TypeScript educator Matt Pocock shares his configuration and steering strategies when using Claude Code with Opus 4.8 on a software development workflow.
Developer and educator Matt Pocock documented his experience experimenting with Anthropic's Claude Code CLI agent using the Opus 4.8 model. He highlighted key configuration settings, such as using custom slash commands like `/to-prd` to generate a Product Requirements Document and `/goal` to kick off implementation, alongside configuring `"autoCompactWindow": 180000` to keep the agent's context within the optimal "smart zone." Pocock noted that while the agent initially fell into the classic trap of trying to build the project in horizontal layers, he successfully steered it to focus on a vertical end-to-end "tracer bullet" implementation instead.
AI coding agents are highly powerful when combined with structured workflow commands, but they still require immediate human steering to avoid the classic pitfall of over-engineering horizontal layers.
* Custom commands like `/to-prd` and `/goal` help structure the agent's path, moving it from specification straight to execution.
* Setting configuration limits like `"autoCompactWindow": 180000` keeps context compact, preventing the agent from getting lost in bloated history or entering high-cost token zones.
* Left unguided, agents default to a "horizontal layer" architecture (building database layers, then APIs, then frontend), whereas human intervention is necessary to force a vertical "tracer bullet" approach for early validation.
* The mention of Opus 4.8 indicates testing Anthropic's latest generation models at medium reasoning effort settings, demonstrating the utility of steering in advanced LLM agent workflows.
DISCOVERED
1h ago
2026-06-09
PUBLISHED
2h ago
2026-06-09
RELEVANCE
AUTHOR
mattpocockuk