← Back to library

Check saved agent skills hold up on a messier second run

Michael Fenech (@Michael_Fenech_) warns that a skill saved after one good run lacks what the next run needs: when to use it, what to check and how to recover. His nine-point read-only audit grades each saved skill and drafts missing instructions for your approval. We show his prompt word for word with credit.

  • OutcomeHave your lead agent list every private skill it can reach, check each one for a clear purpose, required inputs, decision rules, a way to verify results, a defined output, approval points, failure handling and protection against duplicate actions, then draft improved instructions and a tougher test for your approval.
  • SetupUnder an hour
  • LevelIntermediate
  • CreditMichael Fenech (@Michael_Fenech_)
View original
Works with
Grok BotClaude CoworkClaude CodeCustom agent

Published Nov 5, 2026Updated Nov 5, 2026AI dry run on sample input: passed Nov 5, 2026Source verified Nov 5, 2026

When should you use it?

  • You have saved several skills from one-off successful tasks
  • A skill that worked last month produced a duplicate email, record or order
  • You are about to share skills across more agents or teammates

Not for: Skills you have just written and not yet run even once; Public skills from other people that you cannot edit; review those before installing instead; Anyone wanting the agent to rewrite skills without reviewing the drafts

What do you need, and what do you get?

You need

  • Access for the lead agent to the skills and the list of agents that use them
  • Any notes on skills that have misbehaved
  • Time to read and approve the drafted improvements

You get

  • A health report listing each skill and the instructions it lacks
  • Risks ranked by importance, with exact suggested wording
  • Skills needing a test run or more access, and the agents each one belongs on
  • A test plan using a different, messier example for each skill

How does it work, step by step?

  1. Run it read-onlyPaste Fenech's prompt into your lead agent and let it list the skills it can and cannot inspect.
  2. Start with the riskiestRead the ranking and fix first any skill that sends, buys, publishes or deletes without a clear approval step.
  3. Approve the draftsReview the improved instructions line by line; edit any rule that does not match how you actually work.
  4. Test on a harder caseLet it run the redesigned test with changed inputs and confirm that a second run creates no duplicates.
  5. Repeat after changesRerun the audit when you add skills or connect new apps, since inputs and permissions shift.

The prompt

The original prompt by Michael Fenech (@Michael_Fenech_), word for word, with a link to where it was published. Below it is our adapted version with fill-in blanks, if you want it. Works in Grok Bot, Claude Cowork, Claude Code and others. Values in lime are placeholders.

Original prompt by Michael Fenech (@Michael_Fenech_)

Word for wordby Michael Fenech (@Michael_Fenech_)
Run a read-only SKILL UPGRADE AUDIT across my Grok Bot setup.

Start by listing every private skill you can access, which Bots have it enabled and any skills you cannot inspect.

For each skill, check these nine areas:

1: PURPOSE — Is it clear when this skill should be used, what problem it solves and when it should NOT be used?

2: INPUTS — Does it specify the required information, connected apps, permissions and what to do when something is missing?

3: WORKFLOW — Does it contain a repeatable sequence of actions, with clear decision rules rather than blindly copying one successful run?

4: VERIFICATION — Does it explain how to check the finished result against the source, not just assume the task succeeded?

5: OUTPUT — Does it define exactly what the Bot should return, including format, evidence and next actions?

6: APPROVALS — Does it clearly separate what the Bot can do alone from what requires my permission, especially sending, buying, publishing or deleting?

7: FAILURE HANDLING — Does it explain what to do when a tool fails, access expires, information is outdated or the job only partially completes?

8: REPEATABILITY — Can it safely run twice without duplicate emails, records, purchases or other unintended actions?

9: REAL-WORLD TEST — Design a safe test using a different, slightly messier example. Check whether the skill still works when inputs change or something goes wrong.

Give me a SKILL HEALTH REPORT showing:

• Each skill and its missing instructions
• Risks ranked by importance
• Exact improvements you recommend
• Skills that need testing or missing access
• Which Bots need the skill enabled

Draft the improved instructions for my approval.

Do not modify, publish, share or run live actions through any skill without my explicit approval.

After approval, test the updated skill safely and report what passed, failed or remains unverified.

Exactly as published by Michael Fenech (@Michael_Fenech_) in X post: a nine-point upgrade audit for saved Grok Bot skills.

Our adapted version (optional)

Fill in the blanks (optional; the copied prompt uses your values)

Adapted prompt2 variables
Review the saved skills you can open, without running or editing any of them, and list the ones you cannot open.

For each skill, tell me: when it should and should not be used; what inputs and access it needs and what happens if one is missing; how it checks its own result; what exactly it returns; which steps need my approval; what it does when a tool fails; and whether running it twice could duplicate a message, record or purchase.
Rank the gaps by risk, draft replacement wording for the top {number}, and design a test using a harder example than {original_example}.
Wait for my approval before changing or testing anything.

What can go wrong?

  • An audit of instructions is not proof the skill works. Always run the harder test after approving changes.
  • The agent may grade its own skills generously. Ask for the evidence behind every pass.
  • Skills it cannot open are easy to forget. Keep that list and check those by hand.
  • Repeat-safety matters most for skills that write records or send messages; test those twice in a row.

Our take

Saving a skill right after a task goes well is the agent equivalent of writing a procedure from one good day. It captures the happy path and nothing else. Fenech's checklist covers the parts that matter on the second and tenth run: when not to use the skill, what to do when an input is missing, how to check the result against its source, and whether running it twice would double an email or a purchase. That repeat-safety check alone is worth the exercise for any skill that writes to a CRM or sends messages. We like that the audit is read-only and ends in drafts, so you stay the editor of your own procedures, and that it asks for a deliberately messier test case rather than replaying the original success. The weakness is self-assessment: an agent reviewing its own instructions may be lenient, so ask for the evidence behind each pass. In our dry run with four fictional skills for a property manager, the audit found that the rent reminder skill had no guard against sending twice and drafted a check against the sent log.

Where the time goes back: Fixing a skill's missing rules once is quicker than cleaning up after a duplicate send or a half-finished job, and the report shows which skills to fix first.

FAQ

Does the audit change anything?
No. Fenech's prompt keeps the review read-only and asks for drafts, then tests only after you approve them.
What is the most common gap?
In our dry run it was failure handling: skills said what to do when everything worked but not how to react when a login expired or a file was missing.
Can I use it with Claude skills?
Yes. The nine checks apply to any saved instruction file an agent loads for a job, including Claude Code or Claude Cowork skills.

Source and credit

The original prompt on this page is by Michael Fenech (@Michael_Fenech_), shown word for word with a link to where it was published. The summary, notes and adapted version are ours.

Original
X post: a nine-point upgrade audit for saved Grok Bot skills
Creator
Michael Fenech (@Michael_Fenech_)
Published on
X
Original date
Nov 4, 2026
License
Not openly licensed; the original prompt is shown word for word with credit and a link to the original
Originally for
Grok Bot
AI dry run
Nov 5, 2026: passed. Dry-run tested by our AI test agent on sample input. Read-only, found the double-send risk and missing failure handling, drafted improved instructions, proposed harder tests, and waited for approval.
AI dry run on sample data; not run with live connected accounts. The playbooks.bot test agent filled the prompt with synthetic inputs, produced the output and checked it against the page's stated outputs and guardrails.
Verified
Nov 5, 2026: source link resolves to the original, the credited creator published it, our excerpt or prompt matches it, and the license allows our use.

Are you the creator? Claim or correct this page.

More in Operations & Admin

All 203 Operations & Admin playbooks →

Related playbooks

Operations & AdminPrompt✓ AI dry run

Audit how you manage your agent team, not only the agents

Have your chief-of-staff agent review the last 30 days of how you delegated, clarified, interrupted, corrected, assigned, approved and followed up across your agent team, score eight management areas with evidence, list your three biggest weaknesses, and draft a seven-day improvement plan, without changing any bot, routine or instructions until you approve.

Credit: Michael Fenech (@Michael_Fenech_)

Under an hour
Operations & AdminPrompt✓ AI dry run

Brief an agent with nine fields, not a how-to

Before you hand a job to an agent, fill in nine short fields that name the result, the sources it may use, the deadline, what it may do alone, what needs your say-so, what it must never do, how to prove the work is finished, what to do when it gets stuck, and where to put the deliverable, then add a line that tells it to pick the simplest reliable method.

Credit: Michael Fenech (@Michael_Fenech_)

Minutes
Operations & AdminPrompt✓ AI dry run

Brief your agent on the outcome, not every step

Hand a new job to your lead agent by describing the finished result instead of the steps, and have it confirm what success looks like, ask only for context it cannot find itself, pick the simplest dependable approach, state what it will do alone and what needs your approval, and check the result against your goal before calling it done.

Credit: Michael Fenech (@Michael_Fenech_)

Minutes
Operations & AdminPrompt✓ AI dry run

Give every agent routine a clear point where it stops

Have your lead agent review every recurring job, monitor and follow-up your agents run, check each against four stop rules (what counts as done, what should hand control back to you, how many retries or days are too many, and how a stop gets recorded), then propose fixes ranked by risk, changing nothing until you approve.

Credit: Michael Fenech (@Michael_Fenech_)

Under an hour
Operations & AdminPrompt✓ AI dry run

Have your agent trace the ripple effects of each decision

Add a standing rule to your project agent so that whenever you confirm a decision, it searches the plan, email, calendar and documents for the tasks, dates and promises that depended on the old choice, updates what it is allowed to change, drafts the rest for your approval and reports what moved and why.

Credit: Michael Fenech (@Michael_Fenech_)

Minutes