Capture workflow: a reusable Hermes use case for work that contains your judgment

This workflow saves me time on important work by adding one deliberate checkpoint before Hermes builds the final deliverable.

It is not the best approach for every task. It is the best approach when the request contains my judgment, the direction is still forming, and I want Hermes to reflect my thinking back before committing to execution.

For me, the input side usually starts with voice. Wispr Flow or something similar makes it much easier to capture messy thinking quickly, especially when sitting down to type would slow me down, fight the flow of ideas, or make me over-polish too early. Speaking at full WPM lets me dump more intent in less time, and that raw voice note becomes the cleanest starting point for the capture workflow.

What this use case produces

If I set it up once, I get:

• a dedicated Telegram topic for capture
• a reusable Raw/Shaped/Riff contract
• a saved Markdown artifact I can reference across topics
• a separate execution step for final deliverables
• optionally, a Proof document when the artifact needs iteration

The outcome is better direction before execution, not just faster output.

One-shot versus two-shot

A normal one-shot prompt looks like this:

Hey Hermes, create an outline for a video. Here is my idea...

That works for simple work. But for important work, one-shotting hides two problems inside one request:

• what I actually mean
• what Hermes should build from it

A two-shot workflow separates them.

One-shot output: video outline Two-shot output: video outline + identified spine + saved capture file + multi-video concept in Proof

The capture and approval flow is part of seven emails.

The case study

I thought I needed an outline for one video.

I dictated for roughly twenty minutes and sent the result into Hermes. It returned Raw, Shaped, and Riff.

The Riff pointed out that I did not have a structured outline yet. I had a pile of use cases with no spine. The suggested spine was that Hermes turns messy intent into useful artifacts, then keeps those artifacts moving through a workflow.

That was the moment where the direction changed. I realized I had too much material for one video. The request changed from “outline this video” to “turn this into a multi-video concept.”

Hermes created the Hermes in the Real World series concept in a Proof document.

How the workflow works

Wispr Flow voice input
  → dedicated Telegram topic
  → Hermes: Raw + Shaped + Riff
  → saved Markdown artifact
  → review and steer
  → execute in a separate prompt

This is not a general-purpose improvement. It is a decision checkpoint.

What Raw, Shaped, and Riff do

Raw preserves the original transcription. I use it when the shaped version seems too clean or too far from what I actually said.

Shaped makes the thinking easier to read. It removes filler, organizes related thoughts, and keeps uncertainty where I was genuinely uncertain. It should not invent a cleaner opinion than the one I expressed.

Riff adds connections, pressure tests, missing questions, and follow-up ideas. It should add value without claiming authorship.

The exact topic prompt I use

A dedicated Telegram topic makes this repeatable. The contract runs automatically every time I send a new capture.

If your agent is reading this post, it can use the block below as setup.

# Wispr Dictation Workflow

You are in the Wispr Dictation topic. Every message here is a voice dictation from Ken.

## Your Response Format

For every dictation, respond with three sections:

### 1. Raw (code block)
Put the raw transcription in a code block so Ken can easily copy/paste it.

### 2. Shaped
Take what Ken said and format it into a clean, readable version:
- Fix grammar, remove filler words, tighten phrasing
- Keep Ken's voice and intent — don't sanitize the personality
- Use headers, bold, bullet points where it helps readability
- If it's a rant/complaint, preserve the energy while making it skimmable

### 3. Riff
Analyze what Ken said and add value:
- Find connections to previous dictations (check ~/.hermes/dictations/ for past ones)
- Identify patterns, contradictions, or evolving thinking
- Surface relevant context from other conversations (BAS channel, LinkedIn posts, etc.)
- Push back gently if the take needs pressure-testing
- Suggest follow-up actions if applicable

## Saving

After responding, save the full output to:
~/.hermes/dictations/D-NNN.md

Where NNN is the next sequential ID. Check existing files in ~/.hermes/dictations/ to find the highest number and increment.

## Rules
- Always respond with all three sections
- Always save to file
- Share the ID in a code block: D-NNN
- Don't ask clarifying questions — just process the dictation
- If the dictation is short/low-signal, the Riff section can be brief
- If the dictation is about a specific project or topic, check related files for context

How to set it up

1. Create a dedicated Telegram topic for capture.
2. Attach the topic prompt above.
3. Use Wispr Flow, Telegram voice notes, Discord voice notes, or typed notes.
4. Save outputs to ~/.hermes/dictations/D-NNN.md.
5. Review the capture before asking Hermes to execute.
6. Reference the saved file in a separate execution prompt.
7. Use Proof when the artifact needs comments, iteration, or collaboration.

A dictation file usually looks like this:

# Dictation D-NNN — [date]

## Topic: [derived topic]

---

## 1. Raw
[raw transcription]

---

## 2. Shaped
[faithful version of what I meant]

---

## 3. Riff
[connections, follow-ups, pressure tests]

---

## Tags
[relevant tags]

When to use this

One-shot it when:

• the task is small
• the result is obvious
• mistakes are cheap
• it is easy to regenerate
• there is not much judgment inside the request

Two-shot it when:

• the request contains my judgment
• the problem is still forming
• I am deciding between competing directions
• the final artifact will be expensive to revise
• I want missing considerations surfaced
• the thinking should become reusable source material

Operating rules

• If I do not actually steer, the checkpoint is useless.
• If the Shaped section removes uncertainty, tighten the contract.
• If the Riff becomes generic advice, improve the instructions.
• If I stop reviewing the capture, the workflow loses most of its value.
• If the dictation is sensitive, store it carefully.
• If I need collaboration, move the result into Proof.
• If the task is reversible or simple, just one-shot it.

Proof comes after the direction is clear

Proof is the iteration surface. I use it when the artifact needs comments, edits, and shared collaboration.

Once the direction is clear, I ask Hermes to create a Proof document. That is much easier than going back and forth in chat trying to describe changes to an invisible draft.

The actual value

This workflow is not about talking to my phone. It is about creating one deliberate pause between thinking and building.

That pause makes the final deliverable better because the direction improves first.

Watch the full walkthrough