Record-Narrate-Replicate Workflow Capture
Turn narrated work recordings into reusable AI workflows and skills.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 96%
This method captures a workflow from direct evidence rather than asking someone to reconstruct it afterward. A person records the task while narrating what they are doing, why they are doing it, which tools they use, and how they make decisions. The recording is transcribed and supplied to an AI agent, which analyzes the demonstrated process and proposes reusable workflows or skill files. Because actual elapsed time and tool transitions can be visible in the source material, the agent can produce a more grounded process model. The resulting workflow should then be tested against another real instance and refined before being shared or automated more deeply.
Origin
Extracted from Marketing Against The Grain.
Core principles
- 01Observe real work instead of relying on memory
- 02Narration exposes decisions that screen activity alone cannot explain
- 03Use agents to convert demonstrations into repeatable instructions
- 04Start with a representative workflow before scaling
How to run it
- 1
Select recurring work
Choose a task that happens often enough for improved execution or documentation to create meaningful value.
Pro tip Begin with a bounded workflow that has an obvious start and finish.
Watch out Do not start with an entire department when one representative task can validate the method.
- 2
Record the work
Capture the full task on screen so the agent can observe sequence, tools, handoffs, and elapsed time.
Pro tip Use a normal work instance rather than a polished demonstration.
Watch out A staged example may hide the friction and exceptions that matter most.
- 3
Narrate decisions
Explain what you are doing and why as you work, including criteria that would otherwise remain implicit.
Pro tip State what would make you choose a different action.
Watch out Screen activity without narration may reveal clicks but not judgment.
- 4
Transcribe and submit
Create a transcript and provide it to an AI agent with instructions to reconstruct or improve the workflow.
Pro tip Include relevant examples and reference files when they clarify standards.
Watch out Remove confidential information before sharing the material with an external model.
- 5
Generate the reusable asset
Ask the agent to create an AI workflow, operating procedure, or shareable skill suited to the intended use.
Pro tip Specify whether the output is for personal assistance, team use, or automation.
Watch out Do not assume the first generated artifact accurately captures every exception.
- 6
Test and refine
Run the artifact against a fresh instance of the task and update unclear instructions or missing decision branches.
Pro tip Have another person attempt the workflow to expose hidden assumptions.
Watch out Do not distribute an untested skill as an authoritative process.
In the wild
A marketing leader records a two-minute walkthrough of reviewing slides, applying evaluation criteria, consolidating feedback, and drafting an email. The transcript is given to an agent, which reconstructs the current process and proposes an AI-assisted version.
→ The formerly implicit review process becomes a documented workflow that can be tested and improved.
A three-hour coding recording is transcribed and analyzed. The agent identifies the existing data-analysis steps and recommends automating project setup, directory creation, document import, and conversion to Markdown.
→ The recording yields grounded infrastructure and workflow improvements tied to actual work.
Common mistakes
Recording without narration
The agent sees actions but cannot reliably infer the criteria or reasoning behind them.
Capturing an artificial demonstration
A sanitized walkthrough can omit delays, exceptions, and tool friction that the improved workflow must address.
Skipping validation
A plausible generated skill may still misrepresent the real process until it is tested on fresh work.
Is it for you?
Best for
This is best for recurring knowledge-work processes that experienced employees can demonstrate but struggle to document.
Not ideal for
It is not ideal for highly variable work whose critical decisions cannot be observed or narrated.
From the transcript
“you can record yourself just doing work and narrating your work, and you can give that to an agent, and that agent will build skills…”
“So literally you can just record yourself and it will tell you how to do what you're doing in an AI first way.”
“If you just record an hour-long video of yourself doing something, then it's gonna know it takes you an hour.”
From the episode
I Built a $20,000 AI Consultant You Can Have For Free