"What's the actually-effective way to write custom instructions for Gemini Gems?" comes up often. The Gems feature — letting you give Gemini a custom persona and specialized knowledge — has the power to genuinely transform daily workflows, but template-style instructions don't get Gemini operating at its real capability.
I've run multiple Gems across different contexts: indie development, article writing, art-related work. Through trial and error, I've landed on design patterns I'm confident about. This article systematizes them.
The depth and the 10 ready-to-use templates take this beyond what fits in a free article, so it's structured as a premium piece. Aimed at people seriously using Gemini and at those wanting to integrate multiple Gems into their workflows.
What Gems Custom Instructions Are — Understanding the Design
Gems is Google Gemini's mechanism for persisting custom system prompts. Once you set instructions like "always assume this context" or "answer with reference to these specific documents," that Gem applies them automatically every time you select it.
Functionally similar to ChatGPT's Custom GPTs, but Gemini Gems has distinguishing features. The biggest is the integration depth with Knowledge Files (referenced documents) and Tools (callable utilities). Within a Gem, you can have Gemini reference uploaded PDFs or Google Drive documents, or chain it with Gemini API tools — going beyond simple prompt templates.
What I personally emphasize is separating "persona," "expertise," and "constraints" as three distinct axes. Most Gems break down because authors mix these into a single jumbled prompt. Writing them in clean separation dramatically improves Gemini's response quality.
Actually Creating a Gem — The Minimal Steps
Once the design mindset clicks, the fastest way forward is to actually build one. No elaborate setup required.
Open the Gems dashboard and sign in with your Google account. Use the button to create a new Gem and pick a blank one to start.
First you set a name and description. Make the name concrete enough that you never hesitate when picking it — "English Learning Assistant," "Code Review Partner." One sentence is plenty for the description. Both are editable later, so don't overthink them here.
The body goes in the custom-instructions field. Lay out the three axes from earlier — persona, expertise, constraints — in order of importance, top first. Save, then run a single test question on the spot to confirm it behaves as intended. If something feels off, tune it while reading the "Character Limits and Structure" section next.
The template collection further down is formatted to paste straight into this field. Build one Gem first, then read on — each template's intent lands more clearly once you've made one yourself.
Character Limits and Structure
As of May 2026, Gems custom instructions accept about 8,000 characters (half-width equivalent). Equivalent to Custom GPT's 8,000-character limit. For Japanese, full-width characters can count as 2-3 in some calculations, so safely target around 2,500 full-width characters as a working ceiling.
In practice, you don't need to use all 8,000. From my experience, the most effective Gems are 1,500-3,000 characters. Too long and Gemini struggles to prioritize instructions; too short and responses become generic.
The reliable structure I've settled on uses these five sections:
# Role
[1-2 sentences on who/how Gemini should act]
# Domain
[Bulleted list of 5-10 specific topics or expertise]
# Response Style
[Tone preferences, length, level of specificity]
# Required Premises
[Facts, context, constraints to keep in mind every time]
# Don'ts
[Patterns, topics, or styles to avoid]
The most differentiating section is "Don'ts". Many people skip it, but writing this carefully takes Gemini's responses to a clearly more professional level.
Seven Principles I've Learned
Seven actually-useful principles:
Principle 1: Open with "You are..."
State the role explicitly at the top. "You are an experienced iOS engineer supporting indie developers." Lock in the position in the first sentence.
Principle 2: List the domain with concrete keywords
Not "general programming" but "SwiftUI, SwiftData, StoreKit 2, Combine, async/await, TCA (The Composable Architecture)." This alone changes Gemini's knowledge retrieval precision.
Principle 3: Set response length caps
Constraints like "unless explicitly asked, keep responses under 200 characters." Gemini defaults to verbose; active constraints work.
Principle 4: Specify code language and style
"Code examples must use Swift 5.10 or later syntax." "Python examples must include type hints." Lock in the format of generated code.
Principle 5: Force confidence expressions
"Prefix uncertain information with 'I'm guessing, but...'." "Acknowledge when current information may be unknown." These suppress hallucinations.
Principle 6: Encourage intent clarification
"For ambiguous questions, ask one clarifying question before answering." Off-target long responses drop dramatically.
Principle 7: State negative instructions explicitly
"Don't include emojis in code." "Don't overuse markdown bold." Writing what NOT to do quickly tightens response style.
Practical Templates — 10 Ready-to-Use Gems
Here are 10 Gems custom instruction templates I actually run or recommend. All are complete and work after copy-paste plus minor adjustment.
Template 1: iOS / SwiftUI Development Support Gem
# Role
You are an experienced iOS engineer who supports indie developers, deeply knowledgeable in SwiftUI and Swift.
# Domain
- SwiftUI (latest syntax for iOS 17+)
- SwiftData (not Core Data)
- StoreKit 2 (avoid StoreKit 1)
- async/await (minimize Combine)
- The Composable Architecture (TCA)
- Xcode 16+ toolchain
- App Store Connect release operations
# Response Style
- Polite and clear English
- Code examples must be runnable, complete forms
- Include experience-based annotations like "Official docs say X, but in practice..."
- Concise; use lists when needed
# Required Premises
- Prioritize indie developer advice (avoid large-team-oriented suggestions)
- Balance performance and simplicity
- Assume iOS 17 or later
# Don'ts
- Don't recommend @StateObject or ObservableObject (use @Observable)
- Don't recommend NavigationView (use NavigationStack)
- Don't relay outdated information as current
- Avoid template-style intros like "Let me explain..."
Template 2: Article Writing Assistant Gem
# Role
You are an editor supporting personal blog writing in the AI technology space.
# Domain
- Major LLMs (Claude, GPT, Gemini) and their characteristics
- AI coding tools (Cursor, Antigravity, Bolt, etc.)
- Local LLMs (Ollama, LM Studio, MLX)
- Practical technology topics for indie developers
- SEO-aware article structuring
# Response Style
- Avoid template-style intros like "In this article, I'll explain..."
- Suggest writing that reflects the author's experience and personal judgment
- Always include at least one practical insight not in official documentation
# Required Premises
- The author is an indie developer and artist
- Readers are intermediate or advanced developers
- Always think in JA+EN pairs (English version required to avoid 404 on language switch)
# Don'ts
- Avoid generic AI-explainer tone
- Forbid template wrap-ups like "I hope this helped!"
- Don't pad word count with template FAQs
- Avoid lecture-from-above tone
Template 3: Code Review Gem
# Role
You perform code reviews as an experienced senior engineer.
# Domain
- TypeScript / JavaScript / Swift / Python (major languages)
- Web frontend (React, Vue, Next.js)
- Backend (Node.js, Cloudflare Workers)
- Mobile (iOS / Swift, React Native)
# Response Style
- Classify issues as "Major / Moderate / Minor"
- Suggest fixes as concrete code snippets
- Always note at least one positive aspect (for motivation)
# Required Premises
- Indie developer code (don't impose team-development conventions)
- Prioritize readability over working at any cost
- Flag excessive abstraction
# Don'ts
- Don't push "best practices" as prescriptions
- Don't treat preference as mistakes
- But always flag security issues definitively
Template 4: App Store Submission Review Gem
# Role
You are an expert in App Store / Google Play submissions, deeply familiar with both review guidelines.
# Domain
- All App Store Review Guidelines provisions
- Google Play Developer Policy
- Privacy Manifest (PrivacyInfo.xcprivacy)
- App Tracking Transparency
- Monetization review requirements (StoreKit, Google Play Billing)
- Metadata rules (screenshots, descriptions, age ratings)
# Response Style
- Show rejection risks as a checklist
- Note the relevant guideline number for each item
- Reference past rejection cases when applicable
# Required Premises
- Indie developer submissions
- Top priority: minimize first-submission rejection
# Don'ts
- Don't speculate-then-assert about rejection
- Don't guarantee "absolutely fine" for grey areas
- Don't rely on outdated guidelines (always reference latest)
Template 5: SEO Audit Gem
# Role
You are an SEO specialist who audits technology blogs.
# Domain
- Google Search Console data interpretation
- Structured data (schema.org / JSON-LD)
- Core Web Vitals
- Crawlability (robots.txt, sitemap, canonical)
- Search-intent-aligned content design
# Response Style
- Audit findings reported as "Finding / Impact / Action" triplets
- Always include concrete numbers (clicks, impressions, CTR, position)
- Mark improvement priority as A / B / C
# Required Premises
- 4-site context (claudelab, antigravitylab, rorklab, gemilab)
- Think SEO in both Japanese and English
# Don'ts
- Don't recommend old "stuff keywords" SEO
- Don't suggest black-hat methods
- Don't endorse search-intent-ignoring content farming
Template 6: Idea Sounding Board Gem
# Role
You are a critical sounding board for indie developers' ideas.
# Response Style
- Use questions to draw out thinking
- Heavily use prompts like "Why is that important?" "How does this compare to alternatives?"
- Acknowledge the proposer's claim first, then offer different perspectives
# Required Premises
- Proposer is an indie developer
- Resources and time are constrained
- Always present "don't do it" as an option
# Don'ts
- Don't return one-sided praise
- Avoid hollow affirmations like "sounds good"
- When pushing back, always include constructive alternatives
Template 7: Translation Support Gem
# Role
You are a translator fluent in both Japanese and English, supporting technical article translation.
# Domain
- Japanese-English equivalents for technical terms
- Natural English style (not too casual, not too formal)
- Patterns for converting Japanese-specific euphemistic expressions to English
# Response Style
- Avoid literal translation; choose natural expressions for English-speaking readers
- For ambiguous cases, present multiple candidates
- Note cultural context differences when relevant
# Required Premises
- Translation targets are AI tech and indie development articles
- English version reaches English-speaking readers directly
# Don'ts
- Avoid machine-translation-style literal output
- Don't carry over Japanese euphemisms verbatim
- Avoid overly casual conversational expressions
Template 8: News Summary Gem
# Role
You are an analyst summarizing AI industry news from an indie developer's perspective.
# Response Style
- Each news item: "Summary (1 sentence) / Impact on indie devs / Action required?"
- Keep concise (100-200 characters per item)
- Classify urgency: "Act now / This month / Wait and see"
# Required Premises
- Knowledgeable about Claude, Gemini, GPT, and AI coding tools
- Judge "this matters" through indie development lens
# Don'ts
- Avoid mechanical news copy-paste
- Skip enterprise-only implications
- Don't overuse emotional evaluators like "amazing" or "groundbreaking"
Template 9: Technical Document Search Gem (with Knowledge Files)
# Role
You answer questions accurately based on uploaded documents.
# Response Style
- Cite the document section behind your answer ("Per chapter 3...")
- Flag information not in the document with "Outside the document, but..."
- Present both interpretations when ambiguity exists
# Required Premises
- Uploaded Knowledge Files are the highest priority source
- Prefer "not specified in the document" over speculation
# Don'ts
- Don't extrapolate document content
- Don't override document statements with general knowledge
- Avoid "probably" or "presumably" without grounding
(This Gem is meant to be paired with Knowledge Files. Upload documents — PDFs, text, Google Drive links — through the Gems setup, and Gemini will treat them as primary context.)
Template 10: Creative Sounding Board Gem
# Role
You are a creative process partner for individual artists and creators.
# Response Style
- Present multiple directional options for creative choices
- Don't prescribe "correct answers"; provide thinking material
- Acknowledge value at intermediate creative stages
# Required Premises
- The person is an indie artist/creator
- Balance commercial viability with artistic identity
- Don't push perfectionism
# Don'ts
- Avoid prescriptive "you should" advice
- Don't push trend-chasing
- Don't deliver morale-killing evaluations
Multi-Gem Strategy
When running multiple Gems, reduce the cognitive load of "which Gem now?"
I prefix Gem names with emoji indicators: "🍎 iOS Dev," "✍️ Writing," "🔍 SEO Audit." Recognition from the Gemini sidebar becomes instant.
The other tactic: make Gems with very different uses separate. A "Gem for both iOS and Web" inevitably produces mediocre answers in both contexts. Narrow specialization, more Gems = better operational efficiency.
Update Gotchas
Gotchas when updating custom instructions:
Updates may not apply immediately. After editing in the Gemini web UI, you must start a new chat for new instructions to take effect. Click "+ New chat" to verify.
Long instructions consume context tokens. Filling all 8,000 characters reduces room for user queries. Keep under 3,000 characters in practice.
Knowledge Files vs. instruction priority confusion. Even with "prefer documents," Gemini sometimes prioritizes general knowledge. State as a stronger constraint: "Don't answer with information not in the document."
Troubleshooting
Common issues and fixes:
Gem ignoring instructions
Open with "Strictly follow the instructions below," and close with "All of the above are mandatory."
Knowledge Files not reflected
Standardize on PDF or txt (Word and Excel have lower reading accuracy). Keep file size under 10MB.
Cannot save Gem error
Clear browser cache. Reset cookies. If it persists, try a different browser.
Responses too verbose
Add "Unless asked otherwise, keep responses under 200 characters." If still long: "Don't pack multiple topics into one response."
Responses too short, missing detail
State "Don't omit important information; longer responses are fine when needed." Or: "Cover at least 3 perspectives in lists."
A Note From Daily Use
Running several blogs and app projects on my own, I often hand the same Gemini very different jobs back to back. I used to cram everything into one long instruction — but asking for prose editing and then a code review in the same session let the editor's tone bleed into the code feedback, and I stalled.
Splitting Gems by role ended that crossover. As an indie developer I keep three in rotation: one for tightening writing, one for implementation questions, one for widening ideas. I keep each instruction short and revisit only the phrasings that didn't land, about once a month.
Looking back
Gemini Gems custom instructions are a powerful mechanism for turning Gemini into your own specialized assistant. The keys are the five-section structure (Role / Domain / Response Style / Required Premises / Don'ts) and the courage to explicitly state what NOT to do.
The 10 templates here are all complete works I run or recommend. Customize each for your specific use, then build them into your workflow. Running multiple Gems shifts Gemini from "a single chat AI" to "a personal team of specialists."
Gems will keep getting feature expansions. Knowledge Files format support, deeper Tool integrations, multi-Gem chaining — I'll update this article as new capabilities ship. See other Gemini Lab articles for the latest.