How to Train ChatGPT to Do What You Want
Getting ChatGPT to follow instructions consistently means replacing vague requests with examples, constraints, and feedback loops. Here is a practical approach that works across tasks.
Updated 4 min read
Getting ChatGPT to consistently do what you want requires three things: specific instructions instead of vague descriptions, examples that show rather than tell, and a feedback loop that turns repeated corrections into rules. Skipping any of these means re-explaining the same thing in every session.
The rules and examples that work are worth keeping. Tools like fenbs store the instructions that produce the right results alongside the lessons from when they did not, so the next session starts from what already works.
Why ChatGPT ignores your instructions
ChatGPT does not ignore instructions deliberately. It follows the most plausible interpretation of what you asked, drawn from its general training. When your instructions are vague, that plausible interpretation often does not match your actual intent.
"Write a professional email" means something different to everyone. "Write a two-paragraph email to a client explaining a project delay, in a direct and non-apologetic tone, without proposing a new deadline" is specific enough to follow. The specificity is what makes an instruction followable.
Replace descriptions with constraints
Descriptions tell ChatGPT what kind of output you want. Constraints tell it the boundaries the output must stay within. Constraints are more reliable because they are easier to test: an output either fits within the boundary or it does not.
| Description (less reliable) | Constraint (more reliable) |
|---|---|
| "Write clearly" | "Use sentences under 20 words. One idea per sentence." |
| "Be concise" | "Keep the response under 100 words." |
| "Sound professional" | "Avoid slang, contractions, and first-person opinion." |
| "Don't be too formal" | "Use 'you' and 'we', not 'one' or 'the user'." |
| "Include the important details" | "Include: the cause, the impact, and the next step. Nothing else." |
Rewrite your most-used instructions as constraints. Go through each one and ask: is this a description of how I want it to sound, or a boundary the output must stay within? Replace the descriptions.
Use examples instead of explanations
An example input and output demonstrates a format faster and more precisely than a paragraph of explanation. If you want ChatGPT to format meeting notes a specific way, paste one set of meeting notes and your ideal formatted output before asking it to format the next one.
One to two examples is usually enough. More than three tends to over-constrain the model and reduce its ability to handle cases that differ slightly from the examples. The example anchors the format; it should not define every variation.
Build a feedback loop that turns corrections into rules
Every time you correct ChatGPT's output, you are identifying a missing instruction. If you remove hedging language from every response, that is a rule: "Never include hedging language such as 'it could be argued' or 'you might consider'." If you always shorten the output by half, that is a constraint: "Keep responses to three sentences or fewer."
Keep a list of corrections you make regularly. Every month, convert that list into explicit instructions and add them to your standard prompt for that task. Over time, your corrections become shorter because the most common problems are already addressed.
Save the instructions and lessons that work
fenbs Skills and prompts stores the current approved instructions for each task alongside the lessons from when earlier versions fell short. A connected assistant reads both at the start of each session, so it follows the constraints you have built up rather than asking you to re-explain them. Create a free account, add your most-used prompt, and record the correction lessons you have been making manually.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
Using custom instructions effectively
ChatGPT offers a custom instructions feature in settings. These are instructions that apply across all conversations. They are a good place for personal preferences that should always apply: your professional role, how you prefer information structured, what you always want excluded.
Custom instructions are not a replacement for task-specific prompts. Instructions that apply to a particular task — formatting rules for a specific document type, constraints for a specific audience — go in the conversation prompt where they can be kept focused on that task. Mix the general and the specific and they tend to dilute each other.
When ChatGPT still does not follow instructions
If an instruction is being consistently ignored, the most likely cause is that it is too vague, it conflicts with another instruction, or it is buried too far into a long prompt. Test each fix separately: first clarify the instruction, then check for conflicts, then move it higher in the prompt.
Some instructions are genuinely difficult to enforce with text alone. Format rules that involve complex structure work better with an example than with a written description. Tone instructions work better with a sample of the target tone than with an adjective.
How to do this with fenbs
- Create a free account and open a project for the tasks where you use ChatGPT regularly.
- Add the current instructions for each task to Skills and prompts: constraints, format rules, and one or two examples if the task has a specific format.
- After each session, note any corrections you made in Lessons learned: what rule was missing, what you added to fix it.
- Each month, review the lessons and update the approved prompts with the rules they identified.
- Connect your AI assistant from Integrations. It reads the current instructions and lessons at the start of each session, so each conversation starts from the most recent version of what works.