Developer builds a five-step coding-agent loop skill, claiming 30% token savings

On 2026-11-07 at 17:28, a developer posted on V2EX that they had built their own five-step feature-development loop Skill and that it uses 30 percent fewer tokens than OpenSpec. The post describes the Skill as a workflow for independent developers. It can be loaded as a skill into coding agents including Claude Code, Codex, pi, and DeepSeek Harness. The flow has five stages: clarifying the design with an intent card, defining a single independently verifiable end-to-end slice, implementing and verifying, running an independent review with one pass for ordinary features and an extra structural review for framework-level changes, then acceptance and merge.

The developer states that, unlike common spec-driven tools, it does not maintain proposal / tasks / archive directories. The repository keeps only the current design: when the design changes, it is overwritten in place or deleted. It also keeps bookkeeping fully separate from development, so the developer says the coding process is not polluted by stale designs or old ideas. The same post claims the workflow is much lighter to maintain than OpenSpec while still retaining framework depth and delivery detail.

The A/B evidence cited in the post is a single-session open task. In that run, the developer reports model cost down 29.9 percent, input plus output tokens down 36.2 percent, cache-read down 48.6 percent, and retained documents down 51.6 percent, figures the developer says are consistent with the broader claim of more than -30 percent token cost, -48 percent cache-read, and -51 percent documentation maintenance compared with OpenSpec. The material does not include independent verification of these numbers, and no follow-up developments are reported; the event stands at the point where the developer published the self-built Skill and shared its stated test results.