Do-Correct-Crystallize Skill Loop
Turn proven AI-assisted work into reusable, continuously improving skills.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 98%
Start with a small task that already occurs inside a real workflow, such as researching a video, producing a hook outline, or adding citations. Let the agent attempt the task, then use human expertise to identify concrete gaps and request corrections. Once the result consistently meets the desired standard, ask the agent to preserve the successful instructions as a reusable skill. Test that skill on a new input rather than trying to perfect its underlying file format by hand. Each subsequent correction becomes an opportunity to update the skill, allowing the workflow to accumulate context, preferences, and quality controls. The output is a reusable operating procedure that requires only a short invocation while retaining the richer instructions learned through repeated execution.
Origin
Extracted from Marketing Against the Grain, where Riley Brown demonstrates how he turns useful Codex sessions into reusable marketing skills and improves those skills through live corrections.
Core principles
- 01Automate granular tasks rather than vague job descriptions.
- 02Prove a workflow through real use before formalizing it.
- 03Preserve human expertise and judgment during refinement.
- 04Evaluate skills by their outputs, not their internal elegance.
- 05Feed corrections back into the reusable instructions.
How to run it
- 1
Select a granular recurring task
Identify one concrete action that appears repeatedly in your work. Avoid starting with a broad responsibility such as “do marketing.”
Pro tip Write down the small tasks performed during an average week.
Watch out A task that is too broad will produce an unfocused skill.
- 2
Let the agent perform it
Give the agent the real input, desired outcome, and any immediately relevant context. Treat the first result as a working draft.
Pro tip Use an authentic current task so the evaluation reflects real conditions.
- 3
Correct against expert judgment
Inspect the output and specify what is missing, weak, inaccurate, or inconsistent with your standards. Ask the agent to revise it.
Pro tip Describe observable gaps instead of merely saying the result is bad.
Watch out Do not outsource taste or domain judgment to the model.
- 4
Crystallize the proven workflow
Once the revised process works, ask the agent to turn it into a named reusable skill. Preserve the instructions, sources, output format, and quality requirements that produced the successful result.
Pro tip Let the agent create the platform-specific skill structure.
Watch out Do not spend excessive time perfecting a skill template before proving the workflow.
- 5
Evaluate on a fresh input
Run the saved skill on a different example and compare the result with the required output. Identify whether it generalizes beyond the original task.
Pro tip Keep a representative test case with a clearly recognizable good result.
Watch out Success on the source example alone may reflect overfitting.
- 6
Feed improvements back
Whenever a new correction should apply in future runs, instruct the agent to update the skill. This turns routine feedback into cumulative operational knowledge.
Pro tip Update the skill immediately while the reason for the correction is still clear.
In the wild
A creator asks an agent to research a product announcement and draft a one-page hook outline. After noticing unsupported research bullets, the creator requests primary-source hyperlinks and then adds a permanent rule requiring citations for every research-based claim. The corrected process is saved as a hook-outline skill and tested on the next episode.
→ Future outlines arrive in the preferred format with research links already included.
A marketing lead has an agent draft a campaign brief, then corrects its audience assumptions, evidence requirements, and approval section. After two successful revisions, the lead saves the process as a campaign-brief skill and tests it on another product launch.
→ The team generates consistent first drafts without repeatedly explaining the full brief structure.
Common mistakes
Automating an entire role at once
Broad instructions conceal the granular steps and quality judgments that make the work successful. Start with one repeatable task instead.
Outsourcing taste to the model
Unreviewed AI output tends toward generic or spammy work. Domain expertise must guide corrections and final quality.
Optimizing the skill format too early
A technically elaborate skill is useless if its real output has not been tested. Prove and refine the workflow before polishing its structure.
Is it for you?
Best for
It is best for people who repeatedly create, research, analyze, or transform similar assets.
Not ideal for
It is not ideal for rare tasks whose outputs cannot be evaluated against a meaningful quality standard.
From the transcript
“I'm constantly asking AI to do something, and then I'm having the AI take whatever I do that's useful and I turn it into a…”
“So, the best way to create skills, get the agent to do something, then correct the agent, get it to do something better, and then…”
“It's about having the agent create a skill, and then having some basic evaluation step”
From the episode
If You're A Marketer, Copy These Codex Skills (Or Stay Behind)