What This Is
Telling AI to "just do it" in a chatbox isn't nearly enough — that's the core argument of a long post that went viral in the Claude Code community this week. The article argues that documents and scripts are two fundamentally different things.
The former is "work requirements" handed to the AI: it interprets and executes on its own, may misunderstand, skip steps, or improvise — in essence, "advisory." The latter is a "production pipeline" handed to a program: the flow is hardcoded, failures must be retried, and the AI cannot skip steps — in essence, "mandatory." The former is like giving instructions to an employee (who may cut corners); the latter is like writing a machine program (which runs strictly).
When should you upgrade from documents to scripts? Four signals: need for parallel execution, need for automatic retry on exceptions, need to run independently outside a chat session, and need for state persistence — if any one applies, stop relying on the AI's "self-discipline."
Industry View
The progression of "documents for design, scripts for delivery" is the mainstream approach for current AI Agent (an intelligent agent — a program that lets AI autonomously execute multi-step tasks) deployment: first use documents to run through loop (closed-loop workflow) logic in conversation, then once verified, package into scripts and attach to system scheduled tasks (similar to cron) or scheduling systems like Claude Routines (Anthropic's cloud-hosted service). This corresponds to three built-in Claude Code commands: /goal (goal-driven autonomous loops), /loop (timed polling), and /workflows (multi-agent workflows).
But the objections deserve a hearing: first, documents are sufficient for many tasks, and forcing scripts is over-engineering; second, the three values of "observability, fault tolerance, and permission isolation" can actually be a burden for SMB IT teams; third, scripting essentially returns to traditional programming, eating into some of the "natural language as productivity" dividend AI offers. Our implementation advice: run with documents for two weeks first, then check whether you're truly stuck on stability and parallelism.
Impact on Regular People
For enterprise IT: Distinguishing "document-style pilots" from "script-based production" is the most common pitfall in AI Agent projects this year — many teams treat chat demos as production solutions, only to see stability collapse.
For individual professionals: The AI tools you use daily are essentially all "document-style" — the AI loses memory once you exit the session; to truly have it watch things for you, you must accept the engineered form of "scripts + scheduling."
For the consumer market: The AI automation accessible to ordinary users almost all stops at the "conversational" layer; true unattended automation remains mainly in the enterprise space, still one to two years away from reaching consumers.