A Strategy for Predictable Vibe Coding with Claude Code

Vibe coding is a term Andrej Karpathy first coined on X in February 2025, originally describing an extreme approach of barely reading code and just pushing through with "Accept All." In practice, it's used more broadly to include the flow where an LLM generates code and the developer verifies and adjusts it.

The problem is that when you hand off long specs and many rules at once, agents frequently go off track. For example, even if you feed in a 30-page authentication spec, requirements in the middle can get dropped, unrequested features can get mixed in, or existing patterns can regress in the wrong direction as they get pulled toward the prior from training data. The author explains this through Lost in the Middle, ambiguous spec interpretation, error compounding from autoregression, and a 4±1-level limit similar to human working memory.

The key is not to change what's inside the LLM, but to compensate for its limits with an external system. Just as humans handle complex work with sticky notes, checklists, role division, and interim checks, Claude Code can also improve stability through the following combination of tools.

• Multi-turn conversation with small requests: Rather than giving one big instruction at once, check in incrementally.
• Plan mode: Verify the plan before execution to catch directional errors early.
• Todo: Manage multiple files and tasks as a checklist to reduce omissions.
• Subagents: Separate context by domain to increase precision.
• Extended Thinking: Suited for tasks that require figuring out "why?", such as debugging, architecture decisions, and exploring edge cases.
• CLAUDE.md: Hierarchically holds project onboarding rules and per-directory conventions.
• Skills: Don't stop at listing rules—automate actual checks and workflows.
• Compact: As conversations get longer, summarize unnecessary context to restore focus.

The conclusion is clear. To use Claude Code reliably, rather than expecting the LLM to have complete understanding, you should break work into smaller pieces, verify the plan first, and attach external notes and inspection systems to increase predictability.