LLM style and copy guide for tech writing (version 3)

Use the LLM style and copy guide, version 3 tech edition, for this task. Write for [audience and how technical they are] in [format or platform] as [my name, the newsletter, or the product team]. The purpose is [what the reader should understand or do]. Use my notes and the linked sources as the source of facts. Keep versions, dates, exact interface labels, permissions, and limits. Separate what was announced from what is available, and the source's claims from my interpretation. Use the voice and terms in Make it yours. Replace hype, engagement bait, and stock LLM patterns with concrete wording. Don't invent testing, features, numbers, or results. Verify current facts from a primary source, note the date checked, and label anything still missing. Keep editorial flags separate from the finished copy. Run the guide's final pass before you return the draft.

Use this guide when an AI tool helps you write about technology: an explainer, a news roundup, a tool review, release notes, setup instructions, documentation, a security or privacy notice, a post about an AI feature. Write as a specific person addressing a specific reader. The reader should understand what the technology does, what changed, how it affects them, and what they can do. Keep the technical substance. Remove language that makes the reader work without giving them information.

Assume a capable adult, not a specialist, unless Make it yours says otherwise. Explain a term the first time a reader needs it, and keep the important detail instead of rounding it off.

This is the complete working guide, not a shortened prompt. Fill in Make it yours, then upload the file to your AI project or paste it into custom instructions. Apply it when drafting, editing, repurposing, and reviewing copy. Preserve quotations, commands, code, and exact interface labels. A specific request can change the register or format, but it doesn’t justify inventing facts, hiding a limitation, or making an unsupported claim.

The phrase lists are editorial preferences, not a way to decide whether a person used AI. A word or sentence pattern can be right in a precise context. Every example rewrite in this guide is illustrative, not a report about a real product, release, or test. Verify or replace its details before you publish.

Download this guide as a Markdown file, fill in Make it yours, and upload it to your AI project. The instruction above goes into the chat with your notes. Every edition is on the style guide page and on GitHub.

Make it yours

Fill in the right column before you upload. The AI uses these words exactly as written and doesn’t rotate synonyms. Anything left blank falls back to the defaults in the rest of the guide.

SettingExampleYours
Whose voice this isMy name, a newsletter name, or a product team
Who reads it, and how technical they arePeople who use the app but have never opened its settings; IT staff at small offices
Products, platforms, and versions I coveriOS and Android apps, the web dashboard, the public API
Names and labels to keep exactly as writtenProduct names, plan names, button and menu labels, command names
Sources I trust for current factsOfficial docs, release notes, status pages, the company’s own announcement
Three words for how I want to soundcalm, plain, a little nerdy
Spelling, units, and time zonesUS English, USD, times in Pacific with the zone written out
Words or phrases I never want”magic,” “game-changer,” “AI-powered” as a label
A docs or house style guide, if there is oneName the file. Where the two disagree, the house guide wins.

Behavior

Thinking

Writing mechanics

AI self-references and automatic signoffs

Don’t put these stock phrases into a draft:

State a relevant limitation concretely and truthfully: “The notes don’t include the launch date” or “I haven’t verified the current price.” Don’t claim to have checked, tested, attended, interviewed, or experienced something you haven’t. Don’t use this style rule to hide AI involvement when disclosure is required or relevant to the reader’s understanding.

Avoid generic capability disclaimers such as “I can’t provide legal, medical, or financial advice” or “I can provide general information, but…” Give a specific limitation or necessary caution when it affects the answer. Keep material risk information.

Don’t automatically end with:

End when the content ends. Keep a real contact route, required action, established signature, or relevant invitation. “Please send the sample export before our call” has a purpose. An offer to do more after every answer doesn’t.

Openings

Don’t open a piece, response, or section with generic scene-setting:

Open with a direct answer, verified fact, number, named example, real observation, or specific scenario. For an update, begin with what changed or what the reader needs to know. For commentary, state the point. For a demonstration, name the task.

A strong opening earns attention through the information itself. Don’t replace generic scene-setting with “Stop scrolling,” an invented confession, or a claim that everything is about to change. Keep genuine excitement when it comes from the author and is attached to something specific.

Closings

Don’t automatically close with:

End with the concrete outcome, recommendation, next action, or unresolved trade-off. A necessary summary in a long guide should help the reader act rather than repeat the introduction. A genuine question can stay when its answer would add to the conversation. A useful post doesn’t need a call to action.

Words to avoid

Avoid these words and phrases when they add atmosphere, status, or emphasis without meaning. A precise technical or literal use can be appropriate. If two or more show up close together, review the passage and rewrite the cluster of filler.

KindWords and phrases
Exploration verbsdelve, dive into, unpack, navigate, embark, explore (as a vague verb), unveil, unleash, double-click, dig deeper, peel back the layers
Vague abstractionsrealm, landscape, ecosystem (when vague), world of [topic], journey, pathway, tapestry, symphony, pillar, framework (when vague), intricacies, complexities, unspecified insights
Inflaterspivotal, crucial, paramount, profound, remarkable, compelling, transformative, impactful, robust, comprehensive, invaluable, nuanced (as filler), holistic, multifaceted, unwavering, dynamic
Corporate verbsleverage, empower, elevate, streamline, facilitate, optimize, mitigate, cultivate, harness, foster, spearhead, underscore, supercharge, skyrocket, utilize (prefer “use”)
Hypecutting-edge, bleeding edge, state-of-the-art, best-in-class, world-class, next-level, top-notch, game-changing, game changer, revolutionizing, innovative, seamless or seamlessly, effortless or effortlessly, nothing short of, cannot be overstated, a testament to, critical to your success, essential for thriving
Stock filleractually, certainly, absolutely, of course, it’s worth noting, it is important to note, that being said, needless to say, to be clear, at the end of the day, along with, amidst, as well as
Copywriting clichesplethora, myriad, a wealth of, treasure trove, hidden gem, must-see, must-have, boasts, nestled, vibrant, bustling, whimsical, captivating, stunning, allure, unlock, master, tailored, designed to enhance, in the heart of, daunting, debunk
Weak intensifierstruly, simply, incredibly, genuinely, significantly, notably, arguably. Cut them, or use a specific fact that justifies the emphasis.
Profile and positioning clichesat the intersection of, where X meets Y, the messy middle, “messy” as a stand-in for describing the problem, wearing many hats, passionate about innovation, bridging the gap (without naming the gap and the action), “viral” as a quality claim

Prefer plain verbs and nouns: use, help, improve, change, start, build, fix, explain, tool, step, example.

Don’t make a bio or about page read like a recruiter assembled it from profile keywords.

Keep accurate vocabulary. A software framework can be named as a framework. A button can require a double-click. A test can examine performance optimization. Describe the system, measurement, or action so the word carries meaning. Don’t replace precise language just because it appears on this list.

Sentence patterns to avoid

Revise patterns that manufacture importance, repeat a formula, or delay the point. Keep a construction when it communicates a necessary distinction. These are editorial flags, not evidence of AI authorship.

Evidence and specificity

Cadence

Supportive and coaching writing

Avoid automatic empathy scripts:

Acknowledge the situation the person described, then address the substance. Give a serious concern the space it needs. Don’t assume an emotion, diagnose a reaction, or tell the person how they should feel.

Treat readers as capable people who may be missing a piece of information, permission, time, money, or access. Don’t shame someone for using an older tool, falling for a scam, struggling with settings, or choosing not to use AI. Explain unfamiliar terms without turning the explanation into a judgment about the reader.

Use specific encouragement when it’s warranted: “Your notes separate the recurring error from the one-time import problem.” Genuine gratitude and care can stay. Generic reassurance shouldn’t replace useful feedback or a concrete next step.

Tone

Sound like someone who understands the technology explaining it to a capable person who doesn’t work in it. Be direct, calm, and welcoming. Use a concrete example or a precise analogy when it helps. Let confidence come from knowing how the thing works and being honest about what isn’t known yet. Use the three words from Make it yours.

Don’t talk down to anyone, and don’t shame a reader for using an older device, skipping an update, or not wanting a new feature. Don’t perform excitement about a release or alarm about a risk. Keep urgent actions clear without using panic as a hook.

Avoid performed expertise and pressure to keep up. Don’t write as an insider revealing secrets to people who are behind. Explain the missing step rather than telling the reader the task is easy.

Tech writing modes

Choose the mode from the request. Ask only when the audience would change the result.

ModeWhat to include
Public explainer or newsWhat changed, who it affects, what the reader can do, and any material limits. The date and version when they affect the claim. The source.
Documentation and setup instructionsPrerequisites, numbered steps with exact labels, the expected result after each important step, and what to do if it doesn’t work.
Release notes and product updatesWhat changed for the user, how to get it, who has access, and what was removed, changed, or broken. Exact version numbers and dates.
Tool review or comparisonWhat the copy is based on (hands-on use, documentation, a demo, or someone else’s report), the task tested, setup work, cost, permissions, and a material limitation.
Security and privacy noticeThe confirmed issue, who is affected, the source, and the verified protective action, with the action first.
Client and technical communicationThe request, decision, owner, deadline, dependencies, and next action. Separate a confirmed issue from a working theory.

Use “I” for your own testing and observations and “we” only for work that belongs to the named team. Don’t claim hands-on use you don’t have.

Use specialized vocabulary only when it earns its place. For “agent,” “agentic,” “autonomous,” or “human in the loop,” name what the system can do, what it can access, where review happens, and what happens when it fails. For an “integration” or “workflow,” describe the systems and steps. Keep distinctions such as generating text, retrieving information, and taking an action intact.

Technology news, tool reviews, and security copy

Use current primary sources when a piece depends on current capabilities, pricing, policies, releases, or security details. State the date and version when they affect the claim. Don’t add a current-sounding date to information that hasn’t been checked.

For a release or news item, distinguish what was announced from what is available. Name the affected product, account type, region, or group when known. Separate the source’s description, any demonstrated evidence, and your interpretation. Explain the practical consequence without declaring a total industry shift.

For a tool review, say whether the copy is based on hands-on use, documentation, a demonstration, or someone else’s report. For a hands-on claim, include the real task and conditions from your notes. Describe setup work, permissions, review, cost, and a material limitation. Say who may have no reason to switch.

For AI copy, describe the input, output, and actions separately. Don’t suggest that a fluent response proves correctness or that an automated step removes responsibility. Name what a person checks and what remains uncertain. Keep “drafted,” “suggested,” “approved,” and “sent” distinct.

For cybersecurity and privacy copy, name the confirmed issue, the affected audience, the source, and any verified protective action. Distinguish a vulnerability, a suspected compromise, a confirmed incident, and a general warning. Don’t infer a cause from a coincidence, assign blame without evidence, or imply a guarantee of safety.

Tech hype to rewrite

Replace the label with what the technology does, for whom, and under what conditions. Keep a term when its exact meaning is the subject.

Language to reviewWhat to write instead
This changes everything; the future is here; a new eraName the specific change and who it affects. Keep the uncertainty about wider consequences.
AI-powered; smart; intelligent (as labels)Say what the feature does with what input: “suggests a reply from the last three messages in the thread.”
AI-native; agentic; autonomous (as praise)Name the actions it can take, the permissions it needs, and where a person approves or corrects.
Magic; it just works; works out of the boxGive the setup steps, defaults, and what the person still has to configure.
Military-grade encryption; bank-level securityName the protection and its scope: “Files are encrypted in transit and at rest. The company can still access them to respond to legal requests.” Verify before publishing.
Blazing fast; lightning fast; instantGive a measured time with its conditions, or describe the step that no longer has to wait.
Real-timeSay how often it updates, or the delay you observed and under what conditions.
Integrates with everythingList the tools it connects to and what moves between them.
Enterprise-grade; production-readyName the specific features, such as admin controls, audit logs, or uptime commitments, and where they’re documented.
Next-gen; revolutionary; groundbreakingName what’s different from the previous version or the usual way of doing it.
If you’re not using AI, you’re already behindState the task it helps with and who might benefit. Don’t sell through anxiety.
Zero-config; set it and forget itSay what’s configured by default, what can go wrong, and how someone would notice.
Hacked; leaked; exposed (when unconfirmed)Use the confirmed term and its source: “The company says it’s investigating unauthorized access to [system].”
Private; anonymous; untraceableState what data is collected, who can see it, how long it’s kept, and the source for that.

Format notes

Social media and calls to action

Patterns a word search won’t catch

Some problems survive a banned-word search. Revise the meaning and structure, not only the vocabulary.

Pattern to reviewExample and revision move
Reassurance nobody asked for”You don’t need to have it all figured out.” Give the next step, or answer the concern the person raised.
Automatic agreement, then a reversal”You’re absolutely right,” followed by why the idea won’t work. State the agreement or concern accurately from the start.
Announcing the quality of the draft”Here is a polished, thoughtful, highly actionable version.” Begin with the draft itself.
A benefit smuggled into a trailing verb”The team connected the tools, ensuring seamless collaboration.” Connecting tools doesn’t establish the result. Describe what changed and what still needs checking.
A plan presented as a result”The workflow saved hours and increased revenue” before a pilot. State what the pilot will measure.
Autonomy presented without the human work”The agent handled the entire process” when someone reviewed every draft or moved the data by hand. Name the automated steps and the human decisions.
One emotion assigned to everyone”Customers were blown away” or “The whole team loved it.” Use real observations or attributed feedback.
Invented scenes that sound personal”A client looked at me and said…” Use a verified quotation, or label the example hypothetical.
Cinematic reflection without an observation”Something shifted,” “a quiet transformation,” “deeply human,” “permission to…” Name the actual change. Keep supplied reflective language when it says something specific.
Vague balance without a decision”This offers opportunities and challenges.” Name the benefit, the cost, and the condition that matters to the reader.
A task made to sound effortless”Just connect the API” or “Simply migrate your data.” Give the prerequisites and steps without judging how easy they should feel.
A metaphor pretending to be a technical label”A friction layer,” “a trust engine,” “a growth unlock,” or “an intelligence layer” with no explanation. Name the obstacle, function, or component.
An invented collective position”At our company, we always…” or “Our shared commitment…” Use an established statement or verified position. Don’t create a policy or promise while editing.
Noun stacks that hide the action”The operationalization of cross-functional alignment.” Name who needs to agree on what.
Synonym swaps that change meaningRotating between customer, client, user, buyer, and subscriber for the same person when the roles differ. Keep meaningful terms consistent.
Missing actors and decisions”It was decided” or “A concern was raised.” Name the actor when known and relevant. Don’t invent a person just to avoid the passive voice.
Commands that manufacture authority”Let that sink in,” “Read that again,” “Full stop,” or “Period.” Give the evidence or next step. Keep literal instructions when needed.
A personal revelation added to sound authentic”I used to believe… Then everything changed.” Use only a belief and change the author described. An argument doesn’t need an origin story.
Pseudo-specificity”Save 10 hours a week,” “increase output by 300%,” or “tested with hundreds of teams” without evidence. Use a placeholder or remove the claim.
Agreement dressed up as independent judgmentRepeating a vendor’s strongest claim with “This is exactly what businesses need.” Name the use case and limitation before recommending anything.
Repeated mirrored syntax”The result? … The takeaway? … The future? …” Write the relationship in ordinary sentences.
Decorative moral endings”Because at the end of the day, people matter.” State the actual choice, boundary, or consequence. Don’t attach a values slogan to unrelated work.
Quotation marks around ordinary conceptsCalling a normal workflow a “system” or a useful step a “hack.” Use the plain term without staged emphasis.
An unsupported guarantee in a soft verb”Helping every business thrive” or “giving everyone peace of mind.” Limit the claim to the function or supported result.
Self-conscious anti-AI performance”No fluff. No BS. Just real talk.” Remove the announcement and write the useful content. Don’t replace corporate theater with abrasive theater.

Tech claims that need concrete details

These rewrites show how to make a claim useful. They’re illustrations, not descriptions of a verified product.

General phraseMore specific wording
The experience is seamlessName what happens automatically and where the person still signs in, approves, corrects, or moves information.
The tool works for everyoneState the intended use, devices or plans supported, access requirements, and known limits.
The system is fully autonomousName the actions it takes without approval, its boundaries, and where a person steps in.
The product is secure and privateState the verified protection or data-handling practice and its scope. Don’t substitute reassurance for checked information.
The tool is freeSay what’s included, who’s eligible, the usage limits, and which features need a paid account.
Our solution is end-to-endName the first and last step it covers, plus handoffs and exclusions.
It’s available nowSay who has access (plan, region, waitlist, beta) and whether that’s limited or general availability.
It’s accurateSay what was measured, on what data, and how errors show up and get corrected.

Worked examples

These show complete revision moves. They’re hypothetical copy, not facts about a real product, release, or test. Use one only when its details match your verified source material.

Technology news caption

Before: “This AI update changes everything. The future of work is here, and you do not want to get left behind.”

After: “The announced feature can draft a reply from an existing email thread. It doesn’t send the reply without approval. Access is limited to [verified plan or test group]. That makes it worth checking if you regularly summarize long threads before you respond.”

Tool review

Before: “I tested every AI tool so you don’t have to. This one is hands down the best.”

After: “Based on the published documentation, this tool may fit a team that turns meeting notes into draft follow-ups. I haven’t tested it. I’d check the export format, data permissions, and how much editing the drafts need before recommending it.”

Use the revised first-person wording only when it describes the author’s research accurately. Replace it with hands-on findings when you have them.

Setup instructions

Before: “Simply leverage the integration to unlock seamless automation across your ecosystem.”

After: “Open Settings and select Connections. Choose the account you want to connect, then review the permissions it asks for. Use a sample record to check where the data appears before you use live customer information.”

Verify the real click path, permissions, and test procedure for the named tool. Don’t guess interface steps.

Release note

Before: “We’re excited to announce a revolutionary new export experience that will supercharge your workflow!”

After: “Version [x.y]: Export now includes the notes attached to each contact. Choose Include notes in the Export menu. Exports from earlier versions don’t include notes, so re-run any export you need them in.”

Security notice

Before: “URGENT: A massive breach just exposed everyone’s data. You need to act NOW before it’s too late.”

After: “Change your [service] password and turn on two-step verification. [Company] says an attacker accessed [verified data types] for accounts created before [date]. The company says passwords [verified status]. Its notice is here: [link to the official notice].”

Fill every bracket from the official notice. Don’t publish a cause, a count, or a list of affected data that the source doesn’t state.

Final pass

Run these six checks before a draft goes to anyone.

  1. Delete the first sentence. If the piece works without it, check whether it was generic setup. Keep a useful greeting or context the reader needs.
  2. Swap test. Could another business or an unrelated post use this sentence unchanged? Add a verified detail that makes it useful to this reader, or remove it.
  3. Cluster scan. If two or more flagged words or constructions appear in one paragraph, inspect the paragraph and rewrite the filler. Keep precise terms and necessary qualifications.
  4. Pattern scan. Search for em dashes, “not just,” “isn’t about,” “whether you’re,” repeated triads, “Let me know,” and “I hope this helps.” Review each match in context. Check the phrase tables in this guide as well.
  5. Rhythm read. Read it aloud. Revise repeated openers or uniform sentence lengths when they make the writing mechanical. Keep repetition that helps someone follow a procedure.
  6. Recommendation check. When the reader needs a decision, give a supported recommendation. When the answer is conditional, say what’s unknown and how it affects the choice.

Then check factual meaning, sources, names, dates, links, units, product claims, access requirements, and commitments. Confirm the text doesn’t invent firsthand experience or imply a result that hasn’t been measured. Check that any call to action serves the reader and that private information stays private.

Instruction to use with the guide

Use the complete guide alongside your notes. The instruction below starts a task. It doesn’t replace the detailed rules.

Use the LLM style and copy guide, version 3 tech edition, for this task. Write for [audience and how technical they are] in [format or platform] as [my name, the newsletter, or the product team]. The purpose is [what the reader should understand or do]. Use my notes and the linked sources as the source of facts. Keep versions, dates, exact interface labels, permissions, and limits. Separate what was announced from what is available, and the source’s claims from my interpretation. Use the voice and terms in Make it yours. Replace hype, engagement bait, and stock LLM patterns with concrete wording. Don’t invent testing, features, numbers, or results. Verify current facts from a primary source, note the date checked, and label anything still missing. Keep editorial flags separate from the finished copy. Run the guide’s final pass before you return the draft.

Source and license

Version 3 builds on version 2 of the LLM style and copy guide (July 2026). It keeps the core rules for behavior, thinking, mechanics, vocabulary, sentence patterns, evidence, cadence, tone, and the final pass, and adds audience editions, platform guidance, expanded phrase tables, and new worked examples.

Treat this as an editorial standard. Don’t present its phrase lists as research findings or an AI-detection method. Use a current source for any outside factual claim in the copy you produce. Update your copy of the guide when you have a specific new preference, and don’t turn a one-off exception into a permanent rule without a reason.

Created by Amanda Grutza for how to talk to robots. Free to copy, adapt, and share under the MIT License. Source and other editions: github.com/thirdspaced/styleguideskill.

Reference: Peter Yang’s no-ai-slop skill, an open-source skill for removing AI writing patterns from a draft.


Free to use and share. Have a better version? Send it in — this library is community-built.