AI coding best practices: spec-driven development beats pure vibe coding
sujingshen · x · 2026-10-28
Tw93 hand-crafted a set of reusable AI coding rules; Sujingshen endorses it and layers on his Spec-Driven Development (SDD) method, forming a complete workflow.
Tw93's rules
- Shift from implementation-detail coding to "wish-style" programming: state bold goals, let the model figure out details
- Give agents background and purpose (why, for whom, what ideal looks like)
- Use the best models — weak models waste time and goodwill
- Have the agent confirm before adding features/toggles; no unused features
- Split big ideas into minimal viable versions, commit incrementally so you can roll back
Sujingshen's SDD layer
- Pin intent/spec/plan docs before coding to fix boundaries
- CLAUDE.md holds only rules not visible in code, plus verification checklists
- Rules must actually run via scripts, not just sit in prompts
Sagasu's linked essay (part 1 of a 9-part series based on 43 sources) argues the ambiguity of natural language is a systemic property: the clarification time you skip in prompts gets repaid at far higher cost in debugging, since agents fill unstated assumptions with training-data averages. Full article and a ready-to-use SKILL.md are shared openly.
More from coding & agent
- Petabyte pitches sleep-when-idle hosting for 24/7 AI agents with event-triggered wakeups — Petabyte-Cloud · 2026-10-28
- Claude built its own Obsidian plugin, and its human is reviewing AI-edited videos on iPhone — evielync · 2026-10-28
- Dev stitches Claude, GPT, Grok and Gemini into one multi-agent computer-use system — RileyRalmuto · 2026-10-28
- One illustration to a full VTuber model: dev builds open-source kit using Claude and Codex agents — bdsqlsz · 2026-10-28
- Ahrefs Podcast: How Ryan Law Vibe-Coded a Full Website in One Day — 5le · 2026-10-28
- Dev explains Microsoft's four Copilots as nested if, for, while and while(true) — pswider · 2026-10-28