Spec-first LLM coding workflow — plan before code, iterate in small chunks¶
Insight: Experienced LLM developers converge on a spec-first workflow: brainstorm a detailed spec.md with the AI, generate a step-by-step project plan, then implement in small iterative chunks. Osmani calls this "waterfall in 15 minutes" — a rapid structured planning phase before code generation. McKay Johns identifies Plan Mode as Claude Code's most powerful feature, emphasizing that users must understand the plan before execution. At Anthropic, ~90% of Claude Code's own code is now written by Claude Code itself.
Detail: The workflow is: (1) brainstorm spec with AI questions, (2) compile spec.md with requirements/architecture/testing strategy, (3) feed to reasoning model for project plan, (4) iterate on plan, (5) implement step by step. Context packing is critical — "brain dump" all information the model needs including constraints, examples, and warnings about approaches to avoid. Tools like gitingest and repo2txt help automate context packaging for large codebases.
Sources
- Addy Osmani — "My LLM coding workflow going into 2026" (2025-12-18)
- McKay Johns — "Get the Most Out of Claude Code" (2025-12-29)
Related: existing entry "Prompt engineering for coding is a systematic skill" in external/claude-code.md — COMPLEMENTS