Your team is preparing to switch to a new model. You open the prompt used to organize quotation data—the instructions that tell AI how to do the work—and find it full of repeated reminders: check everything again after checking it, and fill in every field. These requirements may have been added one by one after mistakes occurred. Now that you are changing models, should you keep them exactly as they are or rewrite everything?
Intellicon Solutions recommends reviewing prompts whenever you update a model, but acceptance should not be based solely on how much shorter the text becomes. First, distinguish between business rules the company still needs to follow and reminders added to get the previous model to follow instructions. Then use the same data to compare results before and after the revisions. What matters is whether any required conditions have been missed.
First, Separate Company Requirements from Repeated Reminders
According to an INSIDE report, all-caps emphasis and repeated reminders written for older models may increase token usage or even interfere with results when used with a new model. The report introduces the Claude Code command /doctor prompt-audit, which checks for outdated instructions in CLAUDE.md, skills, agents, and commands. It also flags broken paths and conflicting rules across files. Until the user confirms the changes, the tool only lists issues and suggested revisions; it does not modify any files.
That does not mean every strongly worded sentence should be deleted. We recommend starting with a task that someone is actively performing and whose results someone can evaluate, such as organizing quotation data. List the relevant instructions one by one, noting where each is written, the business need it addresses, and why it should be retained. You can begin with the following categories:
- Business requirements that still apply: If company policy requires the currency, tax treatment, and quotation validity period to be specified, retain those requirements. Switching models is not a reason to omit them.
- Checks that need clearer instructions: If an instruction repeatedly asks the model to check its work without specifying what to check, replace it with concrete checks, such as verifying that product names match part numbers or listing items that need confirmation when delivery dates are missing.
- Legacy wording that needs testing before a decision: All-caps text and repeated reminders can go on a list of potential edits or deletions. First, record the errors they were intended to prevent, then test whether those errors return after the wording is removed.
Ask the business owner to confirm the basis for each requirement. If the team cannot explain why an instruction should be retained, mark it as needing clarification. Do not automatically treat it as essential, but do not delete it without confirmation either.
Review Rules Wherever They Are Written
The report notes that rules in different files may contradict one another. Businesses should watch for the same issue when reviewing work instructions. Suppose a skill for organizing quotation data—a reusable procedure—requires missing information to be listed, while the task instructions require every field to be filled in automatically. The team must first decide how missing data should actually be handled, rather than simply deleting a few emphatic phrases.
Keep the review focused on the same task and compile a list of inconsistencies. For each one, record where the rules are written, what conflicts, which business policy should currently apply, and who is responsible for confirming it. If company policy requires checking with the sales team when a delivery date is missing, update both the skill and the task instructions. Specify which missing information must be listed, who should confirm it, and which step cannot be completed until that confirmation is received.
Also verify every folder, template, and operating command referenced in the instructions. If a document has moved or a template has been retired, the references still need updating even when the wording itself is consistent. Keep a change list from this review so that whoever takes responsibility next can track progress, and ask the owner to confirm that every place referencing the same rule has been included.
Use the Same Data to Check Whether Revisions Miss Any Requirements
Keep the original instructions, then have the new model process the same representative data using both the original and revised versions. Apart from the prompt, keep the data, tools, and execution settings as consistent as possible. Otherwise, changing too many things at once may make it impossible to determine why the results differ.
For quotation data, test cases should include complete fields, missing delivery dates, and conflicting information. The business owner should first specify how each situation should be handled, then review the results against the following criteria:
- Have any required conditions been missed? Check that the currency, tax treatment, and quotation validity period are all present. Do not overlook missing information simply because the text reads more smoothly.
- Is missing information accurately identified? When the data contains no delivery date, check that it is flagged for confirmation as required, rather than a date being filled in without a basis.
- Do the format and confirmation steps meet requirements? Check whether colleagues can continue working with the prescribed fields and whether anything requiring confirmation from the sales team is clearly marked.
Run important cases more than once, and retain the input data, instruction version, and output for every run. If the revised instructions omit a requirement, narrow the scope of the changes and test again. Do not replace the instructions currently in use just because one response looks good.
Update Skills and Tasks Only After Testing
Once an approach passes testing, turn it into reusable skills or task instructions rather than leaving it in a single conversation. In EgentWrX, for example, teams can capture established procedures and rules in skills, while using task versioning and publishing policies to manage revised work instructions.
Start by creating a new task version and recording the reasons for the changes and the test results. Then ask the business owner to confirm that all required conditions are covered, and ask IT to verify that referenced paths are valid and operations run correctly. Publish only after these checks are complete. If an employee-created skill is to be made available company-wide, it must also be submitted to the admin review queue and approved before access is opened. The team remains responsible for comparing rules and organizing test cases.
Before the next model change, choose a pilot task and its owner, prepare a rule inventory and representative data, and agree in advance on which test failures would mean keeping the existing instructions for the time being. With this preparation, the team can use the review and test results to decide which instructions to retain, which wording to revise, and when to update the version used in production.
FAQ
Do all existing prompts need to be rewritten when switching to a new model?
No. Start by checking the business purpose of each instruction and why it should be retained. Quotation conditions, data requirements, and confirmation steps that still apply should remain. Repeated reminders and wording tailored to the old model can be added to a testing list. Compare results before and after the revisions, then decide what to edit or remove.
Can instructions be deleted immediately if a review tool recommends removing them?
We do not recommend deleting them immediately. First, ask the business owner to confirm whether the proposed changes could affect compliance with company policy, then arrange testing. A tool’s suggestions cannot replace business judgment. Keep a record of the confirmation, particularly for rules about handling missing data or obtaining managerial approval.
Should business teams or IT be responsible for reviewing existing prompts?
We recommend dividing the review between the business owner and IT, with the AI project owner consolidating the findings. The business owner checks whether the conditions in the instructions and the resulting outputs meet operational needs. IT verifies that referenced paths are valid and operations run correctly. The team then decides whether to publish the new version based on the test results.
References
- When Models Change, Instructions Need Updating Too: Claude Code’s /doctor prompt-audit Checks for Outdated Prompts in CLAUDE.md and Skills — An INSIDE report covering reviews of existing prompts, identification of obsolete content and conflicting rules, and the process of obtaining user confirmation before making changes.



