GitHub Copilot PhD-level tutorial

PhD-level · step 1 of 7

How it works under the hood

General background (not a vendor claim): chat assistants are large language models that read text as tokens and predict the next token, sampling from a probability distribution. Everything they 'remember' in a chat lives in a finite context window. Tools such as web search, file reading or code execution add facts the model didn't have; without them, it answers from training data, which can be out of date.

Why this step matters: Knowing it predicts text rather than looks facts up explains why it can be fluent and wrong at once.

PhD-level · step 2 of 7

API and automation: Custom instructions file

# .github/copilot-instructions.md
This is a Python 3.12 service. Run tests with: pytest -q
Style: small functions, type hints, no new dependencies without asking.
Always run the tests and include the output in the PR description.

• Repository-wide instructions go in .github/copilot-instructions.md and apply to all requests in that repository's context.
• Path-specific custom instructions apply to requests about files that match a pattern.
• You can ask Copilot cloud agent (github.com/copilot/agents) to generate the instructions file for you.

Official docs: GitHub Copilot docs, Add repository custom instructions

Why this step matters: Grounding answers in your own files or live sources cuts made-up details and makes them checkable.

PhD-level · step 3 of 7

Evaluation methodology

Anthropic's prompt-engineering guide puts it plainly: define success criteria and a way to test against them before tuning prompts. That applies to every bot.

1. Write a rubric: correctness, sourcing, format, refusal behaviour.
2. Build a golden set of 20–50 inputs with reference answers.
3. Score blind, ideally with two reviewers; compare pairs (A vs B) rather than absolute scores.
4. Run each item 3 times; answers are non-deterministic.
5. Track regressions after any model or prompt change.

Why this step matters: Without a fixed test set, 'better' just means 'felt better'. Scores let you compare prompts and models fairly.

PhD-level · step 4 of 7

Advanced prompting and agents

• Prompt chaining: one job per call, checked between calls.
• Few-shot examples: show 2–3 input/output pairs in the format you want.
• Tool use: let the model call tools for facts and actions; keep tools narrow and require approval for anything that sends, buys or deletes.
• Protocols like the Model Context Protocol (modelcontextprotocol.io) standardise how assistants connect to tools and data.

Why this step matters: The same model gives much better answers to a specific prompt. This is the cheapest improvement you can make.

PhD-level · step 5 of 7

Failure modes in depth

• Tests that assert the current (buggy) behaviour.
• Silent edits to unrelated files.
• Secrets pasted into chat.
• Credit burn from long agent loops.
• Prompt injection: instructions hidden in a web page or file can steer the model. Treat fetched content as data, not commands.
• Sycophancy: it may agree with a wrong premise. Ask it to argue the other side.

Why this step matters: Knowing how it typically fails tells you where to look when you review its work.

PhD-level · step 6 of 7

Reliability engineering

• Validate structured output against a schema; retry on failure.
• Log inputs, outputs, model and settings so you can reproduce results.
• Make tool actions idempotent and reversible where possible.
• Put a human approval step in front of anything external.

Why this step matters: Retries, fallbacks and logged runs turn an occasionally-right tool into one you can depend on.

PhD-level · step 7 of 7

Reading list (official)

• github.com/features/copilot/plans: plans, AI credits, budgets, alerts
• GitHub docs: custom instructions

GitHub Copilot docs, Add repository custom instructions

Why this step matters: Official docs change; going to the source beats any summary, including this one.