Automating Workflows with Codex and Claude Code
Marketing Against the Graingo watch the original →
the gist
Stop using AI as a chat box and start building reusable 'skills'—text-based automation workflows—using tools like Claude Code and Codex to handle repetitive tasks while maintaining human taste.
Moving Beyond Chat to Systems
Peter Yang argues that the primary bottleneck for creators and professionals is treating AI as a conversational partner rather than an automation engine. By migrating workflows from standard chat interfaces to tools like Claude Code and Codex, users can transition from manual copy-pasting to building a 'Personal OS.' This involves documenting repetitive tasks, mapping them into discrete steps, and codifying those steps into 'skills'—simple text files that act as modular instructions for AI agents.
Building and Refining Skills
The creation of a skill is an iterative process. Rather than writing complex code, users should describe their manual workflow to the AI and ask it to structure the logic. To prevent 'AI slop'—the tendency for models to over-complicate or bloat instructions—Yang recommends a 'skill editor' skill. This meta-skill periodically reviews and refines other skills, ensuring they remain concise, actionable, and free of filler language. The quality of these skills is ultimately gated by the context provided; feeding the AI high-quality examples of past successful work is essential for consistent output.
The Role of Evals in Quality Control
To ensure AI output meets professional standards, Yang emphasizes the use of 'evals' (evaluations). He warns against asking the same agent to grade its own work due to inherent bias. Instead, he uses a separate agent to run a pass/fail check against a defined rubric. He specifically advises against asking AI for subjective scores (e.g., 'rate this 1-5'), noting that models struggle with nuance. Binary pass/fail checks are more robust and allow the agent to iterate on the draft automatically until all criteria are met.
The Human Element and Cognitive Risks
Despite the efficiency gains, the participants discuss the 'uncomfortable' side of AI integration. There is a genuine risk of 'AI brain fatigue' and the atrophy of critical thinking skills, where the user becomes a passive editor rather than a creator. Yang and the hosts stress that human taste remains the ultimate differentiator. They advocate for maintaining the 'genesis' of ideas—writing the first draft or defining the core angle manually—to ensure the work remains authentic and to prevent the loss of the ability to think from first principles.