# The Fable Loop Library: 25 Workflows on Autopilot
**作者**: Machina
**日期**: 2026-07-04T15:43:30.000Z
**来源**: [https://x.com/EXM7777/status/2073432521954697653](https://x.com/EXM7777/status/2073432521954697653)
---
i'm going to teach you how to run Fable 5 on autopilot, using my own library of loops and goals... 25 workflows, each with a prompt and the exact tool it plugs into
the method follows karpathy's recipe for agents: "here's an objective, here's a metric, here's your boundaries of what you can and cannot do. And go"
every workflow below is exactly that... an objective, boundaries, and the proof it's done
Fable 5 is the model built for this: it holds one job for hours without losing the plot, where other models wander off after minutes
remember these two Claude Code commands:
> /loop - a job that repeats on a schedule until it's done
> /goal - a finish line you write once, and the model works alone until it's crossed
one warning before you schedule anything: a model that never gets tired never stops on its own, and Fable is the most expensive model on the market
run it without a budget and a stop rule and the bill will find you... every workflow below carries both
the ready-to-run versions of these loops, and the training on turning them into income, live in the real time AI ops community: weeklyaiops.com
# what's a loop?
a loop is a job your agent repeats on a schedule: every morning, or every time something new lands
five parts, always:
> a schedule - when it wakes up
> ONE change per round - it fixes the single most important thing it found, never everything at once
> the same check every time - so this week's score can be compared against last week's
> a state file - a text file where it writes what it did and what's queued next
> a stop - a hard cap on rounds and a rule for what "done" and "blocked" look like
it's the same shape frontier researchers run their own experiments on: change one thing, test it, keep it only if it improved, write it down, repeat
the state file is the part almost everyone skips, and it's the part that makes every run smarter than the last
the model reads its own history before every round, so it never redoes finished work
every loop below also carries a color:
> green - safe to run alone: it only reads things and writes to its own files
> yellow - it drafts, you approve: replies, page edits, PRs, anything a human ships
> red - never runs alone: money, production, outbound messages, or anything a customer sees
run any loop once by hand before you schedule it
and route it cheap: a smaller model runs the routine rounds, Fable steps in only where the cheap model failed
# what's a goal?
a goal is the other shape of work: not a schedule, a finish line
you type /goal, describe what done looks like, and the model keeps working on its own... a second, smaller model reads along as the judge and confirms when the line is crossed
one detail decides whether this is useful or a token bonfire: the judge only sees the conversation
it can't open your files, can't run your tests, can't check your site
so "done when the tests pass" is a wish... "done when the full green test run is PASTED in the chat" is a contract
agents say "done" all the time without it being true... proof you can read is the only version that counts
every goal below ends the same way: paste the proof, and if you can't, paste the failures and stop
# the library
how to read each entry: the name, the tool it plugs into, why it earns a slot... and the full working prompt, ready for your terminal
two words you'll see:
> MCP - a plug that gives your agent direct access to an app (your inbox, your analytics, your repo, your ad account)
> API key - a password that lets the agent read a service on your behalf
the tools named per loop were researched for agent access specifically... hosted MCPs with OAuth where they exist, cheap pay-as-you-go APIs where they don't
paste each prompt into Claude Code, set the cadence line as your schedule, and swap the bracketed parts for your business
## marketing + content
1. answer-engine gap loop · Exa MCP
your buyers ask ChatGPT and Perplexity questions your site should be the answer to, and every unanswered one is a competitor's citation
this loop finds the gaps and closes exactly one per week
2. programmatic-page quality gate · Search Console MCP
if you generate pages from a template, thin and duplicate pages leak into google's index and drag the whole site down
the gate catches the worst pattern before google does
3. share-of-model brand watch · Perplexity API key
when buyers ask AI engines "best [your category]", you're either in the answer or invisible
every week it asks the same questions and logs where you stand, so you see the trend instead of guessing
4. competitor-content watch loop · DataForSEO MCP + Firecrawl monitor
your competitors tell you their SEO strategy every week... in the articles they publish and the keywords those articles chase
reading both and planning your next moves against their gaps is a friday job... this is the friday job
5. content-brief backlog loop · Search Console MCP
the writing pipeline dies at the blank page, not at the writing
the backlog stays stocked with fully specified briefs mined from your own traffic and questions
6. X content-idea miner · twitterapi.io + Typefully
the accounts you respect are running your content R&D for free... every high-performing post in your niche is a proven angle waiting for your version
a watched list, a ranking of what's working, and your take drafted in your own voice, every morning
7. ad-creative fatigue drafting loop · Meta Marketing API
ad sets decay, and the usual response is noticing two weeks late
the next batch sits drafted before the drop, grounded in what already won for you... and nothing launches without you
8. expectation-gap audit · Zendesk MCP
when a customer writes "i thought it did X", a page somewhere promised them X
this audit starts from the disappointment and walks backwards to the sentence that caused it
9. presale questions loop
the questions people ask BEFORE buying are answers you keep retyping
they're already sitting in your DMs, comments and emails... this loop turns them into an answer bank you never write twice
marketing finds the demand... product decides what gets built next
## product
10. brand-mention feature radar · Reddit API + HN Algolia + Exa
people tell the internet what they wish your product did, they just never tell you
the mentions get swept weekly, the complaints and feature asks clustered, and the loudest one arrives as a drafted implementation plan
11. review-mine to roadmap · app store review APIs
your customers write your roadmap every week in reviews and support messages... in their words, ranked by pain
it just needs compiling
12. drop-point copy loop · PostHog MCP
somewhere in your product, users hit a moment that makes them give up... and the words on that screen are usually half the reason
analytics finds the moment, this loop fixes the words
product listens... operations answers
## business ops
13. inbox-to-decision triage loop · Gmail MCP
a shared inbox is a decision queue wearing a disguise
every unread item becomes decide / delegate / defer / drop, with the top reply already drafted
14. month-close reconciliation prep · QuickBooks API
the boring half of closing the books is categorizing transactions against your own history
this loop does the boring half and hands a human a short exception list... it never files anything
15. SOP-drift catcher · Notion MCP
the written process and the real process drift apart in silence, and the doc is always the last to know
comparing how work actually happened against the SOP is a job, and this one does it weekly
16. proposal / RFP backlog drafter
every RFP asks 80% questions you've answered before and 20% that are genuinely new
this loop drafts the 80 from your answer library and flags the 20, so proposals move daily instead of stalling
17. KPI anomaly watch · PostHog MCP + Stripe
dashboards show everything, which is why nobody looks
silent on normal days, and when a metric breaks its band you get paged with the likely cause pre-investigated
18. unpaid-invoice chaser · Stripe API key
b2b clients pay late because late payment usually costs them nothing
this loop keeps an aging ledger and drafts the next nudge at each threshold... you send, it remembers
operations keeps the machine running... research decides where it points next
## research + decisions
19. overnight intel refresh · Exa MCP + Firecrawl monitor
markets move weekly whether you're watching or not
a cheap model watches your tracked sources on schedule, and the big model only writes the monthly synthesis
20. regulatory / source digest · Firecrawl MCP
some sources you can't afford to skim: the regulators and the two papers that will change your niche
only what materially changed gets digested, with a plain-language "so what"
21. hard-question escalation queue
paying flagship prices for questions a cheap model answers is the default way agent bills explode
this queue makes every question earn its escalation: cheap model first, Fable only on a logged failure
22. kill-criteria loop
options survive meetings because nobody wrote down what would disqualify them
here every live option declares its kill condition upfront... then the evidence hunt begins
23. pre-mortem loop
it's 12 months from now and the decision failed... writing that story before you commit is the cheapest insurance there is
24. repeat-offender digest
the same failure showing up in two different workflows is not two bugs... it's one system problem wearing two costumes
this digest reads every run ledger and hunts the costume changes
25. shadow prompt loop
testing a new prompt on hand-picked examples tells you it passed your imagination, nothing more
run it in the shadows against real traffic and let the disagreements decide
and if your business ships software, the next thirteen live inside your repo and your CI... the system that auto-runs your tests on every change
# start here
the whole system, in order... copy this:
> pick ONE loop that touches your biggest leak: inbox, KPIs, content, errors
> run it once by hand and read the state file it writes
> schedule it, with the cadence line as written
> after a week, add a second loop or hand a cursed one-shot to a /goal with pasted-proof done conditions
> route cheap: a smaller model runs the routine rounds, Fable takes what the cheap model failed
> money, production and outbound messages always end at a human... the prompts are built that way, keep it that way
one more thing on cost: Fable 5 comes off subscription plans on july 7 and moves to pay-as-you-go credits
after that date the cheap-first routing and the hard caps in these prompts stop being style and start being the bill
that endurance is sitting in your terminal right now
these 25 give it a job and a finish line
the smallest start is one loop and one state file
the rest schedules itself
join the best AI community in the world: weeklyaiops.com
(part 2 coming soon with 25 new workflows...)
## 相关链接
- [Machina](https://x.com/EXM7777)
- [@EXM7777](https://x.com/EXM7777)
- [182K](https://x.com/EXM7777/status/2073432521954697653/analytics)
- [weeklyaiops.com](https://weeklyaiops.com/)
- [twitterapi.io](https://twitterapi.io/)
- [weeklyaiops.com](https://weeklyaiops.com/)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [11:43 PM · Jul 4, 2026](https://x.com/EXM7777/status/2073432521954697653)
- [182.2K Views](https://x.com/EXM7777/status/2073432521954697653/analytics)
- [View quotes](https://x.com/EXM7777/status/2073432521954697653/quotes)
---
*导出时间: 2026/7/5 16:16:29*
---
## 中文翻译
# Fable 循环库：25 个自动驾驶工作流
**作者**: Machina
**日期**: 2026-07-04T15:43:30.000Z
**来源**: [https://x.com/EXM7777/status/2073432521954697653](https://x.com/EXM7777/status/2073432521954697653)
---
我要教你如何使用我自己的循环和目标库来实现 Fable 5 的自动驾驶……包含 25 个工作流，每个都配有提示词以及它所接入的精确工具。
该方法遵循 Karpathy 的代理配方：“这是一个目标，这是一个指标，这是你能做和不能做的边界。开始行动吧。”
下面的每个工作流都完全遵循这一点……一个目标、边界，以及完成的证明。
Fable 5 是为此构建的模型：它可以专注于一项工作长达数小时而不跑题，而其他模型几分钟就会走神。
记住这两个 Claude Code 命令：
> /loop - 一项按计划重复直到完成的工作
> /goal - 你只需写一次终点线，模型就会独自工作直到跨越这条线
在你安排任何任务之前有一个警告：一个不知疲倦的模型永远不会自动停止，而 Fable 是市场上最昂贵的模型。
在没有预算和停止规则的情况下运行它，账单会找上门来……下面的每个工作流都包含了这两者。
这些循环的现成可运行版本，以及将它们转化为收入的培训，都存在于实时 AI 运营社区中：weeklyaiops.com
# 什么是循环？
循环是你的代理按计划重复执行的一项工作：每天早上，或者每当有新事情出现时。
总是包含五个部分：
> 一个计划 - 它何时唤醒
> 每轮 ONE 个改变 - 它只修复它发现的最重要的那一件事，从不一次性修复所有问题
> 每次同样的检查 - 这样本周的分数才能与上周的进行比较
> 一个状态文件 - 一个文本文件，它写入它做了什么以及接下来排期了什么
> 一个停止点 - 轮次的硬上限，以及关于“完成”和“受阻”样子的规则
这与前沿研究人员运行自己实验的形态完全相同：改变一件事，测试它，只有改进了才保留，写下来，重复。
状态文件是几乎每个人都跳过的部分，而正是这部分让每一次运行都比上一次更聪明。
模型在每一轮之前阅读它自己的历史记录，所以它永远不会重做已完成的工作。
下面的每个循环也带有颜色标记：
> 绿色 - 可放心独自运行：它只读取内容并写入自己的文件
> 黄色 - 它起草，你批准：回复、页面编辑、PR（拉取请求），任何由人类发布的内容
> 红色 - 绝不独自运行：金钱、生产环境、对外消息，或任何客户能看到的内容
在安排任何循环之前，请先手动运行一次。
并采用低成本路由：一个小型模型运行常规轮次，Fable 只在小型模型失败时介入
# 什么是目标？
目标是工作的另一种形态：不是计划，而是终点线。
你输入 /goal，描述完成的样子，模型就会持续独自工作……第二个、更小的模型作为裁判跟随阅读，并在跨越终点线时予以确认。
有一个细节决定了这是否有用还是一场代币焚毁：裁判只能看到对话内容。
它无法打开你的文件，无法运行你的测试，无法检查你的网站。
所以“测试通过时完成”只是一个愿望……“当完整的绿色测试运行结果被 PASTED（粘贴）到聊天中时完成”才是一份合约。
代理们经常说“完成了”即使并非如此……你能读到的证明才是唯一算数的版本。
下面的每个目标都以同样的方式结束：粘贴证明，如果你做不到，就粘贴失败记录并停止。
# 库
如何阅读每个条目：名称，接入的工具，为何它能占有一席之地……以及完整可运行的提示词，随时准备输入你的终端。
你会看到两个词：
> MCP - 一种插件，赋予你的代理直接访问应用的权限（你的收件箱、你的分析、你的仓库、你的广告账户）
> API key - 一种密码，允许代理代表你读取服务
每个循环命名的工具都是专门为代理访问而研究的……在存在 OAuth 的地方使用托管的 MCP，在不存在的地方使用廉贵的按需付费 API。
将每个提示词粘贴到 Claude Code 中，将 cadence（节奏）行设置为你的计划，并将括号部分替换为你自己的业务。
## 营销 + 内容
1. answer-engine gap loop（问答引擎缺口循环） · Exa MCP
你的买家会向 ChatGPT 和 Perplexity 提问，而你的网站本应是答案，每一个未被回答的问题都是竞争对手的引用来源。
这个循环每周发现并填补一个缺口。
2. programmatic-page quality gate（程序化页面质量门） · Search Console MCP
如果你通过模板生成页面，单薄和重复的页面会渗入谷歌索引并拖累整个网站。
质量门会在谷歌发现之前捕获最糟糕的模式。
3. share-of-model brand watch（模型份额品牌观察） · Perplexity API key
当买家向 AI 引擎询问“最佳 [你的品类]”时，你要么在答案中，要么隐形。
它每周询问同样的问题并记录你的排名，所以你看到的是趋势而不是猜测。
4. competitor-content watch loop（竞品内容观察循环） · DataForSEO MCP + Firecrawl monitor
你的竞争对手每周都在告诉你他们的 SEO 策略……在他们发布的文章和这些文章追逐的关键词中。
阅读两者并根据他们的缺口规划你的下一步，这是一项周五的工作……这就是这项周五的工作。
5. content-brief backlog loop（内容简报积压循环） · Search Console MCP
写作流水线死在白纸前，而不是死在写作中。
积压列表始终保持库存充足，充满了从你自己的流量和问题中挖掘的完整简报。
6. X content-idea miner（X 内容创意挖掘机） · twitterapi.io + Typefully
你尊重的账号正在免费为你的内容做研发……你细分领域中的每个高表现帖子都是一个经过验证的角度，等待着你的版本。
一个监控列表，一个对有效内容的排名，以及每天早上用你自己的声音起草的见解。
7. ad-creative fatigue drafting loop（广告创意疲劳起草循环） · Meta Marketing API
广告组会衰减，而通常的反应是两周后才注意到。
下一批素材在下降之前就已经起草好了，基于那些已经为你获胜的内容……未经你批准，绝不发布。
8. expectation-gap audit（期望差距审计） · Zendesk MCP
当客户写道“我以为它能做 X”时，某个页面向他们承诺了 X。
这项审计从失望开始，向后追溯到导致它的那句话。
9. presale questions loop（售前问题循环）
人们在购买**之前**问的问题是你不断重复输入的答案。
它们已经坐在你的私信、评论和邮件中……这个循环将它们转化为一个你永远不需要写两次的答案库。
营销发现需求……产品决定接下来构建什么。
## 产品
10. brand-mention feature radar（品牌提及功能雷达） · Reddit API + HN Algolia + Exa
人们会告诉互联网他们希望你的产品做什么，只是从来不告诉你。
提及内容每周会被汇总，投诉和功能请求被聚类，声音最大的那个会作为一份草拟的实施计划呈现。
11. review-mine to roadmap（评论挖掘转路线图） · app store review APIs
你的客户每周都在评论和支持消息中写下你的路线图……用他们的话，按痛点排序。
这只需要编译一下。
12. drop-point copy loop（流失点文案循环） · PostHog MCP
在你的产品中，用户会遇到某个时刻让他们放弃……而那个屏幕上的文字通常是原因的一半。
分析工具发现了那个时刻，这个循环修复那些文字。
产品负责听……运营负责答。
## 业务运营
13. inbox-to-decision triage loop（收件箱到决策分诊循环） · Gmail MCP
共享收件箱是一个伪装的决策队列。
每个未读项目都会变成 决定 / 委派 / 延后 / 放弃，并且顶部回复已经起草好。
14. month-close reconciliation prep（月度结账对账准备） · QuickBooks API
结账枯燥的一半是对照你自己的历史记录对交易进行分类。
这个循环做枯燥的那一半，并交给人类一个简短的异常列表……它从不归档任何东西。
15. SOP-drift catcher（SOP 偏离捕获器） · Notion MCP
书面流程和实际流程在沉默中渐行渐远，而文档总是最后一个知道的。
对比实际工作是如何发生的与 SOP 是如何规定的是一项工作，而这一项每周都做。
16. proposal / RFP backlog drafter（提案 / RFP 积压起草器）
每个 RFP 都包含 80% 你以前回答过的问题和 20% 真正的新问题。
这个循环从你的答案库中起草那 80%，并标记出那 20%，所以提案每天推进而不是停滞。
17. KPI anomaly watch（KPI 异常观察） · PostHog MCP + Stripe
仪表盘展示一切，这就是为什么没人看它们。
平时沉默，当指标突破其区间时，你会收到传呼，附带预先调查过的可能原因。
18. unpaid-invoice chaser（未付发票催收员） · Stripe API key
B2B 客户付款晚，因为逾期付款通常对他们没有任何成本。
这个循环保留一个账龄账本，并在每个门槛起草下一次催促……你发送，它记住。
运营保持机器运转……研究决定它接下来指向何方。
## 研究 + 决策
19. overnight intel refresh（隔夜情报刷新） · Exa MCP + Firecrawl monitor
无论你是否关注，市场每周都在变动。
一个廉价模型按计划监控你的追踪来源，大模型只撰写月度综合报告。
20. regulatory / source digest（监管 / 来源摘要） · Firecrawl MCP
有些来源你承担不起略读的代价：监管机构和那两篇将改变你细分领域的论文。
只有实质性发生变化的才会被摘要，并附带“那又怎样”的通俗语言解释。
21. hard-question escalation queue（难题升级队列）
用旗舰级模型的价格去回答一个廉价模型能回答的问题，是导致代理账单爆炸的默认方式。
这个队列让每个问题都证明其升级的必要性：廉价模型优先，Fable 仅在记录的失败时介入。
22. kill-criteria loop（否决标准循环）
选项能在会议中幸存下来，是因为没人写下什么会让它们失去资格。
在这里，每个活跃选项都提前声明其否决条件……然后证据搜寻开始了。
23. pre-mortem loop（事前验尸循环）
现在是 12 个月后，决策失败了……在承诺之前写出那个故事是存在过的最便宜的保险。
24. repeat-offender digest（惯犯摘要）
同样的失败出现在两个不同的工作流中，不是两个 Bug……而是一个披着两套伪装的系统问题。
这个摘要读取每次运行账本并搜寻伪装的变化。
25. shadow prompt loop（影子提示词循环）
在精心挑选的例子上测试新提示词只能告诉你它通过了你的想象，仅此而已。
在影子模式下针对真实流量运行它，让分歧来决定。
如果你的业务运送软件，接下来的十三个工作流将存在于你的仓库和 CI 中……那个在每次更改时自动运行你的测试的系统。
# 从这里开始
整个系统，按顺序……复制这个：
> 选择一个涉及你最大漏洞的循环：收件箱、KPI、内容、错误
> 手动运行它一次并阅读它写入的状态文件
> 安排它，按照书写的节奏行
> 一周后，添加第二个循环，或者把一个棘手的一次性任务交给一个带有粘贴证明完成条件的 /goal
> 路由要便宜：一个小型模型运行常规轮次，Fable 接手廉价模型失败的部分
> 金钱、生产环境和对外消息总是止于人类……提示词就是这样构建的，保持原样
关于成本还有一件事：Fable 5 将于 7 月 7 日脱离订阅计划，转为按需付费信用额度。
在那之后，这些提示词中的“廉价优先”路由和硬上限不再是风格问题，而是实实在在的账单问题。
这种耐力现在就正坐在你的终端里。
这 25 个工作流给了它一份工作和一个终点线。
最小的开始就是一个循环和一个状态文件。
剩下的会自动安排。
加入世界上最好的 AI 社区：weeklyaiops.com
（第二部分即将推出，包含 25 个新工作流……）
## 相关链接
- [Machina](https://x.com/EXM7777)
- [@EXM7777](https://x.com/EXM7777)
- [182K](https://x.com/EXM7777/status/2073432521954697653/analytics)
- [weeklyaiops.com](https://weeklyaiops.com/)
- [twitterapi.io](https://twitterapi.io/)
- [weeklyaiops.com](https://weeklyaiops.com/)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [11:43 PM · Jul 4, 2026](https://x.com/EXM7777/status/2073432521954697653)
- [182.2K Views](https://x.com/EXM7777/status/2073432521954697653/analytics)
- [View quotes](https://x.com/EXM7777/status/2073432521954697653/quotes)
---
*导出时间: 2026/7/5 16:16:29* 
