Copilot instructions

[greenc-FNAL]

No description provided.

[pcanal]

How does the agent know 'when' it is helpful?

[greenc-FNAL]

Probability trees, token prediction, training, context. It doesn't "know" anything, it just predicts tokens. I know that seems glib, but this is all emergent behavior, and guessing things like this can lead to mistaken assumptions and over-reliance on anthropomorphization.

[pcanal]

I wonder if it makes an actual difference (with or without the when helpful).

[pcanal]

What is the value added of this statement? I.e. Conversely this seems to imply that there is an unnamed set of URLs where even if the user is providing explicitly (as said on line 36) the URL, the agent is not allowed to go and fetch it.

[greenc-FNAL]

It often elaborates on a general point with specific examples, because it's not able to "reason" about whether something is in a particular class or not. It's not an exclusive list, just categorical clarification.

[pcanal]

I am still not sure why a category is needed and whether this statement only allows (for better or worse) to fetch things that are within those categories.

[pcanal]

This script and README seems to yet be merge into the repository, is that correct? Or is it provided/created by another tool?

[greenc-FNAL]

See #50.

[pcanal]

If there is a better alternative, why allow the lesser alternative?

[greenc-FNAL]

Best guess is it's allowing two-space line breaks if they already exist, but not in markdown it writes.

[pcanal]

It probably does not matter much (i.e. your preference wins), I would try to phrase it as follow though:

Suggested change

	- Exception: Markdown two-space line breaks (avoid; use proper paragraph breaks instead)
	- Exception: Markdown use proper paragraph breaks unless the existing text already uses two-space line breaks.

[pcanal]

LGFM

[greenc-FNAL]

@pcanal As a general comment, I'd note that this entire document was authored by Claude, with the exception of a couple of conditionalizing tweaks from me. I would ask Claude to preserve something in its instructions in the best way for it to be able to follow them, and this is what resulted. Occasionally I would point out something it did that would seem to contradict the instructions, and ask it to adjust them so that it doesn't happen again. It's all about weights rather than absolutes, and as humans we can't reliably assess what weight any given word or phrase will have.

[greenc-FNAL]

From Copilot itself:

What are the possible filenames and locations of an instructions file to enable copilot to retain context and e.g. conduct guidelines across sessions, and what is the precedence between them?

I'll check GitHub support documentation for Copilot instruction files so I can give the accurate filenames, locations, and precedence. I'll run a GitHub support search for your exact question now.I searched the GitHub Copilot docs and summarized the supported filenames/locations and how Copilot chooses between them. Below are the possible instruction files, where they live, and the precedence rules (with links to the official docs I used).

What I looked up and will summarize:

• GitHub Docs pages about repository custom instructions and Copilot Coding Agent best practices. See: "Adding repository custom instructions for GitHub Copilot" and "Best practices for using GitHub Copilot to work on tasks."
  Adding repository custom instructions for GitHub Copilot
  Best practices for using GitHub Copilot to work on tasks

Supported instruction files / locations

• Repository-wide instructions (repository-level)

  • .github/copilot-instructions.md — repository-wide custom instructions; when present they are automatically added to requests in the repository. See docs: Adding repository custom instructions for GitHub Copilot.

• Path-specific instructions (repository-level, targeted)

  • Any file under .github/instructions/ whose filename ends with .instructions.md (example: .github/instructions/playwright-tests.instructions.md).
  • These files must include a frontmatter block with an applyTo glob (for example applyTo: "**/tests/*.spec.ts") so Copilot knows which files they apply to. See docs: Adding repository custom instructions for GitHub Copilot and examples in the coding-agent guide: Best practices for using GitHub Copilot to work on tasks.

• Agent-specific instruction files (for Copilot coding agent)

  • AGENTS.md, CLAUDE.md, GEMINI.md — used by Copilot coding agent in addition to repository and path-specific instructions. The coding-agent support notes list these as supported agent instruction filenames. See: Adding repository custom instructions for GitHub Copilot and Best practices for using GitHub Copilot to work on tasks.

• User (personal / local) instructions

  • Personal/custom instructions set in the user’s Copilot settings (via the UI) — these are the user-level instructions. Some IDE integrations (JetBrains) also support a local file named global-copilot-instructions.md for local/personal instructions. See: Adding repository custom instructions for GitHub Copilot.

• Organization-level instructions

  • Organization-scoped custom instructions configured at the organization level (configured through organization settings, not as a repo file). The docs describe organization instructions as one of the instruction sources (see precedence below). See: Adding repository custom instructions for GitHub Copilot.

Precedence (how Copilot picks/combines instructions)

• Priority order: personal (user) instructions have highest priority, followed by repository instructions, and organization instructions are lowest priority. However, Copilot combines all relevant instruction sets and sends them together — so they are merged, but highest-priority sources take precedence where conflicts exist. See: Adding repository custom instructions for GitHub Copilot.
• Repository-level details:
  • Repository-wide (.github/copilot-instructions.md) and path-specific (.github/instructions/**/*.instructions.md) files both apply; path-specific files are applied to matching files/directories using the frontmatter applyTo globs. Both kinds are part of the repository instructions set and are merged for a conversation. See: Adding repository custom instructions for GitHub Copilot.
• Agent-specific files (AGENTS.md / CLAUDE.md / GEMINI.md) are supported for Copilot coding agent in addition to the repo and path files; they are considered part of the repository/agent instruction set used by the agent. See: Adding repository custom instructions for GitHub Copilot and Best practices for using GitHub Copilot to work on tasks.
• Local/global personal file (e.g., global-copilot-instructions.md in JetBrains) acts as the personal/user instructions and therefore is highest priority for that user/IDE. See: Adding repository custom instructions for GitHub Copilot.

Practical notes and recommendations

• Multiple instruction sources are combined and provided to Copilot — avoid conflicting instructions across personal, repo, and org levels to reduce ambiguous outputs. See: Adding repository custom instructions for GitHub Copilot.
• Use .github/instructions/*.instructions.md with applyTo frontmatter when you need path/file-type specific guidance (e.g., different rules for tests vs. components). See examples: Best practices for using GitHub Copilot to work on tasks and Adding repository custom instructions for GitHub Copilot.
• For Copilot coding agent, include AGENTS.md / CLAUDE.md / GEMINI.md as needed to supply agent-specific guidance and environment setup alongside the repo instructions. See: Adding repository custom instructions for GitHub Copilot.

If you want, I can:

• Show example contents for each file type (repo-wide, path-specific with applyTo frontmatter, and agent files), or
• Help you create the specific file(s) for a repo (I can draft .github/copilot-instructions.md and a path-specific .instructions.md with the correct frontmatter).