Gemini Gems is a wonderful feature, but a lot of people get stuck in "I made some but never use them" mode. I was the same way for the first two months — make a Gem, abandon it, repeat.
The turning point was a simple rule I gave myself: "If I've typed the same prompt three or more times in a week, turn it into a Gem." The corollary is that ad-hoc Gems made on a whim never get used.
Below I share nine patterns of Gems I genuinely run almost every day. For each, I cover why it's worth Gem-ifying, what custom-instruction elements I include, and the operational lessons learned. The aim is something you can copy tomorrow.
Pattern 1: The "Tone Lock" Translation Gem
Asking Gemini to "translate this" gives you a slightly different tone every time. You want formal, but sometimes you get casual; you want plain, you get marketing flavor.
My fix: build a Gem that pins the translation style. For "translate Japanese blog post to English," the custom instructions look like this:
You are a translator who renders Japanese blog posts into English.
- Tone: warm and informative. Never textbook-style or corporate-PR.
- Keep sentences short. Avoid heavy use of conjunctions.
- Leave domain terms untranslated (Claude, MCP, API, etc.)
- Subjective markers like "I think" / "I feel" — drop them and state directly.
- Output the translation only. No commentary.
This alone collapsed the variance in tone. The trick to tone-locking: write more of what you don't want than what you do want. AIs follow prohibitions more reliably than positive instructions.
Pattern 2: The "Domain-Loaded" Gem with Your Own Glossary
If your work has unusual context, you end up explaining it in every prompt. In my case, I was re-describing the structure of all four sites I run, every time.
Solution: I uploaded a Markdown file summarizing each site's structure, categories, key URLs, and editorial direction as a Gem knowledge file. The Gem ("Dolice Labs Assistant") is instructed to "consult the knowledge file before answering."
Now I can ask "what's tricky about adding a premium article on claudelab.net?" and get an answer that already accounts for the differences across the four sites. Skipping the background paste-in is a small but persistent efficiency.
Pattern 3: The "Draft-to-Polished" Email Gem
Many people aren't comfortable writing English emails. I'm not either. My fix is a Gem that takes Japanese bullet points and turns them into a polite English business email.
The custom instructions are short:
The user sends bullet points in Japanese capturing the gist of a message. Render them as a polite English business email.
- Output subject line and body separately
- Body within 200 words
- Don't use stock phrases like "I hope this message finds you well"
- End with a single call to action
- Don't paraphrase the original Japanese into something with a different nuance
Without the "no stock phrases" line, the output reads like generic AI-flavored English. Just that one constraint moves it noticeably toward something a human would actually write.
Pattern 4: The "Screenshot Explainer" Multimodal Gem
Gemini's multimodal capability is a strong card you should be playing. My daily-use Gem here takes screenshots of error screens or code and explains what's happening, plus next steps.
The user sends a screenshot.
1. First describe what's in the image in one paragraph
2. Then list problems (errors, warnings, inconsistencies) as bullets
3. Finally, suggest the three most likely causes and remedies in priority order
Attach a confidence percentage to each guess.
Throw a screenshot, get a structured answer in 30 seconds — initial troubleshooting time collapses. The "what does this warning mean again?" Google-search round-trip is gone.
Pattern 5: The "My Voice" Style-Mimic Gem
Bloggers know the "I can't think of an opening" stall. To dodge it, I have a "write in my voice" Gem with 30 of my past articles loaded as a knowledge file.
You are a ghostwriter for my blog.
Learn the tone, vocabulary, and structure patterns from past articles in the knowledge file, and follow them.
- First person: "I"
- Sentence ending: polite Japanese desu/masu form
- Always open with a concrete personal experience
- Don't stack abstract argument more than three layers deep
- Don't rush to a conclusion
Output: only the first three paragraphs. The rest is written by me on request.
Get the opening, then write the rest yourself. The trick is using the Gem to get past the first wall, not to ghostwrite the whole piece.
Pattern 6: The "Fixed-Lens" Code Review Gem
Specifying review angles every time you ask Gemini to review code is annoying. My fix is a code-review Gem with the angles baked into custom instructions.
The core of the instructions:
The user pastes code. Review it through the following lenses, in order.
- Security (auth, SQLi, XSS, secrets leakage)
- Performance (N+1, redundant recompute, memory leaks)
- Error handling (uncaught exceptions, inappropriate logging)
- Testability (hard-to-mock structure, side-effect handling)
- Readability (naming, function length, comment density)
For each lens, point to specific lines and propose fixes if issues exist; otherwise write "no issues found." End with overall improvement suggestions.
Paste code, hit enter, get a structured review. A side benefit: with the lenses fixed, you miss less.
Pattern 7: The "Decisions and Actions Only" Meeting Notes Gem
Useful for extracting just the actionable parts from meeting transcripts.
The user sends meeting-notes text. Reply in the following format.
## Decisions
- (decisions as bullets)
## Action Items
- Owner: task description (deadline)
## Open Questions
- (carried to next meeting)
Drop chit-chat, introductions, and discussions that didn't conclude.
Don't add information by inference. Use only what's in the notes.
The "don't add by inference" rule matters. Without it, Gemini helpfully appends speculative material. For meeting notes, that's noise — turn it off explicitly.
Pattern 8: The "Periodic Comparison" Benchmark Gem
You can put benchmark URLs and notes into a Gem's knowledge file, then have it compare new submissions against the benchmark.
I have a Gem comparing my own articles against competitors' titles and structures. Drop a planned title in, and it returns "Competitor A goes from this angle, Competitor B emphasizes this keyword, your differentiated angle is around here."
Key custom-instruction line: "Don't end with comparison alone — always close with a unique-angle proposal." Without it, you get a summary, not a recommendation.
Pattern 9: The "Coach" Gem for Yourself
A slightly unusual use: a coach Gem you converse with when you're undecided about something.
You are my coach. Help me untangle whatever I'm hesitating about.
- Don't jump to solutions. Ask me questions first.
- Maximum two questions per turn.
- Don't reflexively agree with my answers.
- Help me separate "facts," "interpretations," and "feelings."
- Maintain that the final decision is mine.
- Don't rush to a conclusion. Assume at least three turns of dialogue.
The defining constraint here is "don't give answers." AI defaults to answering, so you have to apply explicit brakes. With this in place, Gemini behaves convincingly like a coach.
Operating Rules to Keep Gems Alive
Three rules to avoid the "made it, never used it" trap:
"Delete unused Gems after a month." Unused Gems on the dashboard raise the cost of finding the ones you actually use.
"Name Gems with a verb." "Translate," "Format," "Compare" — verbs make function instantly readable. Avoid "Assistant" and "Helper" — too vague.
"Revise custom instructions every three months." Use reveals small frustrations; address them by appending to the instructions. Treat Gems as something you grow, not something you build once and freeze.
If you only build one tomorrow, start with Pattern 1 — the tone-lock translation Gem. Highest payoff, shortest custom instructions, immediately felt benefit.