By Hiten Shah
·
Jun 24
·
The first mile of AI adoption is workflow capture. Every company is trying to figure out where AI belongs. The common answer is to look for workflows. A team opens a doc, lists recurring processes, draws boxes, assigns owners, and tries to decide where automation could help.
That exercise feels responsible. It also puts the hardest step at the beginning.
Someone has to explain the workflow before AI can help with it.
Friday at 10 AM PT, I’m walking through a better starting point with OpenAI’s new Record & Replay capabilities.
and bring one task you do the same way every week.
The reason this matters becomes obvious when you watch experienced people work. Years of learned judgment show up in tabs, old examples, defaults, skipped fields, quality checks, and small pauses.
Useful details hide in motion.
A salesperson preparing for an important call checks the CRM and reads the last conversation. She notices who actually cares, looks for the unstated objection, and finds the promise that never made it into the notes.
Support leads triage escalations by reading account history, product pain, customer language, timing, contract value, and quiet signals. They know which issue will get larger by tomorrow.
Finance leaders preparing board updates decide which movement needs explanation, which metric deserves context, and which question the board will ask before anyone says it out loud.
Work like this has a method. That method rarely survives the first workflow mapping meeting.
Record & Replay matters because it gives companies a way to capture the method while the work is happening.
The person does the task once. Codex watches. The recording becomes a draft of the method. The method becomes a Skill an agent can reuse.
The workflow map can become the output of doing the work.
The Work Already Has a Shape
Every company has repeatable work that is more specific than its documentation.
Sales has a way of preparing for renewals. Support has a way of sorting urgent issues from loud ones. Product has a way of turning customer feedback into priorities. Marketing has a way of reading campaign performance. Finance has a way of explaining what changed in the business.
Some of that work is written down. Much of it lives across docs, Slack threads, templates, old decks, onboarding calls, and the heads of people who have been around long enough to know where the bodies are buried.
AI makes that hidden method more valuable.
An agent can connect to the CRM and still miss the shape of a deal. Searching every support ticket will not guarantee it recognizes the customer who needs attention today. Opening every product doc can still produce a plan that sounds right and misses the actual decision.
Access gets AI into the work. Skills teach AI how the work gets done.
That was the argument in my
article. A company’s AI advantage will come from the work it teaches the model to do well.
Record & Replay answers the next question.
How does the library begin?
Documentation Loses the Working Version
Workflow mapping usually asks people to describe work after the fact.
That creates a problem because people remember the official process better than the working process. Big steps come easily. The small moves disappear.
Ask someone how they publish a page and they might say they update the draft, add the metadata, schedule it, and check the live version.
Watch them do it and you see more.
One old page gets copied because its template behaves better. The URL gets renamed in a particular way because analytics and search depend on it. Preview images get checked because social cards break quietly. Live pages get opened in another browser because logged-in views can lie. Publishing waits for one last field because that field tends to get missed.
Those details are the workflow.
A map drawn in a meeting is often too clean because the work itself is messier and more useful.
Recording Captures the Work in Motion
OpenAI’s Record & Replay in Codex gives people a way to demonstrate a repetitive workflow on their computer and turn that demonstration into a reusable Skill.
The person performs the task as usual. Codex observes the actions and screen context. When the recording ends, Codex drafts a Skill that can be inspected, edited, replayed, and improved.
That changes the starting point for AI adoption.
A person can begin with one task they already know. Recording captures the work as it happens. Codex creates the first draft of the method. Human review fills in hidden preferences, missing context, and judgment calls.
The workflow map can come after the work because the first draft comes from the recording.
This matters because the screen contains clues people forget to mention.
Tab order matters. Search queries matter. Skipped fields matter. File names matter. Final checks matter. Even the moment someone hesitates before clicking submit can matter.
A conversation about a workflow often gets the official version. Recordings can capture the working version.
The Skill Is the Useful Artifact
Record & Replay is easy to misunderstand because the name sounds like a macro recorder.
A recording is only the capture method. The Skill is the useful artifact. It can describe when to use the workflow, name the inputs that vary, include the procedure, capture preferences, call out edge cases, and define the verification step. People can read it and agents can use it.
A process map helps people discuss the work. Skills help agents apply the method during execution.
Companies have spent years documenting work for humans. AI creates demand for documentation an agent can act on. The useful artifact has to be clear enough for people to review and specific enough for an agent to use.
A Skill sits in that gap.
Replay Turns Gaps Into Edits
The first version of a Skill should be treated as a draft.
A weekly publishing workflow can show why.
Someone records the task. Codex watches the draft get opened, the title checked, the URL updated, the metadata set, the preview created, the canonical link confirmed, the post scheduled, and the live page verified.
The generated Skill may get the main steps right and miss the judgment.
Replay reveals the gaps.
Title format might need more precision. Image size might need a rule. Verification may need to include a logged-out browser. The Skill may need an example of a page worth copying and a warning about the field everyone forgets.
Each correction improves the method. The first replay becomes the interview. It asks the work what the documentation missed.
That is a useful way to think about early Skills. They are created through demonstration and strengthened through correction.
The Adoption Loop Gets Smaller
AI adoption benefits when the loop is small enough to run during a normal week.
Record one task. Replay it. Edit the Skill. Use it again.
A company can start with one repeated task. Someone records it. Codex drafts the Skill. The owner replays it on a real input. Corrections become edits. After a few cycles, the team has a reusable method that came out of the work itself.
Big strategy work has its place, especially for critical systems and regulated workflows. Early adoption also needs something people can try without waiting for a transformation program.
A useful AI habit can start with a single recorded task.
That is what I’m covering in the live session.
if you want to see the loop applied to real work.
The First Workflow Should Be Boring
Useful first workflows are usually boring.
Pick tasks that happen often, follow a stable pattern, include a few human preferences, produce a clear output, and create annoyance when repeated manually.
Weekly reports work well here. Publishing workflows, sales call preparation, support escalation reviews, customer research synthesis, release notes, competitive monitoring, board deck updates, and contract review preparation also fit.
These tasks matter because they repeat. Repetition creates enough surface area for the Skill to improve.
Boring work also reveals the company’s operating system. It shows naming conventions and quality standards. It reveals review habits, hidden dependencies, and the rituals that make work feel consistent.
Once those methods are captured, teams can improve them in the open.
The Library Grows Through Circulation
A Skill library becomes valuable when methods circulate.
One person records how they prepare a renewal call. Another replays the Skill on a different account and notices a missing risk check. A manager adds the question that keeps getting forgotten. Then a seller uses the improved Skill and produces a better brief.
The method moves through the team.
People can challenge it, edit it, and raise the floor.
Private Skills save time for individuals. Shared Skills help the company remember how good work gets done.
The same pattern applies across functions.
Support leads capture how they triage escalations. Product managers show how they synthesize feedback. Finance leaders walk through variance reviews. Marketers break down campaign performance. Engineers outline how they prepare release notes. Founders demonstrate how they draft investor updates.
Each Skill captures a slice of how the company thinks.
Over time, the library becomes a living operating manual for agents. The company teaches AI the methods it wants repeated, reviewed, and improved.
Recorded Work Shows Where to Invest
Record & Replay also helps teams decide what deserves more durable automation. Some recorded tasks will stay personal. They save time and require little structure.
Other tasks will become shared Skills because several people need the same method. A smaller set will graduate into plugins, app integrations, and MCP-backed tools once the workflow proves important enough.
Architecture arrives later, after a captured task has proved its value.
This sequencing keeps teams honest. A recorded Skill creates evidence. People can see how often the task repeats, where replay fails, which inputs vary, and which parts need stronger tooling.
The company can invest in workflows that have already shown demand.
Version one can be rough. It needs to make the work visible.
The Skill Becomes a Mirror
A recorded Skill can expose weak process.
Codex may capture a messy workaround. The generated Skill may reveal that three people do the same task three different ways. Replay may fail because the workflow depends on information nobody trusted enough to write down.
Those failures are useful. They show where the company’s method is unclear.
A task that proves difficult to record may be unstable. Disagreement about the generated Skill may point to an undefined method. Repeated replay failure at the same step may signal a need for better tooling or cleaner inputs.
The Skill becomes a mirror.
Companies should treat early Skills as drafts with owners. Good Skills need review, examples, quality bars, and someone close to the work who can decide what the agent should do when the happy path breaks.
Ownership matters because Skills can spread process debt.
Sloppy Skills can make bad work repeat faster. Strong Skills can make good work easier to recognize and easier to improve. The quality of the library will depend on the quality of the captured methods and the discipline of the review loop.
Management Gets a New Artifact
Managers have always taught by watching work, asking questions, correcting the method, and raising standards.
Skills give that management work a new artifact. The correction can become an edit to the method the agent will use next time.
A manager can ask one practical question every week.
What work did we record?
The answer produces evidence. It shows which workflows repeat, which methods deserve reuse, which teams are learning, and which Skills are becoming part of the company’s operating rhythm.
That question is more useful than asking whether people are using AI. Usage can mean anything. A recorded Skill points to work that changed.
The Skill Library Starts With One Recording
AI adoption gets stronger when the company captures the methods behind repeated work.
Models, tools, and access all matter. A company still needs to teach agents how the business works.
Record & Replay gives that teaching a starting point.
Someone performs the task. Codex drafts the Skill. The team edits the method. Reuse turns the method into company memory.
A small recording can carry a large implication.
The workflow map can emerge through practice.
Ask people to show the work.
Record one task. Replay it. Edit the Skill. Let the method circulate.
The Skill Library starts there.
Friday at 10 AM PT, I’m walking through this live with Record & Replay.
Bring one task you do the same way every week. We’ll look at how to record it, turn it into a Skill, and start building the kind of library your AI can actually use.
