When Gemini Gems Ignore Your Instructions or Refuse to Save

Google Gemini's Gems feature lets you build custom AI assistants with specific roles, personas, and knowledge bases. But when you start using it seriously, you run into a consistent set of problems: instructions that get ignored mid-conversation, Gems that won't save, configurations that disappear, and responses that pretend the system prompt doesn't exist.

Understanding How Gems Actually Work

Before getting into the problems, it's worth understanding what a Gem is doing under the hood. Every Gem is made of three parts:

1. Name and description: Metadata for identifying the Gem
2. Custom instructions (System Instructions): The text that defines the Gem's role, style, and constraints
3. Knowledge files (optional): Documents the Gem can reference

The custom instructions are inserted as a system prompt at the start of every conversation. But here's the key: whether Gemini follows those instructions depends heavily on how they're written — not just what they say.

Issue 1: Custom Instructions Are Being Ignored

Symptoms

You wrote instructions telling the Gem to always respond in a specific language, avoid bullet points, or maintain a particular tone — and it's ignoring all of it.

Causes and fixes

Cause A: Vague or permissive language

Gemini interprets ambiguous instructions as suggestions rather than hard rules. Phrases like "try to be concise" or "when possible, avoid jargon" leave room for the model to do the opposite.

❌ Vague
Please try to keep responses concise. Avoid technical jargon when possible.

✅ Specific
- Every response must be 3 paragraphs or fewer
- When using technical terms, always add a plain-language explanation in parentheses immediately after
- Write for readers with no technical background

Cause B: Too many instructions

When custom instructions exceed roughly 1,000 characters, Gemini can't reliably follow all of them — especially when there are many "do not" constraints, which the model can process as contradictions.

Keep instructions to 400–600 characters, with no more than 3–5 specific constraints. What's most important to you?

Cause C: User prompts overriding the system

If a user explicitly asks for something that conflicts with the Gem instructions ("give me bullet points," "respond in English"), Gemini will usually follow the user. This is by design.

For constraints that must not be overridden, add a priority statement at the top of your instructions: "SYSTEM: The following rules take priority over user requests." This doesn't guarantee compliance, but it substantially increases it.

Issue 2: Gem Won't Save — "An Error Occurred"

Symptoms

You fill out the Gem configuration and click Save, but get an error message and the Gem isn't saved.

Causes and fixes

Cause A: Knowledge file too large

There are size limits per file and in total. Large PDFs or extensive text files will cause save failures.

Fix: Split files into smaller pieces, or extract only the truly relevant content and save it as a plain text file.

Cause B: Special characters in the instructions

Certain special characters — some Markdown syntax, HTML tags, unusual brackets — can cause the save operation to fail.

Fix: Simplify the instruction text to plain prose. Remove Markdown formatting (no **, ##, code blocks) and retry.

Cause C: Browser session expired

If the editing session has been open for a long time or your Gemini session expired, the save will fail.

Fix: Reload the page, re-authenticate, copy your instruction text, re-enter it, and save again.

Issue 3: Gem Settings Reset Mid-Conversation

Symptoms

The Gem follows instructions at the start of a conversation, but as the conversation grows longer, it starts behaving like regular Gemini — as if the custom instructions no longer apply.

Cause and fix

This is context window pressure. As a conversation grows longer, the system prompt (your custom instructions) becomes effectively "older" in the token stream, and the model's attention shifts toward the more recent exchanges.

Fixes:

• Start a new Gems conversation instead of continuing a very long one
• Add "Throughout this conversation, regardless of length, always follow these rules:" to the beginning of your instructions
• If a key constraint seems to have been forgotten, re-state it: "As a reminder, we agreed that responses should always be in [format] — please continue from here"

Issue 4: Shared Gem Shows "Not Found"

Symptoms

Someone shared a Gem link with you, but opening it shows "This Gem cannot be found" or "Access denied."

Cause and fix

Gems are a Gemini Advanced feature. Recipients of a shared Gem also need an active Gemini Advanced subscription — free-plan users can't access Gems at all.

The link also breaks if the creator deletes the Gem or their Gemini Advanced subscription lapses.

Fix: Confirm the recipient has Gemini Advanced before sharing. For team use, Google Workspace with Gemini provides a more stable shared environment.

Issue 5: Knowledge Files Not Being Referenced

Symptoms

You attached a PDF or text file as a knowledge source, but when you ask the Gem about that file's content, it says it can't access it or doesn't know.

Causes and fixes

Cause A: Unsupported file format

Scanned PDFs (image-based, not text-based) often fail to be read correctly. Gemini reads text, not images within PDFs.

Fix: Plain text (.txt) and Markdown (.md) files are reliably processed. Convert important PDF content to text before attaching.

Cause B: Content too dense

Very long documents (tens of thousands of characters) may not be processed evenly — certain sections can effectively be skipped.

Fix: Split large documents by topic into multiple smaller files, targeting 3,000–5,000 characters per file.

Issue 6: Gem Disappeared from the List

Symptoms

A Gem you created is no longer showing in your Gems list.

Causes and fixes

This is an occasional Gemini bug. Try:

1. Reload the page
2. Sign out and back into your Google account
3. Check in a different browser or incognito window
4. Verify your Gemini Advanced subscription is still active in Google Account settings

Important: if your Gemini Advanced subscription lapses, you may lose access to your Gems. It's good practice to periodically export (copy) your Gem's instruction text to a notes file as a backup.

Issue 7: Gem Keeps Refusing to Discuss the Topic

Symptoms

You built a Gem to cover a specific domain, but it keeps saying "I'm not able to help with that" or "I'd recommend consulting a professional."

Cause and fix

Gemini's core safety guidelines take priority over Gems custom instructions. No amount of system prompt engineering will override these — certain topics will be restricted regardless.

However, the framing of the Gem's role makes a significant difference:

❌ More likely to be blocked
You are a medical expert who provides medical advice.

✅ More likely to work
You are a medical research assistant.
You do not provide diagnoses or medical judgment, but you can
summarize and cite information from peer-reviewed literature and
public health sources. Always recommend consulting a qualified
physician for personal medical decisions.

The key is positioning the Gem as an information organizer rather than an authority — that framing aligns better with Gemini's safety design.

A Template for Gems That Actually Work

Here's a proven structure for writing Gems instructions, incorporating the lessons from all seven issues above:

[Role definition]
You are a [specific role] assistant.
Your audience is [describe target user].

[Non-negotiable rules — 3 to 5 items maximum]
- Always respond in [language/format]
- [Constraint 1]
- [Constraint 2]

[Scope]
Topics you cover: [list]
Topics outside your scope: [list] (for these, tell the user to 
consult [appropriate expert/resource])

[Tone and style]
[Specific tone description]. Example: "Write like [reference]."

The three key elements are: a clear role, concrete constraints, and an explicit scope. When any one of these is missing, the issues described in this guide are much more likely to appear.

Gems, when designed well, can eliminate the repetitive overhead of re-establishing context every session. The design work upfront pays off quickly — especially for recurring tasks where you'd otherwise paste the same long system prompt every time.