How to Create a Gemini Gem: Persona, Task, Context, Format

You create a Gemini Gem on gemini.google.com by filling Instructions with Persona, Task, Context, Format, and Fences. Persona names who speaks. Task names the repeating job. Context holds stable background. Format lists section headers. Fences are hard limits. This how-to walks paste order, preview and Save steps, and week-one maintenance. It differs from gemini-gems-guide, which explains the full Gem system, and from gemini-gem-prompts, which ships wording examples. You leave with a field checklist, paste-ready blocks, split rules, and optional scaffolding at https://promptmake.net/gemini-gem-generator before you open New Gem. PromptMake outputs config text only. It does not create Gems inside Google, upload knowledge files, or press Save for you.

Before you open New Gem: one-job prep

Pick one weekly job you already repeat in Gemini chat: support reply editing, lesson warm-ups, exec one-pagers, research claim checks, social caption drafts. If you only need a clever one-off ask, skip Gem creation and stay in normal chat.

Write three offline sentences before the builder. Sentence one: outcome in plain words. Sentence two: audience and tone. Sentence three: three fences such as no invented pricing, no legal advice, mark Unknown when facts are missing.

Gather knowledge PDFs only when the job needs house style or FAQs you refuse to paste each session. Ticket IDs and today's customer name belong in the chat message, not in standing Instructions.

Confirm create rights on your Google account. Consumer Gemini allows personal Gem builds as of mid-2026. Workspace availability depends on admin policy. Model picks inside Gems often include Gemini 3.5 Flash for fast drafts and Gemini 3.1 Pro for heavier reasoning. UI labels shift. The field order below stays stable.

Access and naming gates

Ask IT whether knowledge uploads may contain customer data before you attach playbooks.

Name the Gem by job, not vibe: Support Reply Rewriter beats Magic Helper.

If two jobs need different section headers, plan two Gems now. One mega-Gem rots faster than two sharp profiles.

Field 1: Persona — who speaks

Persona is the voice and expertise layer in Instructions. It answers who the Gem pretends to be when a cold chat opens.

Strong persona lines: You are a B2B support editor who writes calm, direct email replies. You are a grades 6-8 math teacher who plans practical warm-ups. Weak persona lines: helpful assistant, creative genius, friendly AI.

Keep persona to one or two sentences. Stack adjectives only when they change output shape. If persona and task fight, fix persona first. A storyteller persona will resist bullet-only exec briefs.

Paste persona at the top of Instructions even when Google does not require a Persona label. Headers help you edit later.

Persona paste-ready examples

Support Gem persona: You are a customer support editor for a B2B SaaS team. Tone is calm, direct, and empathetic without fluff.

Lesson Gem persona: You are a middle-school math teacher who favors short activities and clear exit checks.

Research Gem persona: You are a cautious research assistant who flags missing citations and never invents statistics.

Field 2: Task — what repeats every chat

Task names the standing job on every Gem open. It is the verb chain the Gem runs before you paste today's ticket.

Strong task: For every draft I paste, return two subject options, a revised body, three risk flags, and open questions. Weak task: help with email.

One primary task per Gem. Secondary tasks belong in a second Gem or as a one-time override in chat.

Pair task with persona in adjacent lines. Preview fails when task is vague even if persona sounds smart.

Task paste-ready examples

Support task: Rewrite pasted ticket drafts into on-brand replies with subject line, customer-facing body, and internal note.

Lesson task: From the topic I paste, produce objective, warm-up steps, materials list, differentiation note, and exit check.

Exec task: Compress pasted meeting notes into Bottom line, Decisions, Owners, Risks, Open questions.

Field 3: Context — stable background

Context holds facts that change rarely: audience band, product category, channel, grade level, region policy.

Good context: Our readers are small-business owners in the US. Replies go to email, not chat widgets. Lessons target 45-minute class blocks.

Bad context: today's ticket 8842, customer named Alex, promo ends Friday. That detail belongs in the chat message for this session only.

Context answers what a new hire needs on day one. Keep it short. Expand only when preview shows wrong audience assumptions.

Context paste-ready examples

Support context: Product is project management software for teams under fifty seats. Never promise features not listed in attached FAQ.

Lesson context: District uses Common Core alignment language. Avoid copyrighted worksheet text in full.

Marketing context: Social captions target Instagram and LinkedIn. No medical claims.

Field 4: Format — section headers every time

Format is the contract for reply shape. List headers explicitly so preview cannot drift into essay mush.

Support format: Subject options, Customer body, Internal note, Risk flags.

Lesson format: Objective, Warm-up, Materials, Differentiation, Exit check.

Exec format: Bottom line, Decisions, Owners, Risks, Open questions.

Format fixes are the fastest repair when preview looks fluent but unusable. Add headers before you add more persona adjectives.

Format fences inside the list

State order: headers appear in this order every time.

State brevity: Customer body under 150 words unless I ask for long form.

State bullets vs prose per section when it matters for your team.

Field 5: Fences — hard stops

Fences protect you from confident wrong answers. They belong in Instructions, not in polite chat reminders you retype weekly.

Examples: Never invent pricing or SLAs. Write Unknown when knowledge files omit the fact. Do not give legal or medical advice. Ask once for missing grade band before drafting.

Run preview with a missing-info sample before Save. If the Gem invents policy, strengthen fences in plain words.

Regulated teams should name escalation paths in fences: escalate billing disputes to human review.

Fence paste-ready block

Never invent product features, pricing, or SLA terms. If the answer is not in attached knowledge, write Unknown and list what file would be needed. Do not paste full copyrighted text. Ask one clarifying question when severity or grade band is missing.

Build loop: New Gem to Save in under thirty minutes

Open gemini.google.com. Choose Explore Gems or New Gem. Paste Name and Description for humans. Paste Instructions built from Persona, Task, Context, Format, and Fences. Attach only knowledge files that serve this job. Pick default tools if your account shows Canvas or Deep Research for every chat on this Gem. Preview twice: happy path and missing-info path. Save. Open a fresh chat and run today's real task.

Kickoff detail stays in chat: Draft below. Channel email. Severity medium. Standing behavior stays in Instructions.

After you edit Instructions, start a new chat so the updated block loads. Old threads may keep older behavior.

Step-by-step paste order

Step 1: Name and Description for human search in My Gems.

Step 2: Persona line at top of Instructions.

Step 3: Task line with clear outcome verbs.

Step 4: Context paragraph with stable audience and channel facts.

Step 5: Format header list in exact order.

Step 6: Fence block at bottom in plain words.

Step 7: Attach knowledge files. Name them inside Instructions if preview ignores retrieval.

Step 8: Preview, edit one rule per failure, Save.

When to split into two Gems

Split when formats disagree: email rewrites vs chat snippets need different headers.

Split when audiences disagree: exec briefs vs classroom warm-ups should not share persona.

Split when knowledge files serve different products. Mixed FAQ PDFs blend retrieval.

Merge only when two Gems differ by tone adjectives alone but share sections and fences.

Scaffold drafts on PromptMake before Google paste

When the blank Instructions box stalls you, draft structure at https://promptmake.net/gemini-gem-generator. Describe the repeating job in three messy sentences. Generate once. Edit Persona, Format, and Fences with real product names. Paste into New Gem. Preview with two real samples from last week. Save.

Guests get about three runs per day on generator paths. Free accounts get about five. Quotas are separate per tool path. Keep proprietary playbooks out of public tools when policy forbids it. Redact names and use generic product categories if you only need structure help.

Honest limits: PromptMake does not open gemini.google.com, does not upload knowledge, does not set sharing, does not save Gems inside Google. You finish every production step in Google's builder.

Read gemini-gems-guide for system overview and gemini-gem-prompts for deeper wording patterns. This page owns the how-to field order for how to create a gemini gem.

Common mistakes when you create a gemini gem

Mistake 1: Vague persona with no Format list. You get fluent mush.

Mistake 2: Packing today's ticket into Instructions. Next chat inherits stale details.

Mistake 3: One company-wide mega-Gem. Split by outcome.

Mistake 4: Saving without missing-info preview. Fences fail in production.

Mistake 5: Continuing an old chat after Instructions edit. Open fresh chat.

Mistake 6: Dumping unrelated PDFs into Knowledge. Retrieval blends topics.

Mistake 7: Treating scaffold output as finished law. You own fences and Save.

Mistake 8: Claiming a third-party tool created the Gem inside Google. Only Google's Save button does that.

FAQ

How do I create a gemini gem?

Open gemini.google.com, choose New Gem, fill Name and Description, then paste Instructions built from Persona, Task, Context, Format, and Fences. Attach optional knowledge files, preview with real samples, and Save. Open a fresh chat to use the Gem. PromptMake can draft field text at https://promptmake.net/gemini-gem-generator but does not create the Gem inside Google.

What are Persona, Task, Context, Format, and Fences?

Persona names who speaks and expertise tone. Task names the repeating job on every chat open. Context holds stable background like audience and channel. Format lists section headers every reply must include. Fences are hard limits such as no invented policy and mark Unknown when facts are missing.

How is this different from gemini-gems-guide?

Gemini-gems-guide explains what Gems are, how they load in chat, and how they compare to everyday Gemini. This page is a how-to create checklist with paste order for each field. Read the guide for system context. Use this page when you are at the builder ready to paste.

How is this different from gemini-gem-prompts?

Gemini-gem-prompts focuses on example wording inside Instructions for specific jobs. This how-to covers field order, preview loop, and split rules without duplicating that prompt list.

Do I need knowledge files to create a Gem?

No. Knowledge files help when the Gem must follow house style or FAQs you will not paste weekly. Many sharp Gems run on Instructions alone. Add files only when preview fails without them.

Can PromptMake create my Gem automatically?

No. PromptMake generates English scaffolding for Persona, Task, Context, Format, and Fences. You copy into Google's builder, attach knowledge, preview, and Save yourself.

Which Gemini model should I pick inside the Gem?

Gemini 3.5 Flash suits high-volume editing Gems where you want fast first passes. Gemini 3.1 Pro suits analysis Gems where missing nuance hurts. Pick one default per Gem when the builder allows it.

How do I test a new Gem before sharing it?

Preview with a happy-path sample and a missing-info sample. Save. Run three live chats the first week with real tasks. Log failures by type: format drift, invented policy, wrong audience. Change one rule per failure type.