The Bottleneck Dissolved

Two-week follow-up on The Nerve. Results first. I have several projects running in parallel now — the scale matters more than the list, so I’m skipping specifics — all on a single $200 Claude Max plan. Three outcomes worth pulling out:

(1) Token burn dropped hard. The $200 Max Plan feels hard to exhaust now. The real cause, in retrospect, was me. When I was in the loop, my own context would drift (I’d forget what I’d been doing five minutes earlier), my decisions came out low-quality and mostly unhelpful, and the AI had to burn tokens re-clarifying, re-answering, re-explaining. Pull the human out of the loop, that cost disappears.

(2) KPI-driven pushes actually work. The orchestrator watches the worst row and dispatches on it on its own. The number moves.

(3) My own hands-on time dropped hard. I’m mostly out of the daily loop. Below is the recipe.

Nerve is pure scripts. Hard-coded rules. The agent just calls the CLI. Simpler logic means less agent thinking. A complex one-shot flow—check KPI, read the Delta (Δ), dispatch, verify, fire the TG ping—becomes one script. One or two batch commands run the whole chain. When something breaks, fix the script first. Not CLAUDE.md. Not memory. Those drift. The script is the asset. Swap Claude for another model tomorrow, the system still runs.

The orchestrator’s game is the loop. A single line in the prompt changes everything. This is the current wakeup prompt:

Action (every wake):
1. `nerve recap` + `nerve board daily` → read top pick + KPI Δ
2. **Push top pick**: dispatch to owner pane, even incremental work
3. No top pick → pick the worst KPI row, unblock / double down
4. Truly nothing to do + all workers busy → sleep. Schedule a purposeful next wake before sleeping.
5. Need something from user → `nerve user-ask add "title" --tag X` to queue. Don’t shout in chat.

And this one matters more than it looks: the post-compact rule is what decides whether the agent can run long-term at all. After compaction, most agents drift hard. You need something watching at that exact moment — an LLM, an external script, whatever — but it has to self-monitor. Even then it can drift. When nothing else works, bolt on an external script that catches bad tool-calls and fires a wake-up prompt to drag the agent back into the intended logic. That prompt is what replaces me being there in real time.

A useful frame: one compact cycle = one “day” in the agent’s life. Within the day, scripts constrain what it does. At “the next day” (post-compact), a fresh set of rules + scripts has to take over, or the agent doesn’t survive the night.

Running an orchestrator isn’t new — any hierarchical-agent system has one. But mine sits entirely inside the $200 plan. And it scales. I can run several. An orchestrator can have orchestrators. As long as the tokens hold, $200 moves a lot of weight. It’s like a superpower now — which is why I’m building a big thing.

Core thesis unchanged: your ceiling isn’t how smart the AI is — it’s whoever’s driving it. Long-held view. Used well, it’s a lever that moves everything.

The bottlenecks I wrote about in The Bottleneck Moved and Wetware Upgrade are gone. Wetware is no longer the constraint. I distilled my own logic into the orchestrator. It took over the part of my brain that was frying. My head is clearer now.

给The Nerve 那篇的两周 follow-up。先讲结果。我现在有好几个项目在并行跑，规模不小（具体是哪些先不展开），全部挂在一个 $200 的 Claude Max plan 上。三个最值得拎出来讲的结果：

(1) Token 消耗大幅下降，$200 Max Plan 基本用不完。回头想，真正的原因是我：之前人在 loop 里，自己的上下文会丢（我会忘了之前在干嘛），给出的决策质量都很低，没起到任何帮助 — 反倒让 AI 反复去澄清问题、回答问题，不停地解释。人撤出 loop，这块开销就没了。

(2) KPI 推送效果很好，Result 非常强。Orchestrator 自动盯 KPI 最差那条去 push / unblock / 加码，数字是真的能被推动的。

(3) 我个人插手管理的时间减少了很多。基本不用待在每日 dispatch 的 loop 里。下面讲 recipe。

现在 nerve 纯依赖脚本，脚本写死规则，让 agent 调用 CLI，逻辑越简单越好，给 agent 减少认知负担。比如说有一些比较复杂的逻辑，那“一条龙”流程下来，既要看 KPI，又要看 Delta，然后还要做 dispatch。Dispatch 完了还要负责验收，验收完之后还要发 TG 消息。这一整套流程全部通过脚本实现，只需要执行一两个 batch 命令就可以全部完成。出问题时第一优先级就是改脚本，不会通过什么 CLAUDE.md 或者 memory 去改这些 — 说实话，这些东西都很不稳定。时间长了，这个脚本实际上就是沉淀下来的资产。以后套用其他任何模型其实都能跑，不需要依赖这个 Claude。

Orchestrator 重点在 loop 怎么设置，一行提示词差异都差很多。目前的 wakeup prompt：

Action (每次必做):
1. `nerve recap` + `nerve board daily` → 看 top pick 和 KPI Δ
2. **主动推 top pick**: dispatch 到 owner pane, 哪怕增量工作
3. 无 top pick → 挑 KPI 最差那条主动 unblock / 加码
4. 真没事做 + workers 全 busy 才 sleep. Sleep 前先 schedule 下次有目的的 wake.
5. 需要用户做什么 → `nerve user-ask add "title" --tag X` 入队, 不在 chat 喊.

然后这条比看上去重要得多：一定要有 post-compact 规则 — 一个 agent 能不能长期跑，看的就是这条。很多 agent 一 compact 性能就大幅漂移。所以 after-compact 阶段必须有东西一直盯着 — 可以是一个 LLM，也可以是别的，但它必须能自己监测。就算这样它可能还是会漂。实在没办法就挂外部脚本兜底：发现它工具调用开始乱，就通过 wake-up prompt 把它拉回正确的逻辑，代替我在现场引导。

一个好用的理解：一次 compact 相当于 agent 的“一整天”。一天内脚本约束它该做什么；到了“第二天”（after compact），必须有一套规则 + 脚本接着约束，不然 agent 活不过第二天。

Orchestrator 打通其实也不是什么新鲜事 — 其他有 hierarchy 的 agent 都有。但我这个是在 $200 的 Max Plan 里面全包的。未来的 integration 还可以有很多：我不光可以只有一个 orchestrator，我可以有几个，甚至 orchestrator 还可以有 orchestrator。只要 token 够，$200 可以撬动非常多的能力。It’s like a superpower now — 也是为什么 I’m building a big thing。

核心观点没变：你的上限不取决于 AI 多聪明，取决于用它的人。这是我长期的看法。用得好，它就是一根能撬动一切的杠杆。

有点感慨 — 之前博客里提到过的那些瓶颈（瓶颈转移了、Wetware 升级），其实都解决了。Wetware 升级现在也不是我的瓶颈了。用一个 Orchestrator 把自己蒸馏了就替代了，我现在精神状态比以前好很多。

The Bottleneck Dissolved

瓶颈都没了