Guiding AI to Do the Work Right, Then Saving the Workflow as a Skill

How much does saving past workflows really help AI in future tasks? A study called Agent Workflow Memory explored turning past operational trajectories into reusable workflows. The result: in WebArena web task tests, the success rate surged by 51.1% over the baseline. Of course, this is a specific metric from a controlled test. It can’t be directly converted into how many hours you’ll save in the office, nor does it guarantee that once you write a Skill, AI will never make a mistake again.¹

For tasks that inherently follow fixed rules and steps, I actually use this method quite often. I start a conversation, guide AI through my process step-by-step, and ask AI to run through it. If AI makes a mistake along the way, I point it out immediately and ask AI to show me its exact thought process.

Only when AI can smoothly complete the task from start to finish with the correct result do I ask AI to write the entire workflow into a Skill. Next time I face the same type of task, I simply instruct AI to execute based on this Skill. Whether it’s reviewing documents, fetching data, performing routine checks, or even verifying materials submitted by team members when managing employees—this method is highly effective.

graph TD;

    A["Human explains workflow steps"] --> B["AI performs a test run"];

    B["AI performs a test run"] --> C{"Is the result as expected?"};

    C{"Is the result as expected?"} -- "Omissions or misunderstandings found" --> D["Correct immediately and clarify conditions"];

    D["Correct immediately and clarify conditions"] --> B["AI performs a test run"];

    C{"Is the result as expected?"} -- "Completely correct" --> E["Write the entire workflow into a Skill"];

Watch AI Do It to Find the Missing Steps

When you first explain steps to AI via text, you’re mostly giving it a general direction. You won’t truly know how AI interprets those words until AI actually gets to work.

Take document checking, for example. The simple instruction “check the data for issues” could be interpreted by AI in multiple ways. Pointing out empty fields is one way, but cross-referencing attachments with the spreadsheet to see if they match is another. If your job actually requires the latter, but AI only highlights the blank fields, then no matter how clean its checklist looks, AI has still missed a huge chunk of the work. This is a hypothetical example, but it perfectly illustrates why you must watch AI do it once.

By having AI show you its process, you can pinpoint exactly what data AI uses to make decisions. If AI clearly only did half the job but wrapped up early, it’s much easier to catch which step was missed. Many people suddenly realize at this stage: There are conditions I’ve long taken for granted and didn’t even mention, and the other party simply doesn’t know them.

This is the moment you can refine vague requests into actionable commands. For instance, if you need a record for future reference after the check, the instruction shouldn’t just be “list the problems,” but rather “mark which file and field have issues, and include the basis for your cross-check.” If a missing attachment blocks the check, you must explicitly tell AI to pause, state what’s missing, and wait until the data is complete before continuing.

How to handle the details depends on your actual work. My habit is to execute and tweak simultaneously until AI gets the entire sequence right. If you want to try this, I suggest pausing during every correction to ask yourself: Was a step missing? Was the criteria unclear? Or is this batch of data just special? Depending on the reason, what you eventually write into the Skill will change.

When Formatting the Skill, Bundle the Corrections Too

Once the conversation is over, it’s time to review the formatted result.

Because you might have tried several approaches earlier or added new conditions halfway through, you must check whether AI actually used your final, approved version when formatting the workflow.

If AI just leaves you with generic fluff like “read data, perform check, output results,” it will likely fall into the exact same trap next time. Why did you have to cross-check an extra attachment last round? Under what circumstances should it declare it can’t make a judgment? These all impact the final outcome. So, when asking AI to format the Skill, directly tell AI to review the places you just corrected and explicitly bake those reusable requirements back into the corresponding steps.

There is another thing to separate: Which parts are fixed practices, and which are just the specific data you handed AI this time?

Details like the file name, date, or person-in-charge of a spreadsheet will likely change with the next batch of work. If you hardcode these into the Skill, AI might fetch the wrong file next time. Conversely, which two sets of data to cross-check, or what to do when something is missing, are usually universal rules that must always be followed. The Agent Workflow Memory study handled this similarly by replacing specific values from single operations with replaceable variables, allowing the extracted workflow to be applied to other tasks.¹

Another study, Learn-by-interact, works backwards from the interaction trajectory between an agent and its environment to deduce instructions, using those interaction experiences to fuel subsequent tasks. They use system-generated data—which differs from my hands-on, step-by-step AI guidance—but the underlying spirit is identical: turning the process of successfully completing a task into a useful tool for next time.²

So, what exactly is a Skill? You can think of it as an “instruction manual” that the tool automatically loads when executing a specific task. As Anthropic officially states, it can include prompts, scripts, or templates, and it even supports version control.³ But remember, formatting a workflow into a Skill does not mean the AI model has been retrained. Next time you need it, you still have to pull out this manual for AI and provide AI with the necessary data and permissions.

Therefore, after formatting a Skill, the most critical part to review is the step you previously got wrong. Were the original corrections retained? If this conversational context isn’t backing it up next time, just looking at this Skill, will AI still know what to base its judgments on?

Swap the Data to Test the Workflow’s Resilience

Even if the task is still called “check data,” what you receive each time might not look identical. A spreadsheet might be reformatted, attachments might be missing, or even the underlying business rules might have just updated. So even with a saved workflow, you still need to confirm the working conditions before starting each time.

In the context of managing employees, you can set very specific checking standards: Are the submitted materials complete? Do the two sets of data match? Where do they need to make additions? If the data on hand simply isn’t enough to make a judgment, the check result should honestly reflect that limitation. You cannot casually criticize an employee’s ability or attitude in a report just because they missed an attachment. How far the data check goes depends entirely on the task scope you defined upfront.

If you want to know whether a formatted Skill is clear enough, I highly recommend grabbing another batch of similar data, opening a new conversation, and asking AI to run through it using the Skill. This acts as an insurance policy for this method, and it also helps catch any requirements that were actually left behind in the previous chat and never made it into the Skill.

During this rerun, keep your eyes peeled on the places that went wrong last time. If AI misses the same cross-check again, it means the requirement for that step wasn’t written in plain enough language and needs to be revised. If a new situation arises, you need to make a judgment call: Is this a universal rule that applies across the board, or is it a special case unique to this data? Never rush to turn every edge case into an ironclad rule for all tasks, or you’ll end up making the workflow too rigid to function.

Conversely, if you confirm that this situation is an exception that will happen again in the future, you should add the corresponding handling method into the Skill. A useful Skill inherently requires iterative updates. Every time you encounter a special situation and add a countermeasure, the entire workflow becomes more complete, and AI will be able to block more trouble for you next time.

My most frequent habit now is to guide AI to complete the work correctly first, and then ask AI to write down the whole experience. If you want to bring this method into your own work, why not pick a document you normally check, follow your usual steps, start a conversation, and let AI try it out. When you see AI miss a field, stop right there, clearly explain your checking method to AI, and then let AI finish the rest.

Further Reading

• The First Step of AI Implementation: Deconstructing the Four-Layer Structure of Tasks

References