GitHub Copilot ghost text in VS Code keeps proposing a library your project does not use — which documented levers steer it back?
answer
- what is on the desk
- current file and open tabs
- the import line speaks
- a pasted sample primes it
- instruction files do not reach it
basics
~20 sCopilot builds ghost text from the current file and open tabs, so fix the editor: add the real import, open files using the right library, close unrelated ones, add a top comment or sample. Instruction files do not reach it.
solid answer
~40 sVS Code's docs say inline suggestions draw on the **current and open files** in the editor; GitHub's responsible-use page adds that the code around the cursor is combined with snippets from open tabs. So the levers live in the editor: add the import or module reference you actually use (with no import the docs' example suggests `console.log`, with a `Log4js` reference it logs through Log4js), open a file that already uses the right library, close irrelevant ones, add a top-of-file comment, and if Copilot keeps writing an older API, paste a short sample of the newer one and delete it once suggestions follow. What does **not** help: custom instructions such as `.github/copilot-instructions.md` are not taken into account for inline suggestions, and a file under a content exclusion never informs suggestions in other files.
go deeper
Recall that Copilot reads your current file and open tabs, so imports, nearby code and what you have open shape every suggestion.
Explain each documented lever — imports, open and closed tabs, top comment, sample priming — and why instruction files and chat requests do not reach ghost text.
Diagnose a team-wide wrong-library pattern: stale open files, old code the model latches onto, content exclusions, and instruction files mistakenly trusted to steer completions.
Decide what the team writes down where: instruction files steer chat, agents and review, while consistent imports and current code samples are what steer completions.
## Where Copilot's ghost text gets its context GitHub Copilot's **ghost text** — the dimmed suggestion at the cursor — is generated from a prompt Copilot assembles automatically. Two pages describe what goes in it: - VS Code's inline-suggestions page: Copilot "looks at the **current and open files** in your editor to analyze the context". Having related files open helps it see a bigger picture of the project. - GitHub's responsible-use page for inline suggestions: the **surrounding code** at the cursor is combined with contextual information such as **code snippets from open tabs** and sent to the model. That is why a suggestion that keeps reaching for the wrong library is usually a context problem, and the fix is in what the editor shows Copilot, not in a setting. ## The levers the docs name 1. **Imports and references.** The VS Code prompt-engineering guide's example: with no imports, Copilot suggests `console.log` to log a value; once the file references `Log4js`, it suggests logging through that framework. The guide says to set the includes yourself because you know which dependencies — and versions — you want. 2. **Open the right files, close the wrong ones.** GitHub's prompt-engineering page: open any relevant files and close irrelevant files, because Copilot uses the open files to understand your request. A tab left open on an old module that uses the rival library keeps feeding it in. 3. **A top-level comment.** A short description at the top of the file tells Copilot what the file is for. 4. **Meaningful names and scoped comments.** A function called `fetchData()` tells Copilot little; a specific name and a focused comment narrow the body it writes. 5. **Prime with sample code.** Paste a small, correct example close to what you want, let Copilot pick up the pattern, then delete the sample. The guide calls this especially helpful to jumpstart Copilot to a **newer library version** when it defaults to older code. The guide's caveat applies throughout: Copilot latches on to the existing code, so "garbage in, garbage out" — a file full of the old pattern keeps producing it. ## Comments as an explicit hint When context alone is ambiguous, say it in the file. VS Code's inline-suggestions page shows Copilot generating a whole TypeScript class from a comment that lists its properties and methods, and suggests naming the approach in the comment — "use recursion", "use a singleton pattern". For the wrong-library problem the comment can name the library outright: a line such as "parse dates with the project's date helper" right above the function is read as part of the current file, unlike a rule kept in a separate instruction file. ## Levers that do not reach ghost text | What a developer tries | Does it change ghost text? | Why | |---|---|---| | Writing "use our HTTP wrapper" in `.github/copilot-instructions.md` | No | VS Code's docs: custom instructions are not taken into account for inline suggestions as you type | | Asking in Copilot Chat to prefer the library | Not a documented input | The inline-suggestion pages name the current file and open tabs, not chat history | | Running **Change Completions Model** | Not reliably | It swaps the model; the model still sees the same open files | | Keeping the right file open while it is under a **content exclusion** | No | Excluded content does not inform inline suggestions in other files | Instruction files are still worth writing for Chat, agents and code review — but a team that adds a "never use library X" rule and expects ghost text to obey it will be disappointed. ## A worked diagnosis A developer's suggestions keep importing a date library the service dropped last quarter. Going through the levers in order: - The file has no import yet, so Copilot falls back on whatever the open tabs show; two old test files with the retired import are open. - Closing those tabs and adding the current library's import line changes the next suggestion. - For an API whose newer version differs from the older one Copilot keeps producing, a five-line pasted sample, deleted afterwards, finishes the job. ## Limits worth stating - These levers **shift probabilities**; they do not guarantee a library. Reviewing what you accept stays your job. - The docs describe the inputs at the level above; they do not publish a ranking of how much each open tab counts, so do not claim one. - If suggestions stop entirely in some files rather than going wrong, the cause is different — per-language settings, Snooze or a content exclusion.
- Copilot keeps writing the previous major version of a library's API. What does VS Code's guide suggest?Prime it: paste a small sample of code written against the newer API into the open editor, let suggestions start following it, then delete the sample. The guide names this as a way to jumpstart Copilot to a newer library version when it defaults to older suggestions. Adding the correct import helps too.
- Why might keeping a reference file open fail to help when it sits in a sensitive folder?If that file is covered by a Copilot content exclusion, its content does not inform inline suggestions in other files, whether or not it is open. The status-bar icon in the excluded file shows a diagonal line, and hovering it names the setting that applied the restriction.
Copilot's ghost text is like a pair programmer who can only read the papers spread on your desk: put the right documents on it and clear the stale ones, because the team handbook in the drawer — your instruction file — is never read while you type.
- Adding a library rule to .github/copilot-instructions.md fixes Copilot's ghost text.
- Copilot ghost text only ever reads the line you are typing.
- Open editor tabs have no effect on what Copilot suggests inline.
- Changing the completions model is the main fix for a wrong-library suggestion.
- An excluded file still informs suggestions as long as it is open in a tab.