Copilot refuses to follow custom mode prompts

[mindplay-dk]

Select Topic Area

Question

Copilot Feature Area

Copilot Agent Mode

Body

I was unhappy with the agent's default process of attempting to write everything at once, and then spending 80-90% of the (very expensive) tokens trying to debug it's own broken code. The results aren't good - because the process is obviously bad. No one successfully or effectively builds anything using this totally unrealistic approach.

Everyone keeps telling me "you must put more effort into the prompts" - I'm providing it with clear specs and plenty of context, but it just doesn't seem to take process directions.

So I decided to try to write my own chat mode. (based on the popular "beast mode" prompt.)

Copilot doesn't seem to follow anything I specify in a custom chat mode? 🤷‍♂️

It seems like there's an internal process that's private to the agent, and apparently that's "secret" or "proprietary", and can't be accessed at all? Who even knows. I tried the Export chat command, and it doesn't even mention the mode that was used - it doesn't provide any insight beyond what you can plainly see in the chat panel anyway, which isn't much.

I tried to gain insight by switching to "ask" mode and performing a kind of "post mortem", and I'm not even convinced it's actually seeing the full .chatmode.md file contents?

Note how it weirdly paraphrases my exact directives, and then generally just seems to do whatever it wants - I strongly suspect it's never actually seen the mode prompt? Does this get "compressed" or modified somehow before being passed to the agent? If it's just going to modify my instructions in ways that aren't even disclosed, how are you supposed to diagnose or work with this feature?

This whole product being closed-source, it seems impossible to gain any understanding of what it's doing, or how, or why.

I've had plenty of success writing custom LLM-based workflows, both via tools and by writing code.

So I don't think I'm a complete dunce at this. 🥴

But when I'm building my own agentic workflows or scripts using LLMs, I can see and control every exchange with the model.

How are you supposed to work with a proprietary, closed-source model that won't tell you what it's doing? 🤷‍♂️

You don't even have proper documentation to explain what it does?

An "agent" can be a gazillion things.

I don't understand how you could possibly be expected to do meaningful work with a product that hides all the important details.

Wasted another sunday on this, and I am none the wiser.

I don't get this product.

For that matter, I don't get how the default modes work for anybody.

"Ask" is pretty useful, "Edit" is sometimes useful for a first draft, but "Agent" is just a black box slop machine? 🤔

[soul-abhi]

Hey, I just wanted to say you're absolutely not alone in this — your post really resonated with me. I’ve been running into similar problems with Agent Mode too.

It feels like the model just pretends to follow structure (like your Section 9), but then paraphrases or invents the content instead of quoting it properly — which makes it hard to know if it actually read anything at all. I also tried customizing .chatmode.md, but the agent didn’t seem to consistently follow it, or at least didn’t give any clear sign that it had read and applied the instructions.

I agree that it’s super frustrating not being able to see what prompt the agent is actually working from. If it's going to modify or ignore the directives we give it, there should at least be a way to verify that and understand why.

Would love to see GitHub give us a transparent mode — something like a debug view to show how the agent interpreted the prompt or what internal structure it's using. Without that, it really does feel like a black box.

Thanks for taking the time to write all this out. You’re not the only one scratching your head over how this is supposed to work.

[Anubhavick]

I ran into the same issue with Copilot Chat in Agent mode. It doesn’t seem to fully follow custom .chatmode.md files. Even when I gave it clear instructions (like quoting section 9 exactly), it paraphrased or ignored them.

From what I can tell, there’s an internal system prompt that takes priority, and there’s no way to see or override it. The .chatmode.md content might be truncated or not passed through as-is.

"Ask" mode works better for direct control. For full transparency and behavior control, I’d recommend using the OpenAI API or building your own agent outside Copilot.

Let me know if you'd like help with that.

@mindplay-dk
marks this as an answer if this help in someway please.

[happymoodhub]

• Keep the custom prompt very short and blunt. (if need help, use an ai from the happymod android hub)
• Avoid layered logic—Copilot seems to skip nested steps.
• Try testing your .chatmode.md with variations in "Ask" mode first.
• Check for invisible formatting issues (especially if copy-pasting).

[SriramBharath-7]

GitHub Copilot Agent Mode is frustrating for power users because you can’t really control or see how your prompts are handled behind the scenes. It often ignores or paraphrases your instructions, and there’s almost no way to audit what’s actually sent to the AI or change its core workflow. The product is closed-source and under-documented, so you can’t dig into its inner workings like you can with your own LLM setups. For complex, process-driven coding, custom workflows or open tools give you much more control and transparency than Copilot’s Agent Mode ever will.

[Anica-blip]

I thought I was the problem and I noticed the same thing... it would 'repeat' exactly what I asked but would add something completely out of context, then deliver whatever it wanted. Back and forth, hours and days, and basically almost reached the final line but only me realizing that I would spend another few hours before it sort itself out, eventually and by accident.

It blamed everything outside even if we went through all the steps, verified and confirmed and after going around, it would try that root again and I would put my foot down... say 'nonsense' :)))) or that's 'gibberish' etc.. believe me, it angered me a lot sometimes so I had to put it back into 'solution' mode. As a non coder I am sure the wool can be pulled over our head until WE pushback as we see all the tactics, IT manuals being thrown at us to pretend its doing 'something' but still nothing works.

I've basically had to 'baby sit copilot' to focus and behave and so far did manage to get some repositories sorted so my knowledge is catching up, not in code but the 'white noise' that is wasting energy and time.

Maybe I am too demanding but its always at the end that I have to find ways to change strategies to make it work the way I imagined it should.

[Rudraprasadd]

You're not missing something obvious. The tool just doesn't yet serve power users who want agentic, transparent development loops.

You can build that — just not inside Copilot Chat (yet). You're better off moving to frameworks that treat you like a developer instead of a consumer.

Let me know if you want help prototyping your own assistant or rewriting your .chatmode.md for better results (to the extent possible).

[mgulati3]

Totally get your frustration, Agent Mode often feels like a black box. Even with clear .chatmode.md instructions, it tends to ignore process guidance and rush through tasks, wasting tokens on self-debugging. You’re not alone in finding it hard to trust or control.

The lack of transparency around how prompts are parsed or modified makes it tough to build anything meaningful. Without visibility, it’s hard to iterate or improve.

Appreciate you laying it all out, hope GitHub takes this feedback seriously. It’s a powerful idea, but right now it’s too opaque to be useful for real workflows.

[mindplay-dk]

I switched to Kilo code, where I can actually preview the whole system prompt.

Man, this thing is night and day vs Copilot.

I tried to question Copilot some more today, and it's just crazy - if you directly add a document and ask it something about that document, it never quotes that document verbatim? It seems to think it's quoting your document, but I strongly suspect it's never actually shown the real document - it seems like it's getting some sort of summary/distilled version of it instead??

That's bizarre. 🥴

Context is the most important thing, and for some reason, Copilot cripples the LLM in favor of, what, saving tokens or something?

No idea. No clue what this thing is doing.

Whatever it's doing, it doesn't work, that's for sure.

Kilo with the same Claude model, same prompt, and same context, completes my sample task in half the time.

Go figure. 🤷‍♂️

[RoFz]

Adding another data point: User-level instruction files also being ignored

I'm experiencing the same issue, but with user-level instruction files (stored in vscode-userdata:/Users/.../Code/User/prompts/*.instructions.md) rather than repository-level .chatmode.md files.

Specific examples from today's session:

1. verify-before-suggest.instructions.md - Requires verifying command-line flags and technical details before suggesting them. The agent violated this multiple times within the same conversation, even after being corrected:

   • Suggested GH_PROMPT_DISABLED=1 and gh config set prompt disabled without verification (both incorrect)
   • Made claims about "known LLM limitations" without providing authoritative sources

2. grammar-check.instructions.md - Requires mentioning noticeable grammar errors politely. The agent completely ignored this throughout a multi-hour conversation with several grammar issues present.

The agent acknowledges the files are loaded - it can read them, reference them, and even apologize for not following them. But it doesn't apply them consistently.

What's particularly concerning:

• The agent violated the same instruction multiple times in succession
• Even after being corrected, it violated the same instruction again moments later
• The files are in context and accessible, but application is unreliable

This suggests the issue isn't just about .chatmode.md being compressed or truncated - even when instruction files are fully loaded and accessible in context, the agent doesn't reliably follow them.

Similar to @mindplay-dk's experience, this makes it extremely difficult to establish reliable workflows or behaviors. The lack of transparency into why instructions are ignored makes it impossible to diagnose or fix.

Would love to see either:

1. A debug mode showing which instructions were actually applied
2. Better reliability in following loaded instruction files
3. Transparency into what's modifying/filtering instructions before they reach the model

[mindplay-dk]

I moved on to OpenCode - your Copilot subscription is officially supported, it takes 10 seconds to set up, and I feel like I'm getting way more value for my tokens here.

(the whole text-based UI thing in OpenCode was off putting to me, but I kept hearing good things, and it really isn't far from a GUI - it accepts mouse input, and there's only one or two keyboard shortcuts you really need to learn. it's free, so definitely worth a try.)

[razali-sudo]

found it helpful