A skill gives an AI agent a reusable procedure. It can describe how to review campaign copy, supply examples of acceptable work, and include code for parts of the task. The agent application still has to load it, work with the relevant information, and carry out the permitted actions.
That distinction does not make a skill a trivial contribution. If you configure an existing agent app to review your campaigns, you may have built something useful without creating a new model or hosting another service. The app already supplies much of the machinery.
The confusion between skills and agents becomes costly when the promise grows. A saved review procedure, a review you start yourself, and a system that keeps watching campaigns are different commitments. They can share the same skill while leaving very different amounts of work with you.
The Procedure Can Outlive the Conversation
Suppose your team is promoting a webinar. The approved brief describes a 45-minute beginner session, with a recording sent afterward. You want the landing page, invitation email and social posts to make promises that agree with that brief.
In a one-off chat, you could explain what to compare, supply the files and ask for a report. While reading the reply, you might realize that you also need the exact offending passage, its location and the approved wording it contradicts. Without those details, a finding such as inconsistent offer still leaves someone searching through the copy to understand the problem.
Saving that procedure as a skill preserves the standard for the next review. The next campaign can have a different duration, audience or benefit; the requirement to identify and explain discrepancies stays useful. The campaign facts belong in its brief. Otherwise, the reusable procedure can quietly become a way of applying last month’s offer to this month’s work.
An Agent Skill packages instructions in a SKILL.md file, alongside optional scripts, references and templates. That packaging matters because it makes a procedure available beyond the conversation in which you developed it. It also gives you something concrete to inspect and revise when your expectations change.
What Exactly Is Being Reused?
The valuable part is often a decision that previously stayed in the reviewer’s head. Consider an email promising a certificate when the webinar brief says nothing about certificates. Does silence mean the claim is wrong, or that somebody needs to confirm it?
A useful procedure preserves that distinction. The reviewer can identify a promise without authorization, but the supplied evidence does not establish that certificates will never be issued. Automatically replacing the line with a promise that no certificate is available would manufacture a different unsupported claim.
Writing this down does more than save typing. It exposes a choice the team can disagree about before relying on the output:
- One editor may want unapproved benefits removed from drafts.
- Another may want them held for the campaign owner’s decision.
Both approaches can be coherent, but a skill cannot reconcile that disagreement merely by sounding more expert.
Nor does saving the file ensure that every later task uses it. For example, ChatGPT and Codex initially use a skill’s name and description, then load its full instructions when they select it. Having the procedure available and applying it in a particular run are distinct events.
A skill therefore gives you a reusable expression of how the work should be done. It does not establish that the current review read the right brief, applied every criterion or reached a supported conclusion. Those are questions about the work that actually happened.
A practical next step is to test the reusable prompt against fixed cases before switching models.
A Clear Procedure Can Still Require Judgment
The saved procedure is not the model doing the work. In an agent loop, the model can choose a tool action, the surrounding application executes it, and the result becomes information for the model’s next decision. That application and its execution machinery are often called the runtime. A request to read a document and the act of reading that document are different parts of the exchange.
For our webinar, the next action might depend on an unexpected promise in the email. The agent could inspect an approved benefits document, find a certificate policy, then return to the copy with better evidence. If the document is missing or contradicts the brief, the appropriate result may instead be a question for the owner.
Anthropic’s distinction between workflows and agents is useful here: a workflow follows predefined execution paths, while an agent lets the model direct how the task proceeds. Product terminology varies, so the presence of an AI model alone does not tell you how much control it has over the sequence.
Predictable Steps Still Have a Place
A fixed workflow can still use a model. It might always extract the promises from each asset, compare them with the brief, then format a report. An agent could choose to investigate an unfamiliar claim before deciding whether the comparison is complete. The difference concerns who determines the next action, rather than whether one system is intelligent and the other is not.
There are good reasons to retain predictable steps. A required date field can be checked without asking a model to invent an inspection strategy. Whether a phrase such as a session for experienced practitioners agrees with a beginner brief calls for interpretation. A sensible review can combine those approaches instead of treating maximum model discretion as the goal.
Discretion Has an Evidence Boundary
The certificate question shows where flexibility stops helping. More searching might locate an approval that the reviewer did not initially receive. It cannot create the business decision if nobody has made it.
This matters when we describe an agent as owning a task. Ownership could mean gathering the evidence, explaining the discrepancy and getting the question to the right person. It need not mean deciding every unresolved issue itself. For a copy reviewer, asking a well-supported question can be completion of its assignment, even though the campaign still needs a decision.
The opposite mistake is to make the skill so prescriptive that it confuses agreement with identical wording. A 45-minute session and three-quarters of an hour describe the same duration; a short session leaves that duration unspecified. Whether the latter is acceptable depends on the channel’s requirements, not simply on whether the checker can find the digits 45.
The useful tension is between consistent standards and flexible investigation. A procedure should make clear what counts as evidence and what outcome is required. It can leave room for an agent to work out which permitted action will get there. A detailed skill and an adaptable agent are compatible.
More Context Does Not Settle Authority
That investigation raises another question: what information can the agent reach? If the approved brief is already a readable file, the review needs no new integration. If the current version lives in a shared content system, continually exporting it by hand may become the weak point.
Model Context Protocol, or MCP, provides a standard way for an AI application to connect to capabilities supplied by servers, including tools, resources and prompts. It describes that exchange; it does not decide how the application should reason about what comes back.
An Accessible Brief May Be the Wrong Brief
For the webinar team, a connection could make a brief available without another manual export. It would not, on its own, establish that the retrieved file was the approved version. A draft, a proposal and a superseded approval can all be accessible. Even the most recently edited document may not be the one that controls the offer.
That is why adding access can make a review look more convincing before it makes it more dependable. A detailed report can quote a real document accurately and still compare the campaign against the wrong commitment. The missing information is the document’s authority for this campaign, not another page of background material.
The same distinction applies to action. A connection that can read a landing page and one that can update it expose different consequences. A rule telling the agent to report issues helps describe the assignment; it is not equivalent to withholding the ability to change the page. Review and correction can be connected, but that connection deserves a decision about what may change and who may approve it.
The related choice is whether to let an agent operate an existing app or replace it.
A Second Worker Can Repeat the Same Mistake
More agents can help with the volume of evidence. In Claude Code, for example, a subagent has its own context and tool access and returns its results to the main conversation. This lets a side assignment work through material that would otherwise fill the main discussion.
If the webinar campaign also promotes a technical white paper, checking its product claims against long reference documents could be a useful separate assignment. The campaign reviewer needs the resulting evidence and unresolved contradictions, rather than every intermediate search. By contrast, giving each short social post its own worker creates several handoffs around the same small brief.
A separate context is not automatically an independent source of truth. If both reviewers treat the same obsolete brief as authoritative, they can agree for the same wrong reason. A second confident report has added an opinion without resolving the shared assumption.
Delegation becomes more useful when the assignment changes what can be established. Two workers could take different questions:
- One checks the claims against the supplied brief.
- The other investigates whether that brief has been superseded in the approved records.
Those results answer different questions. Their value lies in the evidence returned and how it is reconciled, not in counting two agents as two votes.
These components can also be combined. Claude Code can run a skill in a subagent context. An MCP tool could call a service that itself runs an agent. A skill is therefore not necessarily passive prose, and a tool is not necessarily a simple lookup. The package name alone does not reveal how much work, access or delegation sits behind it.
The Job Is Larger Than the Run
So far, someone has brought the campaign to the reviewer. That person chose the files, started the task and remained available to interpret the findings. A useful amount of coordination still sits outside the agent, even if the review itself requires little supervision.
Now imagine asking the system to check every campaign as it becomes ready for launch. The review criteria may remain unchanged. The new responsibility is to recognize eligible work, use its current inputs, preserve the outcome and make an incomplete review visible to someone who can respond.
A schedule answers when to start. It does not, by itself, answer which campaign should be reviewed. A weekly task can execute its instructions successfully while repeatedly inspecting last month’s folder. Persistent instructions and a current assignment are different kinds of information.
Monday’s Review Does Not Cover Tuesday’s Changes
The passage of time creates another complication. Suppose the agent reviews the invitation email on Monday and the copy owner adds a certificate promise on Tuesday. Monday’s report may have been entirely correct. It no longer establishes that Tuesday’s copy agrees with the brief. A system expected to keep the campaign checked needs a way to relate the reviewed result to the version people intend to send.
This is a more demanding promise than producing a report on request. It needs a rule for what becomes due again when inputs change, as well as a visible distinction between reviewed copy, unresolved questions and work that was never inspected. A quiet system is useful only when you can tell whether there were no issues or no completed review.
Allowing the agent to edit the campaign broadens that promise again. A corrected export is not enough if the invitation waiting in the email platform still contains the old offer. The task has to reach the actual draft people will use, and someone must decide whether the correction is within the agent’s authority. Reliable follow-through depends on that relationship between the report, the approved decision and the saved work.
An On-Demand Reviewer Can Be Enough
None of this makes an on-demand reviewer an inferior achievement. For a team preparing an occasional webinar, a person may be the most practical way to supply the current brief and resolve exceptions. The agent can still do valuable interpretive work within that arrangement. Removing the person from those moments is a separate design choice, with responsibilities that have to move somewhere.
When you save a good skill, you have made part of your team’s judgment reusable. When you build an agent workflow around it, you are deciding how that judgment meets real evidence, permitted actions and changing work. The meaningful claim is what the system will take care of for you, including what happens when it cannot finish. That is what another person would need to understand before trusting it with the next campaign.