Why Your Copilot Instructions Stop Working - Split Behaviour From Project Context

The Situation#

We had a good copilot-instructions.md. Project overview, tech stack, folder structure, coding conventions, security rules, the works. Over 300 lines. Then we added a custom chat mode for package upgrades and a handful of prompt files for bug fixes and features.

And Copilot got worse.

The Problems#

Context window gets full too fast#

Every request now carried:

• the full custom instructions
• the chat mode’s system prompt
• the inline prompt for the task
• plus the actual code we were working on

Then we hit the wall. The model ran out of room to remember the conversation.

What we noticed:

• Had to re-explain the problem multiple times
• Responses got generic and lost all the project context
• Stopped following the response guidelines we defined at the start

Instructions fighting each other#

We set up nice instructions everywhere, then they started contradicting.

Real example we hit:

• copilot-instructions.md says “always use async/await for consistency”
• A legacy prompt file mentions “callbacks for compatibility”
• Copilot gets confused and produces a weird mix of both

The mess we ended up with:

• Code that’s half async, half callback
• Different patterns in the same function
• Copilot switching approaches mid-conversation

Wrong instructions win#

Here’s what we learned the hard way: prompts have a priority order.

1. Whatever you just typed (inline prompts)
2. Your chat mode’s system prompt
3. copilot-instructions.md

Where it breaks: you use a quick prompt to format something, and suddenly it ignores the syntax rules from copilot-instructions.md.

It gets worse with weaker models:

• They follow the latest prompt blindly
• Forget your coding standards
• Skip the security checks you carefully set up

The Solution#

After a few trials, the best approach we found (for now) is to split the prompts apart by what they’re for.

1. Separate instructions from context#

Instead of cramming everything into copilot-instructions.md, split it:

.github
├── copilot-instructions.md     # AI behaviour rules only
├── copilot-project-context.md  # Technical details, stack, structure
└── prompts/
    ├── bug-fix.prompt.md        # Task-specific procedures
    └── add-feature.prompt.md

What goes where:

• copilot-instructions.md: how the AI should behave. Ask first, plan before coding, state confidence, when to stop.
• copilot-project-context.md: what the project is. Tech stack, folder structure, conventions, no-go areas.
• Individual prompt files: the procedure for one kind of task.

Copilot only auto-loads copilot-instructions.md. The context file is read on demand, when the instructions tell the model to go read it.

2. Keep instructions short and focused#

We learned to be brutal about cutting. Our copilot-instructions.md went from 300+ lines to under 120.

What we kept:

• Critical workflow rules (like “never edit code immediately”)
• Security guidelines
• A confidence assessment framework
• Fallback procedures

What we moved out:

• Project tech stack → copilot-project-context.md
• Architecture details → copilot-project-context.md
• Task-specific steps → individual prompt files

3. Cross-reference instead of duplicate#

Instead of repeating info everywhere, each file points at the others.

In copilot-instructions.md:

## 📋 Project Technical Context

For project-specific details (tech stack, architecture, restrictions), read:
→ [copilot-project-context.md](copilot-project-context.md)

In copilot-project-context.md:

## 🤖 AI Behavioral Rules

For AI workflow rules and behavioral guidelines, see:
→ [copilot-instructions.md](copilot-instructions.md)

And at the top of every prompt file:

## Before Starting
- Follow behavioral rules: [copilot-instructions.md](../copilot-instructions.md)
- For project technical context: [copilot-project-context.md](../copilot-project-context.md)

Each file stays focused, one fact lives in one place, and there’s nothing left to contradict.

What It Looks Like#

The trimmed instructions file we settled on (copilot-instructions.md) looks like this:

# Copilot Instructions

## 📋 Project Technical Context
→ read copilot-project-context.md

## 🚨 CRITICAL: NEVER EDIT CODE IMMEDIATELY 🚨
Ask clarifying questions → present a plan → wait for approval → one small change at a time.
For new tasks, write the plan to `.dev-notes/` first.

## AI Confidence Assessment
High (>90%) / Medium (70–90%) / Low (<70%), and what to do at each level.

## Fallback
When stuck: stop, summarise what you know, ask.

Everything about this project (stack, structure, “do not touch src/utils/BLE/”) lives in the context file, and each task procedure lives in its own prompt.

Summary#

If Copilot starts ignoring your rules, it’s usually not the model. It’s a 300-line instructions file competing with a chat mode and an inline prompt for the same context window. Split behaviour from project context, cut both hard, and cross-reference instead of repeating.

References#

• How My GitHub Copilot Setup Evolved (this blog)
• GitHub Copilot Prompt Files (this blog)
• GitHub Docs: Adding repository custom instructions for GitHub Copilot ↗
• VS Code Docs: Custom instructions ↗