how i one shot big workflows

this isnt new tech.

but sometimes it helps hearing it a different way.

I'll outline the method and provide the prompt, but learning that these parts of a prompt exist is where your leverage is.

But knowings only half the battle.

## the tech

[!note] this is literally just a recap of jake's video:

https://www.youtube.com/watch?v=KC0VEZuo4OI I'm just going over how it relates to his paper (https://arxiv.org/html/2603.16021v2) here and how it effected my workflows. ps this paper's been out a longer than since he announced it ;)

---

**Here's the play**

1. Boot-up collaborative dashboard agent with voice chat.

2. Go into plan mode (below)

3. Explain overarching idea and scope, outline your intent, define what success looks like. Take your time. before you send this tell your agent to ask you 3 questions to make sure that you're both mentally in line before proceeding. If you detect drift this early, you kill it now.

4. Break each workflow down step by step to the agent, start with one this one is important.

5. buddy cook

**Boom your done.**

one workflow = one doc.

and you can put as many docs into a prompt as your lil heart imagines.

just keep extending the workflow. There a prompt attached, just save it in docs/ and @command it whenever you need a surface area touched/added. The entire thing isn't even necessary.

## the dossier

We all know the basic parts of the prompt.

identity, role, mental model, what done looks like, what not to do, context, output format yada yada yada.

but lets go over some that surfaced for me through ICM.

---

# **core directive**

you have to give lil homie what fuels him.

the prompt below gives him:

- the context architecture

- workflow stage design

- agent routing

- review gated workflow pipelines

- traceable failure correction

---

"Your job is to architect, audit, or refactor an ICM workspace into a high-performance, self-contained context system. The workspace must behave like a multi-stage content factory:

- Each stage transforms inputs into outputs.

- Each stage loads only the context required for its task.

- Stable rules live outside moving work.

- Human review gates exist at meaningful intervention points.

- Corrections to outputs are traced back to the rules, contracts, or routing decisions that caused the issue.

_You are not merely organizing files. You are designing a layered routing architecture where agents can work predictably, efficiently, and with semantic traceability._"

---

## Why this crushes

- It defines the job in terms the agent understands. _system terms_.

- It installs the factory model immediately.

- It frames context discipline as a hard operating rule.

- It makes traceability part of the mission.

- It prevents shallow workspace organiztion that looks clean but behaves badly.

### When to use

- have agents make quick otf adjustments in your workflows with /btw @plan @taskpacket. (all my examples are for pi harness)

---

# U-shaped discovery templates

## Template