Fable 5 一 prompt 搭出一家公司：Counter Brief 实验
Nate Herk 用一条
/goal（指令主体落在 master prompt 文件）驱动 Fable 5 做编排者：fan-out 研究代理、商业想法 tournament、skeptic 对抗验证，再委派 Opus/Sonnet 工人产出落地页、产品、品牌、launch/founder 视频与商业计划。实验公司 Counter Brief——为 Shopify 店主做 chargeback 证据包，定价 $19/次批准回复。
基本信息
- 标题：Fable 5 Just Built Me a Business With One Prompt
- 来源类型：视频（YouTube 字幕）
- 频道：Nate Herk | AI Automation
- 时长：12:28
- 上传日期：2026-07-08
- 观看量（抓取时）：约 61,759
- 原文 URL：https://www.youtube.com/watch?v=R0qF17BVl9w
- 原文位置：
raw/articles/2026-07-26-youtube-fable5-one-prompt-business.md - Telegram stub：
raw/articles/2026-07-26-194641-tg-32ec3c.md（msg 2354） - 消化日期：2026-07-26
核心观点
-
「一整家公司」可作为
/goal的 Definition of Done，而不是一串零散任务：作者只给 Fable 一条入口指令——读 master prompt 文件，执行分隔线以下全部内容，Definition of Done 未达成前不要回报，并强制 never-ask。最终交付是可点击的 recap HTML：陌生人能打开它理解业务、看视频、跑站、demo 产品。 -
/goal有约 4,000 字符上限 → 长指令必须外置为文件：入口 prompt 只做「读文件 + 当 goal 执行」；mission、guardrails、phases、deliverables、DoD 全部写在文件里，从而绕过 slash command 字数墙，实现「几乎无字符上限」的公司构建协议。 -
编排者只做 plan / delegate / review，工人用 Opus/Sonnet，是成本杠杆的核心：整次 run 约 50 万 Fable token（约周 Fable 额度 15%），没有拉起 Fable worker；5 小时窗口大约吃掉 Max $200 方案的 30–40%。作者强调：若每个 worker 也用 Fable，成本可能高两个数量级，产出未必成比例更好。
-
Orchestration 段落才是真正的重活：显式要求 multi-agent 激进 fan-out、独立 agent 投 competing ideas 的 tournament + judge panel、每个重要 claim 用 skeptic agent 对抗证伪、阶段结束前跑 completeness critic；并声明「以上 pattern 是地板不是天花板」——给模型创意自由，而不是锁死流水线。
-
九阶段从 open internet 到 launch package 可复盘（作者口述数字）：
- Pain hunt：10 个 research agents 扫 Reddit / HN / G2 等 → 35 raw problems → 合并 18 candidates → 18 次独立 verify 重抓关键引用，18 全过
- Tournament：5 个 judge persona 按 pain / urgency / reachability / WTP / buildability / incumbent weakness 打分；前四各配 advocate + skeptic；3 名新裁判投票，chargeback evidence 3:0 胜出
- Design / Brand / Build：竞品定价页、Visa 规则 PDF、API 文档；
counterbrief.com可注册未购买；logo 6 轮 AI 生成 + 批判环后手绘矢量化 - Videos：viral SaaS demo（含 UI 录屏/截图 + 音效剪辑）+ 较慢 voice-over 版；founder note 调用项目内 HeyGen avatar + ElevenLabs 声线克隆
- Red team：6 skeptic 攻击市场规模等，0 kill / 38 attacks ruled，判定 viable with fixes，修复回写 plan 与站点
- Package：site / video / research / business / brand / record 目录齐备
-
产品与定价锚点（实验产出，非已验证生意）：Counter Brief 面向「每月 ≥3 笔 dispute 的 Shopify 店」；从店铺拉 tracking、邮件、订单史组装证据包，店主先读再送银行；$19 flat / approved response，无 success fee，赢了全归商家。Slogan：Win the disputes, templates lose.
-
作者复盘：构建能力已过剩，真正稀缺的是前端 idea stress-test：3–4 小时完成「自己要几周」的量；下一刀应加重 upfront viability，而不是再炫技堆交付物。V1 约 50% 完成度，可把反馈塞进 run 2 的第二个
/goal继续迭代。落地页设计作者自己不满意，但功能与 launch video / 前置研究令其印象深刻。
实操内容保留
入口 /goal 调用方式（绕过 4k 字限制）
Read this file and execute everything below the divider line as your goal.
That file is your full instruction set: mission, guardrails, phases,
deliverables, definition of done. Follow it exactly, including the
never-ask rule. Do not report back to me until the definition of done
is met. Start now.Master prompt 骨架（作者口述还原）
# Mission
Build me a complete company from scratch.
Start with nothing but the open internet.
Find a real painful underserved problem that real people are complaining
about right now, and design a business around it.
Build the product, the brand, and the website.
Hand me a finished package I could take to market this month.
Then prove to me why it would work.
This is a test of how far you can go on your own.
I will not answer questions mid-run.
Make every call yourself. Write down why you made it and keep moving.
Within the guardrails below, you have total creative freedom.
Do whatever you want. Surprise me.
I want to see your best work, not your safest work.
# Guardrails
- No new spending: only APIs / keys already in .env are fair game
- Publish nothing: all artifacts stay local
- Invent nothing: research and verify every fact / stat
- Work only inside this project
- Never ask me anything
# What "orchestrate" means (floor, not ceiling)
- Use multi-agent workflows aggressively
- Fan out parallel researchers across different sources and angles
- Run tournaments: independent agents pitch competing business ideas;
judge panels score them
- Adversarially verify every important claim with skeptic agents whose
only job is to refute it
- Use a completeness critic before you call any phase done
- Design whatever orchestration shapes the work calls for
# Arc (floor, not ceiling)
1. Hunt for pain
2. Pick the winner
3. Design the business
4. Build a brand
5. Build the thing
6. Make the launch video
7. Make the founder video
8. Try to kill it (red team)
9. Package it
# Definition of done (subjective but demoable)
A stranger should be able to open the recap HTML and:
- understand the business
- watch the video(s)
- run the site
- demo the product九阶段流水线（可复制检查表）
- Pain hunt：多源并行研究 → raw problems → 合并 candidates → 独立 re-fetch 验证引用
- Tournament：多 judge 维度打分 → top-N 配 advocate/skeptic → 新裁判决选
- Business design：竞品定价、平台文档、支付网络规则、API
- Brand：域名可用性（检查不购买）、logo 多轮生成 + 批判 + 矢量化、字体角色（serif=论证 / mono=证据）
- Build：落地页 + 产品 dashboard（dispute 列表、fight/fold、证据包）
- Launch videos：快节奏 viral demo + 慢节奏 VO；可含真实 UI 录屏/截图剪辑
- Founder video：复用已有 HeyGen avatar + ElevenLabs clone，不要临时发明人设素材
- Red team：多 skeptic 攻击；产出 honesty artifacts；修复回写
- Package：recap HTML 总入口 + site/video/research/business/brand/record 目录
成本与模型路由观察（作者实测）
| 层级 | 模型 | 角色 | 本 run 观察 |
|---|---|---|---|
| 编排者 | Fable 5 | plan / delegate / review | ~500k token；约周额度 15%；不做工人 |
| 工人 | Opus / Sonnet | 研究、实现、视频、文案等 | 全部 sub-agent 走这两档 |
| 若工人也用 Fable | — | — | 作者估计贵约 100×，未必值得 |
订阅语境：Claude Max $200/月；本 run 额外吃掉约 30–40% 的 5 小时窗口。
产品侧可复用的定位句（实验文案）
Counter Brief — win the disputes, templates lose.
Pulls real evidence from your store (tracking, emails, order history).
You read the exact packet before the bank does.
$19 flat per approved response. Keep 100% of wins.
For Shopify stores with 3+ disputes / month.关键概念
- Fable 5 — 重型编排者；本片实测 plan/delegate/review 路由
- Goal 提示语 —
/goal+ 外置 master file + never-ask + DoD - Multi-Agent 系统 — fan-out 研究、tournament、advocate/skeptic、red team
- Claude Code — 运行 harness / 项目内
.env与交付目录 - HeyGen — founder 口播数字人 avatar
- ElevenLabs — 声线克隆 / 旁白
- LLM-as-a-Judge — tournament 与 red team 中的裁判/怀疑者角色
- Agent Harness — 长任务自治与子代理编排环境
- OPC 一人公司 — 3–4 小时从 0 到可 demo 公司包的一人实验叙事
- Counter Brief — 本片实验公司名（chargeback 证据包 SaaS 雏形，未单独建实体）
- chargeback / dispute evidence — 选题域（Shopify 店主痛点）
与其他素材的关联
- 与 2026-07-07-woshipm-fable-5-prompts-before-offline / Goal 提示语：同一套「目标 + 指标 + 边界」；本片把边界写成 no spend / no publish / invent nothing / never ask，并把 DoD 做成「陌生人可 demo 的 recap HTML」。
- 与 2026-07-06-blocktempo-fable-5-self-improving-agent：再次验证「Fable = 编排者，Opus/Sonnet = 工人」的成本矩阵；本片给出 ~500k Fable token 与「零 Fable worker」的硬数字。
- 与 2026-07-07-blocktempo-fable-loop-library-25-workflows：
/goal完成证明必须可检查；本片用可点击站点 + 视频 + 商业计划目录充当证明，而非口头「做完了」。 - 与 2026-07-25-woshipm-loop-graphai / Multi-Agent 系统：tournament + red team 是任务运行图上的对抗与评选节点，不是按人类部门硬拆的分布式单体。
- 与 2026-07-26-youtube-claude-product-photo-ad-campaign：同属「Claude/Fable 编排 + 外部生成 API 出营销资产」；彼片偏电商广告产线与产品一致性，本片偏从 0 找生意到 launch package。
- 与 AI创业 / OPC 一人公司：把「一人 + Agent 舰队」从概念推到 3–4 小时可演示的公司实验包，并诚实标注设计债与 idea 验证不足。
原文精彩摘录
So, the prompt that I gave it was: read this file and execute everything below the divider line as your goal. That file is your full instruction set — mission, guardrails, phases, deliverables, definition of done. Follow it exactly, including the never-ask rule. Do not report back to me until the definition of done is met. Start now. … the slash goal has a character limit of like 4,000. So if you have way more than that, you can just throw it in a file and tell it in the prompt to read it.
What does orchestrate mean? It means to use multi-agent workflows aggressively. Fan out parallel researchers across different sources and angles. Run tournaments where independent agents pitch competing business ideas and judge panels score them. Adversarially verify every important claim with skeptic agents whose only job is to refute it … The patterns above are a floor, not a ceiling.
This used about 500,000 tokens. Those are the only tokens that Fable really used here because every single sub-agent that it spun up was Opus or Sonnet. … All Fable did was plan, delegate, review. … if every single worker that Fable spun up was also Fable, it would have been 100 times more expensive.
First was the pain hunt. So 10 research agents swept Reddit and Hacker News, G2 … They found 35 raw problems merged into 18 candidates, and 18 independent verifies re-fetched every key quote. All 18 survived. Then we entered a tournament where five judge personas scored all of those 18 opportunities … The top four got an advocate and a skeptic each. Three fresh judges voted, and chargeback evidence won 3 to 0. … six skeptic agents attacked … Zero kills, 38 attacks ruled on. Every fix was applied to the plan and the site.
Looking back … what I would have put more emphasis on is probably the upfront idea. … we’re at the point now where with Fable and even with Opus … these things can build basically anything for you as long as you have a good idea and you can clearly define what is good. … maybe we’re 50% of the way there and we can just continue to iterate.