Which to run first in fault exclusion and site health check in AI skills

In the background AI skills, which should be run first, fault diagnosis and site health check - when there are no exceptions, run the site health check to establish a baseline first, and then proceed with fault diagnosis when exceptions occur.These two skills are often treated as the same, but in fact, the input and output are different: the health check performs a full-site inspection, reads the configuration and status, and produces a list of items to be repaired;Troubleshooting for an already occurred issue involves reading logs and recent changes, resulting in a localization conclusion. The most common result when the order is reversed is - after the health check is completed, a pile of historical leftover items is obtained, but no one knows which item is related to the current fault.

The problems solved by the two types of skills are different

Skill system covers content planning, SEO analysis, batch operations, site health checks, troubleshooting, template development and API development. Health checks belong to 'active patrol', it does not assume that the site has a problem, checking the column structure, publishing configuration, whether the backup is available, whether the static files and attachments are complete, whether the search related switches are set, and whether the permission grouping is too wide. Troubleshooting belongs to 'passive location', it starts from a phenomenon: the page cannot be opened, the release does not take effect, the interface call fails, and the style is lost, and then look back at the recent changes and logs.

Establish a baseline first and then handle exceptions.

The value of a baseline lies in comparison. After a health check, you know what tasks and configurations are running in the normal state of the site; the next time there is an anomaly, just look at the difference between the current phenomenon and the baseline, and you can narrow down the range to the part that was changed most recently. Without a baseline, fault diagnosis can only rely on guessing, and the common practice is to try all possibilities, which takes a long time and is prone to causing a second problem in the process.

Suggested rhythm: After the new station goes online, run a health check first, and separate the results into two categories: 'must repair' and 'can observe'; then determine the cycle according to the amount of column updates, and shorten the interval during dense content periods. No cycle is set for troubleshooting, and it is run when the phenomenon occurs.

How to convert the inspection results into tasks

The checklist provided by the health check should be actionable, otherwise it is just a report. The method to convert it into a to-do list is to specify a responsible person and a judgment standard for each item, for example: the backup item should be written as 'recover once and record the time consumed', the permission item should be written as 'list the publishable user groups and check the recent changes', and the configuration item should be written as 'verify in front of the screen after the modification is complete'. Anything that cannot be written should be returned and asked again, what is this check preventing?'

When processing batch tasks, use batch operation skills to merge similar changes into execution, but check each type individually rather than releasing the entire batch. SEO and content tasks are arranged according to the priority of the column, first handling parts that affect readability and indexability, and then handling the additional configuration.

Why is round-trip confirmation needed for writing operations

When the background capability is opened for AI client calls, it goes through the MCP interface, and the tool surface converges according to the intent directory. The significance of convergence lies in the verifiability of the call scope: what can be done, what cannot be done, with a clear directory rather than scattered interfaces. High-risk areas (such as backup, upgrade, and multi-site management operations at the whole site level) are closed by default and need to be explicitly enabled when necessary.

On this basis, the operation should be approved and merged in rounds: summarize multiple changes that AI plans to make in one confirmation, and people see the result of this batch as a whole, rather than agreeing item by item. Actions such as batch modification of columns and batch replacement of keywords are easy to find that the direction is wrong after several 'agrees'.

The built-in AI Chat intelligent conversation here plays the role of inquiry and explanation: ask it about its current status, let it explain the meaning of an inspection, and analyze log fragments for it. The actual action of changing the site content still requires approval, and the intelligent body can arrange calls, but the decision-making power remains with humans, which is also the reason why the content, SEO, operation and maintenance, and business domain tool sets are managed separately.

Common questions

Will the health check miss any ongoing problems?Yes. It inspects according to preset items and is not responsible for discovering unknown phenomena. When there is a clear abnormality, it goes through fault elimination first, and then adds the exposed inspection items to the baseline.

How often should a health check be performed?Based on the frequency of changes. The period of frequent adjustments of columns and permissions should be shortened, and it can also be quarterly during the stable operation period, as long as records are kept of each change.

Can AI directly perform repairs?Read and low-risk write operations can be performed, and full-site operations must be explicitly enabled and confirmed in rounds. The default status remains closed, and risks are determined by configuration rather than by model judgment.

What should be done when there are more and more to-do lists?Regularly merge similar items, control the number of 'must repair' items within an executable quantity, and convert the rest into configuration adjustments. The length of the list itself is not the goal; it is meaningful only if it can be closed item by item.