Workflow-to-Skill Capture
Turn narrated real work into a reusable AI skill.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 97%
Workflow-to-Skill Capture begins with evidence of real work rather than a speculative automation brief. A practitioner records a recurring task while narrating actions, inputs, judgments, exceptions, and desired outputs. The recording is transcribed, then an AI analyzes the transcript to identify the capabilities, decision rules, tools, and reusable instructions involved. Those findings become a focused skill or workflow that is tested on another instance of the same task. The practitioner compares the generated result with the standard they would produce manually, adds missing context, and repeats. This mechanism captures tacit knowledge that is often absent from formal process documents while grounding automation in an activity that already creates value.
Origin
Extracted from Marketing Against The Grain, where the hosts describe recording daily work, transcribing it, and converting the demonstrated process into AI skills and workflows.
Core principles
- 01Capture how work actually happens, not how a process document claims it happens.
- 02Narration exposes tacit decisions that ordinary documentation misses.
- 03Derive skills from recurring work rather than imagined use cases.
- 04Improve the generated skill through repeated use and feedback.
How to run it
- 1
Select a Recurring Task
Choose one workflow that occurs often enough to justify refinement and has a recognizable successful output.
Pro tip Start with a task you expect to perform again within the next week.
Watch out Do not begin by attempting to map an entire role or department.
- 2
Record and Narrate the Work
Complete the task normally while explaining what you are doing, why each decision matters, and what would cause you to change course.
Pro tip Narrate exceptions and rejected options, not only visible clicks.
Watch out A silent screen recording will omit much of the judgment the skill needs.
- 3
Create a Transcript
Transcribe the recording so an AI can inspect the workflow as structured evidence.
Pro tip Preserve references to inputs, tools, and intermediate outputs.
Watch out Remove confidential information before sharing the transcript with an external model.
- 4
Extract Skills and Decisions
Ask the AI to identify the capabilities, steps, decision rules, dependencies, and quality criteria demonstrated in the transcript.
Pro tip Challenge the AI to distinguish reusable rules from one-off actions.
Watch out Do not treat every action in the recording as a separate skill.
- 5
Build One Skill
Convert one coherent workflow into explicit instructions, required inputs, tool usage, constraints, and an output format.
Pro tip Include examples of both acceptable and unacceptable outputs.
Watch out Avoid producing a large repository of untested skills.
- 6
Test and Refine
Use the skill on a fresh task, compare its output with expert work, and revise the instructions around observed gaps.
Pro tip Track recurring corrections so they become permanent instructions.
Watch out A convincing first result does not prove the skill is reliable.
In the wild
A product marketer records a narrated session in which they turn a feature brief into launch-page copy. The transcript reveals how they identify the audience, select proof points, reject vague claims, and structure the page. An AI converts those decisions into a draft-copy skill, which the marketer tests against the next product launch and revises using their edits.
→ The marketer gains a reusable first-draft process grounded in their actual standards and judgment.
A growth lead records themselves collecting campaign data, checking anomalies, selecting meaningful changes, and writing a weekly summary. The recording is transcribed and converted into a skill that specifies sources, calculations, anomaly checks, and reporting structure.
→ Future reports require less manual assembly while preserving the lead's interpretation criteria.
Common mistakes
Automating an Imaginary Process
Designing a skill without observing real work omits tacit decisions and often solves a workflow nobody consistently performs.
Recording Actions Without Reasoning
Capturing clicks but not the reasons behind them produces brittle instructions that cannot handle ordinary variation.
Generating Too Many Skills at Once
A large batch creates AI clutter before the team has tested whether any individual skill improves real work.
Is it for you?
Best for
Small teams and individual knowledge workers who repeatedly perform judgment-heavy digital tasks.
Not ideal for
Rare, highly variable, or safety-critical work that cannot be represented reliably from a small number of demonstrations.
From the transcript
“record the work that you do on a daily basis in video, put that in Gemini or get the transcript and put it in Claude…”
“literally just do work and narrate yourself doing work and you can turn that into skills and workflows.”
From the episode
Perplexity Computer: The Super Agent Playbook (5 Real Workflows)
Perplexity Computer