@ThisisHan1_: 最近做了一条开发 pipeline,想分享一下背后的想法。 我是被 loop / goal engineering、还有 auto-goal(让 agent 自己写 goal、自己 spawn 子任务)那一串东西启发的。但真正让我想通的,…
摘要
这个开发管道通过先制作粗糙原型引发用户反馈,将每次“不对”的反应转化为可检查的规则,然后由AI代理独立开发并验收,以尽早发现问题并避免自欺欺人。
查看缓存全文
缓存时间: 2026/06/17 13:57
最近做了一条开发 pipeline,想分享一下背后的想法。
我是被 loop / goal engineering、还有 auto-goal(让 agent 自己写 goal、自己 spawn 子任务)那一串东西启发的。但真正让我想通的,是一件关于我自己的事:我的 taste 是“一看就知道”型的,不是“能描述”型的。很多东西我得看到了才知道“不对,不是这个”,事先根本写不清楚。冷著问我要什么,我挤不出来;可给我一个丑原型,我立刻知道哪里不对。
所以这条 pipeline 是反过来走的:先丢一个丑到不行的原型逼出我的反应 →把每个“不对,我要…”抄成一条可检查的规则 →这些规则变成验收标准,让 agent 一项一项开发 →而且每一项都由另一个没参与建造的 agent 独立验收,它没办法自己宣布自己做完了。
对比一下不这样做:你丢一句“帮我做 X”,agent 埋头一路跑到底,最后给你一个打磨得很漂亮、却不是你要的东西——还偏偏是在最贵的时候(结尾)才发现。这条 pipeline 做的,就是把“发现不对”这一刻,挪到最前面、最便宜的时候,而且全程不准 agent 放水、自欺。
灵感来源都很值得读:Lance Martin 讲长时间跑的 agent 会偷懒、自恋、偏离目标;Addy Osmani 讲怎么把 session 里反覆出现的修正,提炼成可复用的规则。
http://github.com/H1an1/idea-to-shipped…
H1an1/idea-to-shipped
Source: https://github.com/H1an1/idea-to-shipped
idea-to-shipped
A taste-driven build pipeline for Claude Code — a package of six skills that take a half-formed idea all the way to a self-running build, without asking you to write a spec, and without losing the parts of “what you want” that you can only recognize once you see them.
It’s built on one fact about how people actually work: your taste shows up when you react to something concrete, not when you’re asked to describe it in advance. So the pipeline never makes you specify everything up front. It makes you react, and it writes down what your reactions reveal.
Formerly
roadmap-to-loop— that skill is now just one stage of the larger pipeline.
The pipeline
idea
↓
1. idea-to-prototype → an ugly, fast mock-up to react to. You say "no, not that"; it learns.
↓
2. prototype-to-roadmap → once the picture's firm, it slices it into a buildable, ordered plan.
↓
3. roadmap-to-loop → turns the plan into LOOP.md: what "done" means, and who checks it.
↓
4. loop-to-goal → the agent writes its own goals and builds, unattended, checking itself
against your criteria, stopping only when it needs your eye.
↓
shipped
taste-to-rubric → runs underneath all of it. Every time you react to something — the first
mock-up, a mid-build check, final sign-off — it captures the preference
into RUBRIC.md so nobody has to ask you twice.
idea-to-shipped → the front door: it routes you to the right stage by what you already
have, and protects the principles the pipeline is built on.
RUBRIC.md — your taste, written down
Each project grows one file, RUBRIC.md. Every “no, not that” that turns out to be a standing
preference (not a one-off tweak) gets distilled into one checkable line, tagged revealed (you only
knew it on sight) or known. Over a project it becomes the rulebook the agent checks itself
against; rules graduate from “only the owner can judge” to “a verifier subagent can judge” over time.
Over many projects, those rulebooks can be compared to distill the constants of your taste — but
that’s later, once a few exist.
How to use it
You don’t pick a stage. Just say what you want — “I want something that…” — and the front door
(idea-to-shipped) figures out where you are and starts you at the right place. A vague idea starts
at the mock-up; if you already have a firm concept or a plan, it jumps in further down.
Your job, start to end, is small and only yours: react (at the first mock-up) and judge (at the gates the agent stops you for). Everything in between is the machine’s.
Install
These are Claude Code skills. Copy the ones you want into your personal skills directory:
git clone https://github.com/H1an1/idea-to-shipped.git
cp -R idea-to-shipped/skills/* ~/.claude/skills/
Each skill is a self-contained skills/<name>/SKILL.md; its description controls when Claude reaches
for it. You can also drop them into a project’s .claude/skills/ instead, to scope them to one repo.
The skills
| Skill | What it does |
|---|---|
| idea-to-shipped | the front door — routes you to the right stage |
| idea-to-prototype | vague idea → cheapest rough prototype that provokes your taste |
| prototype-to-roadmap | firm concept → sliced, ordered roadmap with your rules loaded in |
| roadmap-to-loop | roadmap → LOOP.md with acceptance criteria tagged machine / verifier / you |
| loop-to-goal | LOOP.md → unattended build that writes its own goals and verifies itself |
| taste-to-rubric | any reaction, anywhere → a written rule in RUBRIC.md |
What’s intentionally not built yet
- The cross-project master rubric (needs several projects’
RUBRIC.mdto compare). - Auto-loading rules back in / scoring — its shape will reveal itself once real rules pile up.
These come with mileage, not design. Run the pipeline on real things first; the next pieces will show you their shape.
相似文章
@freeman1266: https://x.com/freeman1266/status/2064702757773496552
本文介绍Loop Engineering概念,即通过设计自动化系统让AI代理自主循环工作,包括自动化任务、工作树、技能、插件、子代理等要素,从而取代手动提示的方式,提升开发效率。
@justloveabit: https://x.com/justloveabit/status/2070338139441484053
文章宣布Prompt Engineering已死,提出Loop Engineering作为2026年AI开发的新范式,强调设计自主循环系统(Plan-Execute-Verify-Iterate)让Agent自主完成复杂工作流,并给出了实际案例和上手方法。
@yanhua1010: https://x.com/yanhua1010/status/2072170080969138616
Claude Code团队发布博客介绍/goal和/loop功能,文章分析了AI编程从一次对话到循环操作的范式转变,详细解释了四种循环类型(回合制、目标制、定时制、主动式)及其适用场景,并提出了Harness Engineering的执行层概念。
@cellinlab: https://x.com/cellinlab/status/2064144608242679822
这篇文章介绍了 Loop Engineering 的概念——不再直接给 AI agent 写 prompt,而是设计一个系统(loop)来递归地让 agent 迭代工作,直到任务完成。文章详细对比了 Claude Code 和 Codex 在 automations、worktrees、skills、sub-agents 等五个构建块上的实现,认为这可能是未来与 coding agent 协作的趋势,但仍需警惕 token 成本和 AI slop 问题。
@jakevin7: 分享一下maka 最近在做一件很有意思的事:让 agent 自动优化自己的 system prompt,完全闭环,不需要任何人工介入。Karpathy 的 autoresearch、AEGIS 等都做过类似方向,一个有目标的自我强化学习系…
Maka是一个本地优先的桌面AI工作台,其新功能让agent自动优化自己的system prompt,通过生成变体、容器评测和acceptance policy迭代改进,无需人工介入。