Skip to content

Using GitHub Copilot

Jump to bottom
Alexander Refsum Jensenius edited this page Sep 21, 2026 · 2 revisions

GitHub Copilot is an AI assistant that works inside VS Code. It writes drafts, explains what a file does, translates text, and fills in the fiddly parts of frontmatter.

For this site it is more than a convenience. MishMash studies creative uses of AI, and the website is deliberately built with AI assistance and open about it. Using Copilot here is in keeping with the project, provided you declare it and check its work. See Web Philosophy.

You need Editing in VS Code working first.

Getting it free

Copilot has a free tier that anyone can use, with a monthly cap on how much you can ask of it. Through GitHub Education, academic staff and students get the full version at no cost.

If you have not applied yet, go to github.com/education and do it now. Approval takes a few days. Step 3 of Start Here has the detail.

Once approved, check your entitlement at github.com/settings/copilot.

Turning it on

In VS Code, click the extensions icon and search for GitHub Copilot. Install it, and install GitHub Copilot Chat alongside it.

You are already signed in to GitHub from setting up VS Code, so Copilot picks that up. A Copilot icon appears in the status bar at the bottom of the window. If it shows a warning, click it and follow the sign-in prompt.

Two ways to use it:

Suggestions appear in grey as you type. Press Tab to accept, Esc to dismiss.

Chat opens with Ctrl+Alt+I, or Cmd+Alt+I on a Mac. This is the more useful one for writing content.

Prompts that work well here

Copilot is much better when it can see what you are working on. Open the relevant file first, and mention the file by name in the chat.

Frontmatter

Write the frontmatter for a new event in site/_events/. It is MeshUp #12 on machine listening, 5 November 2026, 14:00 to 16:00 Oslo time, on Zoom, belonging to work package 3. Follow the pattern in the other files in this folder.

Getting the timezone offset and the categories right is exactly the sort of thing Copilot is good at, and exactly what people get wrong by hand.

Norwegian translation

Translate site/about/index.md into Norwegian bokmål for site/no/about/index.md. Keep the frontmatter structure, set lang: nb, keep all links and Liquid tags unchanged.

The site publishes machine-translated Norwegian pages and marks them as such. A human must read the result before it goes live.

Understanding what you are looking at

What does site/_includes/events_wp_pages.html do, and what happens if I remove it from a work package page?

This is the best use of Copilot when you are new. Ask before you delete something you do not recognise.

Reading levels

Rewrite this section at the simple reading level for the adaptive content system: short sentences, concrete examples, no jargon. Keep every factual claim identical.

See Adaptive Content for how the levels work and how the blocks are marked up.

Diagnosing a failure

Paste the error from the Actions log into chat and ask what it means. It is usually faster than reading the log yourself. See When Things Go Wrong.

Where Copilot gets it wrong

It is confidently wrong about details that matter here, and the pattern is predictable.

It invents plausible-looking frontmatter fields that this site does not use. Check any field you do not recognise against Adding Events, Adding News or the person template.

It invents links. A URL to a person or project that Copilot supplies may not exist. The link checker in CI catches internal ones; external ones are only audited, not blocked.

It does not know our conventions unless you tell it. Slugs rather than paths in directory cross-links, work packages in categories rather than tags, headings starting at ##, plain prose without inline bold. Name the rule in the prompt, or fix the output.

It invents facts about MishMash. Anything about people, funding, partners or research content has to come from you or from an authoritative source, never from the model.

The rule is simple: Copilot drafts, you decide. You are responsible for what you commit, exactly as you would be if you had typed it.

Declaring AI assistance

Pages on the site carry an "About the page" block in the footer showing who last edited them, where the data came from, and whether AI was involved. When you use Copilot substantially on a page, record it in the frontmatter:

page_about:
  ai_support:
    agent: Copilot
    model_name: Claude
    model_version: "Sonnet 4.6"

For a machine-translated page:

page_about:
  translation:
    automatic: true
    source_url: /about/description/
    model: "Copilot"

The full set of fields is in Page Metadata.

This is not bureaucracy. Being able to show exactly where AI was involved is one of the reasons the site is built on GitHub at all, and the AI colophon is generated from this history. Declaring your use is the price of the experiment being honest.

A light touch, such as Copilot finishing a sentence or reformatting a table, does not need declaring. A drafted page, a translation, or a generated reading-level variant does.

Related

Clone this wiki locally