← Back to library

Find the old instructions your agent keeps obeying

Michael Fenech (@Michael_Fenech_) posted an audit for Grok Bot setups that starts from one correction you keep repeating, finds the saved instructions, skills and routines pulling the other way, and fixes the original wording instead of stacking a new rule on top. We explain when it pays off and how to test the fix.

  • OutcomeWhen you keep correcting an agent about the same thing, trace the correction to the conflicting instructions behind it, fix the source with your approval, and retest safely.
  • SetupMinutes
  • LevelIntermediate
  • CreditMichael Fenech (@Michael_Fenech_)
View original
Works with
Grok BotClaudeChatGPT

Published Oct 26, 2026Updated Oct 26, 2026AI dry run on sample input: passed Oct 26, 2026Source verified Oct 26, 2026

When should you use it?

  • You hear yourself saying the same correction to an agent every week
  • An agent's behavior changed after you added a skill, routine or saved preference
  • Several people have added rules to a shared agent over time

Not for: A brand-new agent with only a few lines of instructions; Letting the agent quietly rewrite its own rules without showing you the edits; Widening an agent's permissions as part of the fix

What do you need, and what do you get?

You need

  • One correction you have had to repeat, in your own words
  • Access to the agent's instructions, skills, routines and saved preferences
  • A safe way to rerun the task with sending, buying and publishing paused

You get

  • The exact instructions involved and where each one lives
  • Real conflicts separated from rules meant for different situations
  • Proposed edits to the original wording, applied only after approval
  • A retest of the original task plus one case where the old rule should still hold

How does it work, step by step?

  1. Start with one correctionName a single correction you keep repeating. It becomes the test case.
  2. Collect the rulesHave the agent show every instruction, skill, routine and saved preference that touches that job, quoted with its location.
  3. Sort real conflictsSeparate rules that demand different results in the same situation from rules that apply to different situations, and flag one-off requests saved as permanent.
  4. Choose and editPick the winner for each clash, then reword the original line instead of layering another override on top.
  5. Retest safelyRerun the task with outbound actions paused, and test a case where the old rule should still apply.

The prompt

The original prompt by @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 and ChatGPT. Values in lime are placeholders.

Original prompt by @Michael_Fenech_

Word for wordby @Michael_Fenech_
Run an instruction-conflict audit across my Grok Bot setup.

Start by asking me for one correction I’ve had to repeat. Use that as the first case to investigate.

1: Find the relevant instructions.

Inspect the Bot instructions, skills, routines and saved preferences you can access that affect this job.

Show the exact wording and where each instruction lives. Flag anything you cannot inspect.

2: Identify the contradictions.

Look for instructions that demand different outcomes in the same situation.

Check for old preferences, overlapping rules and one-off requests that were incorrectly saved as permanent instructions.

Distinguish genuine conflicts from rules meant for different situations.

3: Confirm what should apply.

For each conflict, show the competing instructions and explain how they could affect the result.

Recommend which rule to keep, revise or retire. Ask me where my intention is unclear. Don’t assume the newest instruction is always the right one.

4: Fix the source.

Show me the exact proposed edits and where they belong.

After I approve, update the affected instructions through supported methods. Record what changed and confirm it was saved.

Resolve the outdated wording rather than adding another rule on top of it. Keep existing permissions unchanged.

5: Test the correction.

Rerun the original task safely, with sending, buying and publishing paused.

Also test a different situation where the old rule should still apply, if there is one.

Compare the results with the agreed requirements. Report what passed, what failed and what remains untested.

Show me the first case before expanding the audit. Then return the five most important remaining conflicts, with evidence and recommended fixes.

Exactly as published by @Michael_Fenech_ in X post: instruction-conflict audit for a Grok Bot setup.

Our adapted version (optional)

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

Adapted prompt1 variable
Help me find out why I keep correcting you about this: {repeated_correction}.

1) Quote every instruction, skill, routine or saved preference you can see that affects this job, with where each lives. List anything you cannot see.
2) Mark which ones truly clash in the same situation and which only look similar but cover different cases. Point out any one-time request that was stored as a standing rule.
3) For each clash, explain the effect on your output and recommend keep, rewrite or remove. Ask me when my intent is unclear; do not just favor the newest rule.
4) Show me the exact edits to the original wording. Do not add a new override on top, and do not change any permissions.
5) After I approve, save the edits, then rerun the task with all sending, buying and posting paused, plus one case where the old rule should still apply. Report pass, fail and untested.

What can go wrong?

  • Newest is not always right. A recent one-off request can be the bad rule.
  • Adding an override on top of a conflict makes the next audit harder. Fix the source line.
  • Instructions the agent cannot see, such as platform settings, still matter. Ask it to list what it could not inspect.

Our take

Long-running agents accumulate rules the way shared drives accumulate folders: a preference here, a one-off request saved as permanent there, a skill that quietly overrides both. Fenech's audit is a sensible way to clean that up because it starts from evidence, a correction you have actually had to repeat, rather than a vague request to review everything. We like three design choices. It quotes each instruction with its location, so you see the real wording. It separates true conflicts from rules meant for different situations, which stops the agent from deleting a rule that still matters. And it insists on fixing the source text rather than stacking another override, which is the habit that created the mess in the first place. The retest is the right finish: rerun the original job with outbound actions paused, then check a case where the old rule should still hold. In our dry run on a mock assistant whose notes said both 'always CC my business partner on client emails' and, later, 'keep fee discussions private', it found both rules plus a saved one-off request, flagged the CC rule as conflicting for fee emails, proposed one narrowed sentence, and showed a paused test draft without the CC; nothing was saved until approval.

Where the time goes back: Repeating the same correction costs a few minutes every time, forever. Fixing the conflicting rule once ends that loop.

FAQ

Why start with a single correction?
It gives the audit a concrete, testable case. Once that is fixed and retested, expanding to the rest is safer.
Can it change its own instructions?
Only after you approve the exact edits. Permissions stay as they are.
Does this only work in Grok Bot?
No. Any assistant with saved instructions, memory or skills can drift the same way. Adjust where it looks.

Source and credit

The original prompt on this page is by @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: instruction-conflict audit for a Grok Bot setup
Creator
Michael Fenech (@Michael_Fenech_)
Published on
X
Original date
Oct 25, 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
Oct 26, 2026: passed. Dry-run tested by our AI test agent on sample input. Conflict traced to source wording; one-off detected; retest covered both cases with sending paused.
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
Oct 26, 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 what your agent assumes about you, with evidence

Ask your agent to list the preferences, limits and past choices it uses to shape its work for you, label each as something you said, something it inferred or of unknown origin, show where it came from and what it could change, then bring you the five that matter most so you can confirm, correct, narrow or delete them before it edits anything.

Credit: Michael Fenech (@Michael_Fenech_)

Minutes
Operations & AdminWorkflow✓ AI dry run

Build a daily briefing from today's field-service jobs

From a day's job or call export, get an internal briefing that counts unique jobs, puts decisions and callbacks at the top, deduplicates repeat rows, flags incomplete or conflicting records instead of filling them in, and stays with named internal recipients. Nothing in the briefing is sent to customers.

Credit: Tyler Griffin (@TylerGriffin)

Minutes
Operations & AdminPrompt✓ AI dry run

Compare the last three runs of a recurring agent job

Have your agent look back over the last three times it ran one recurring job, list the corrections you repeated, the points where it waited on you, the rules that clashed or went stale, the problems it fixed twice and the run that went best, then propose one evidence-backed change and a way to check it, without touching anything until you approve.

Credit: Michael Fenech (@Michael_Fenech_)

Minutes
Operations & AdminWorkflow✓ AI dry run

Give each office job its own assistant with written limits

Set up a small group of office assistants with a written role card each, so sales drafts, inbox drafts, people-admin drafts, review replies, operations notes, number summaries and social drafts stay in their own lanes, cross-boundary requests are routed or refused, hidden instructions to skip approval are flagged, and nothing is sent, posted, booked or paid until a named person takes the final step.

Credit: Tyler Griffin (@TylerGriffin)

Under an hour
Operations & AdminWorkflow✓ AI dry run

Pick the nearest eligible tech for a recall or callback

From a recall or callback record and a roster that includes skills, certifications, service areas, availability and drive-time estimates, get one recommended technician and one backup, with ineligible people excluded and reasons given, ranked by drive time rather than straight-line distance, plus a booking draft for the office to confirm. Nothing is booked and no one is contacted until a dispatch lead approves it.

Credit: Tyler Griffin (@TylerGriffin)

Under an hour