checks regressions, another inspects security boundaries, another runs the real interface, and another compares documentation to behavior. Each lane needs a precise input, evidence format, and stop condition. Without that structure, parallelism only multiplies similar noise.

The harness is the compounding asset. It assembles context, tools, task state, permissions, validators, and feedback. A better model may raise first-pass success; a better harness prevents a local failure from destroying the whole run. Because models change quickly, the reusable operating system around them often compounds longer.

Recovery is part of that system. Long tasks will be interrupted by network failures, process exits, model drift, or a changed requirement. Each run should leave a handoff packet: objective, confirmed facts, modified artifacts, checks already run, unresolved assumptions, and the exact next step. The best context is not the longest transcript; it is compressed state from which work can resume.

A solo builder can implement the method with five explicit roles: scout, designer, executor, critic, and verifier. One model may play several roles across a project, but each turn should have one clear responsibility. Start with simple routing rules, preserve intermediate artifacts, and require independent evidence for high-risk changes.

The deeper lesson is that model rankings move faster than a good work system. GPT and Claude can trade places, while role definitions, task graphs, validators, checkpoints, and handoff protocols remain useful. The durable question is not whether to use Codex or Claude Code exclusively. It is what role this step needs, how it receives context, how it recovers, and who verifies the result.