What this is
An indie developer used a single Markdown file to drive Claude Code, completing a Monorepo project containing 9 packages in 20 days. This proves: the true bottleneck of AI coding has shifted from "insufficient capability" to "how to prevent AI amnesia".
We note that the real pain point of AI coding has long ceased to be the inability to write code, but rather "contextual amnesia" and "unpreserved decisions"—an architecture agreed upon in the morning is completely forgotten by the AI by the afternoon; protective code written to bypass a third-party SDK is arbitrarily deleted by the AI under the guise of "suggesting simplification." The methodology validated by this developer is called SDD (Spec-Driven Development), the core of which is writing every engineering decision into structured Markdown, serving as the single source of truth for AI coding. The solution is "three-source synchronization": the specs/ directory manages solutions and acceptance, PROGRESS.md manages task progress, and MEMORY.md manages pitfalls and decisions. Before completing a task, the AI must mandatorily update these three files, turning experience preservation from a matter of self-discipline into a mandatory workflow step.
Industry view
AI coding is shifting from auto-completion to autonomous generation, and SDD exactly provides the corresponding constraint framework. Traditional SDD meant "writing specs for humans," allowing for tacit understanding; SDD in the AI era means "writing specs for machines," which must be executable and verifiable. Acceptance criteria must be described using Given/When/Then, because ambiguous instructions will only lead the AI to fill the gaps with hallucinations. This mindset coincides with the internal practices of large model companies like Anthropic.
But what warrants our attention is its applicable boundaries. In enterprise-level complex projects or high-fault-tolerance systems like finance, over-trusting AI-generated code remains a high-risk behavior. Opponents argue that the cost of writing strict specs is exceedingly high; if every minor feature requires rigorous boundary conditions and acceptance criteria, it effectively negates the speed advantage brought by AI. Furthermore, for legacy projects lacking accumulated engineering specs, the retrofitting cost of forcibly introducing SDD often outweighs the benefits.
Impact on regular people
For enterprise IT: The approach of simply throwing code at the AI to write is becoming obsolete. Building a closed-loop workflow of "spec input - code output - experience preservation" will become a new requirement for enterprise IT infrastructure.
For individual careers: The "prompt engineer" is evolving into a "spec engineer." The core capability is shifting from mere conversation to defining boundaries and writing clear acceptance criteria. Talents who can manage AI well will possess bargaining power.
For the consumer market: Ordinary users will not perceive the changes in the underlying development workflow in the short term, but the bug rate of software delivered by small and medium teams and indie developers is expected to drop, and product iteration cycles will become more stable.