Turn Claude Sessions Into Skills With a Daily Audit

Open Claude Cowork and start a new task or chat thread. Keep this focused on one job: auditing your recent AI work for skill opportunities.

The goal is not to make Claude invent a giant list of possible skills. The goal is to find repeated friction. If you keep correcting the same output, rewriting the same prompt, or explaining the same workflow, that is the kind of thing that might deserve a skill.

**Pro tip:** You can also run this in Claude Code or Codex. In Codex, start a new thread outside any specific project, then paste the audit prompt there. Keeping it outside a project helps make the audit personal instead of tied to one repo.

Use this prompt:

Scan my sessions from the past 7 days. If any personal skills should be created or improved, suggest them with evidence. Personal skills only, no repo skills. Prefer no change unless there is a clear reason. Show me what you would create or update before making changes.

The key phrase is "prefer no change unless there is a clear reason." Without that, the assistant may try to be too helpful and suggest skills for one-off tasks.

What you want back is a short audit with evidence. A useful answer should tell you what happened repeatedly, what skill might help, and whether the best move is to create something new, update an existing skill, or do nothing.

In the example audit, the assistant found two concrete moves:

• Update an existing `weekly-marketing-review` skill because the working drafts consistently required extra sections that the skill did not yet ask for.
• Delete a duplicate quick-reference skill that competed with the canonical `SKILL.md`.

It also skipped ideas that did not have enough evidence. That is a good sign. A skill audit should not create busywork.

Do not approve every recommendation automatically.

Approve only the ones that meet at least one of these tests:

• You repeated the same task several times.
• You corrected the same kind of output more than once.
• The assistant found a clear gap in an existing skill.
• The workflow is likely to happen again.

Reject anything that feels like a one-off task. A single unusual request does not need a skill. It might just need a note, a template, or nothing at all.

Also watch for scope creep. A good skill is narrow. It should have a clear trigger and a clear job. If the recommendation sounds like "make a skill for all of my marketing work," ask Claude to make it smaller before you approve it.

Once you like the audit process, ask Claude Cowork to turn the audit itself into a skill.

Use this prompt:

Turn this skill audit process into a personal skill.

The skill should:
- scan recent sessions or recent work
- look for repeated friction, repeated corrections, and repeated workflows
- recommend personal skill updates only
- avoid repo skills unless I explicitly ask
- prefer no change unless there is clear evidence
- show proposed changes before editing files
- report what it created, updated, skipped, or rejected

Keep the skill narrow. It should audit skill opportunities, not manage every part of my workspace.

This is the part that makes the workflow reusable. Instead of remembering the full audit prompt every week, you can trigger the skill and get the same conservative review process.

If Claude suggests a broad skill, tighten it. The skill should not rewrite your whole setup. It should inspect recent work, recommend improvements, and wait for approval.

After the skill exists, set up an automation to run it weekly or daily.

A weekly schedule is usually enough. Daily can work if you use AI heavily and want more frequent cleanup, but the audit should still be conservative.

Use a prompt like this:

Set up an automation that runs this personal skill audit once a week.

Rules:
- Review recent sessions or recent AI work.
- Recommend personal skill changes only.
- Do not edit repo skills.
- Do not create or update anything unless there is clear evidence.
- Tell me what you recommend before making changes.
- It is okay to report that nothing should change.

In Claude Cowork, you can open the Scheduled Task tab after the automation is created. From there, adjust the timing, change the cadence, or move the task to a lower-cost model like Sonnet if the audit does not need your strongest model every time.

This is also where you can tune the task. If it runs too often, move it to weekly. If it produces too many recommendations, add a rule that it can recommend at most one new skill and one existing skill update per run.