split-llm-codegen-into-idea-plus-per-item

Imported from kjuhwa/skills-hub (skills/workflow/split-llm-codegen-into-idea-plus-per-item/SKILL.md). Install upstream with npx skills add kjuhwa/skills-hub --skill split-llm-codegen-into-idea-plus-per-item. Copyright stays with the author.

Split LLM Code Generation: Idea + Per-Item Calls

When generating multiple large code artifacts via LLM CLI, a single call requesting N items produces N small results. Splitting into (1 + N) calls — one lightweight idea call plus N individual code calls — gives each artifact the full response capacity.

When to use

• Generating 3+ substantial code artifacts (apps, components, modules) in one workflow
• Each artifact needs 500+ lines
• Single-call approach produces undersized results

Pattern

Call 1 (lightweight, 2min timeout):
  "Generate N ideas with names, descriptions, features"
  → Parse N idea blocks

Call 2..N+1 (heavy, 30min timeout each):
  "Build ONE complete app for idea[i] with full spec"
  → Each gets full response capacity

Key decisions

• Idea call is cheap (short response) — use tight timeout (60-120s)
• Code calls are heavy — use generous timeout (15-30min)
• Pass idea metadata (name, features) into each code call for consistency
• Concatenate results after all calls complete

Anti-patterns

• Requesting all artifacts in one call ("generate 3 complete 2000-line apps") — each gets ~1/3 capacity
• Not parsing idea output before code calls — fails silently if ideas malformed
• Equal timeouts for idea vs code calls — wastes time or causes premature timeout