Case 34
Should This Become a Custom GPT or Skill?
Scenario
You notice you paste the same long prompt every week to prepare client review notes. A colleague suggests turning it into a Custom GPT.
Why this framework
Uses PIN: what it's for, the context and source of truth, and a narrow ask. Enough for most everyday work.
Example context
My weekly client review prompt (current text): "Using the attached account notes, produce a one page client review for my Monday meeting: summary of the week, open issues with owners, upcoming renewals, and three suggested talking points. Use British spelling and a neutral tone." How I use it: I run it every Friday afternoon for about eight clients, so around 8 times a week, for roughly a year. The instructions have not changed in the last four months apart from the headings. Input is account notes pasted from our CRM, which include client names and contract values but no personal data about individuals or payment details. I read and edit every review before the meeting. The worst mistake so far was a wrong renewal date, spotted before the meeting. The prompt never takes actions in any system.
Prompt
Purpose: Help me decide whether my weekly client review prompt should become a reusable assistant such as a Custom GPT or skill.
Context: Use the attached prompt and a description of how often and how I use it.
Narrow the ask: Score it on how often it repeats, how stable the instructions are, what data it touches, and the cost of a mistake. Then recommend: keep as a saved prompt, turn into an assistant, or redesign as an agent that acts in systems. If an agent, tell me to work through BLUEPRINT first.
Materials (pasted here in place of the attachments):
My weekly client review prompt (current text): "Using the attached account notes, produce a one page client review for my Monday meeting: summary of the week, open issues with owners, upcoming renewals, and three suggested talking points. Use British spelling and a neutral tone."
How I use it: I run it every Friday afternoon for about eight clients, so around 8 times a week, for roughly a year. The instructions have not changed in the last four months apart from the headings. Input is account notes pasted from our CRM, which include client names and contract values but no personal data about individuals or payment details. I read and edit every review before the meeting. The worst mistake so far was a wrong renewal date, spotted before the meeting. The prompt never takes actions in any system.Example output from ChatGPT
Weekly Client Review Prompt - Reusability Assessment
Recommendation: Reusable assistant
Consequence: Medium
Your weekly client review prompt is a strong candidate for a reusable assistant, such as a Custom GPT or skill. It runs frequently, the instructions are stable, and a person checks every output. There is no clear reason to build an autonomous agent.
1. Scorecard
Scale: 1 = low, 5 = high. Scores measure the characteristic, not overall quality.
| Dimension | Score | Assessment |
|---|---|---|
| Repetition | 5/5 | Around 8 uses weekly, or approximately 416 runs annually |
| Instruction stability | 5/5 | Consistent for four months, with only heading changes |
| Data sensitivity | 3/5 | CRM-derived client names and contract values; commercially confidential |
| Cost of mistake | 3/5 | Incorrect renewal dates could affect client discussions, but reviews are manually checked |
The repetition and stability scores strongly favour reuse. The confidentiality and error scores favour controlled access, explicit verification rules, and human review.
2. Choose the right approach
| Option | Suitability | Decision |
|---|---|---|
| Saved prompt | Moderate | Works, but requires repeated setup and discipline |
| Reusable assistant | High | Recommended |
| Agent acting in systems | Low | Unnecessary autonomy and integration risk |
A reusable assistant can standardise the review format, British spelling, checks for missing information, and handling of renewal dates.
The important distinction is that reusability does not require autonomy. This workflow needs consistent drafting, not independent action.
3. Important platform consideration
As of November 2026, OpenAI is transitioning from Custom GPTs to plugins. New Custom GPT creation is no longer available on personal accounts, and affected Enterprise workspaces have creation restrictions.
For a new implementation, favour a reusable skill or plugin within an organisation-approved environment rather than investing in a new Custom GPT.
4. Recommended improvements
Before making the prompt reusable, add five controls:
- Source discipline: Use only the supplied CRM notes. Do not invent dates, contract values, or commitments.
- Renewal verification: Reproduce renewal dates exactly and flag conflicting or missing dates.
- Clear ownership: Never invent issue owners or assign responsibilities without evidence.
- Consistent output: Keep the review to one page with the same four sections.
- Human approval: Mark every review as a draft until checked by the relationship manager.
Use only an environment approved for processing confidential client information. The absence of individual personal data does not make contract values public.
5. Proposed reusable assistant instruction
Weekly Client Review Assistant - v1.0
Purpose: Prepare a one-page client review for Monday relationship meetings.
Input: CRM account notes supplied by the user.
Instructions:
-
Use only the supplied notes.
-
Write in British English with a neutral, professional tone.
-
Produce four sections:
- Summary of the week
- Open issues and owners
- Upcoming renewals, including exact dates
- Three suggested talking points
-
Do not invent facts, dates, amounts, owners or commitments.
-
Clearly flag missing or conflicting information as "Requires verification".
-
Distinguish recorded facts from suggested talking points.
-
Keep the review to one page.
-
Never access, change or send information through external systems.
Final check: List any dates, values or unresolved issues requiring human verification.
Status: Draft - requires relationship manager review before use.
6. When would an agent be justified?
An agent becomes worth considering if you later want the system to retrieve CRM records automatically, cross-check renewal dates against contract systems, update tasks, or distribute completed reviews.
At that point, first work through BLUEPRINT, the agent engineering framework, to define behaviour, limitations, interfaces, controls, and memory before granting system access.
Final decision
Build a reusable, human-reviewed drafting assistant now. Do not build an autonomous agent.
Pilot it across your existing eight weekly reviews for four weeks. Measure preparation time, factual corrections, missed renewal dates, and edits required.
The success criterion is simple: consistent, accurate reviews with less manual preparation and no reduction in human oversight.
