Workflow · Audit and reporting
The audit and reporting loop
Real work Rebuilt from a Claude plugin I built for a content team
Every week, one run reads the whole help centre once, checks it against the Content Design System, and publishes one report. Each issue comes with a ready-made prompt, and the report shows what moved since the last run. The only step that changes anything waits for your approval.
Try it
- Select Run the audit to run stages one to five.
- Read the report, and copy a prompt.
- Approve the fix pass, then run next week's audit to measure the movement.
Not run yet
What this workflow cannot tell you
- Whether an article is correct. That needs a person who knows the product.
- Usage for a week until the second run. Views come from the difference between two saved snapshots.
- Whether a flagged word is wrong. A word count is a starting list, and some uses are right in context.
Safeguards
What keeps the workflow safe to run on a real knowledge base.
- The audit is read onlyIt reads the help centre and writes only its own report and snapshot. A scheduled run can never change an article.
- Writing is a separate step, behind approvalThe fix pass only runs when you approve it, and only fixes mechanical issues. It never changes the words.
- One first, then the batchThe fix pass edits one article, checks it, and stops on the first failure.
- It republishes what it editsOn many platforms an edit returns a published article to draft without warning. The fix pass republishes each one and checks it is live.
- Fixes and drift stay separateThe next run counts what the fix pass fixed apart from what people fixed, so the numbers stay honest.
- Checked by lookingA success code is not proof. The run opens every page it publishes and checks it.
- It states its limitsEvery report says what it cannot tell you, so nobody quotes it as more than it is.
Take it to the next level
How to get results that match the best teams, and what to build next.
- Run it on a scheduleRun it weekly and post a short summary to the team channel, with the three tasks that would help most.
- Join it to what readers doFeed in failed searches and questions the chat bot handed to people, so real reader needs become tasks.
- Measure time to fixTrack how long each type of issue stays open, and agree a target with the team.
- Hold new problems at the doorCheck every new or edited article before it goes live, so mechanical issues never reach readers.
- Version the rulesEach check names the rule and version it tests. A rule change then explains a jump in the numbers.
- Check the checkerEach run, a person reviews a sample of flagged items. Show the rate of false alarms on the report, and tune the checks.
Rebuilt from a Claude plugin I built for a real content team. Keystone is an invented payments product, and every name, figure and conversation here is made up.