Thanks again for joining me at ACES.
Want to build an agent that helps you look up and apply your publication’s style rules? Start with a small set of rules, clear instructions, and examples you can check.
This guide focuses on things to work on as you set up your agent; it doesn’t go much into how to choose an agent (or create one locally if you have tech resources to support that). Once you decide which kind of agent you’d like to create, the guidelines on this page will help you make it a success. You can read up on a few of the available models in my Gemini Notebook. If you haven’t used this tool, formerly known as NotebookLM, check it out. You can ask it questions about the source material I’ve added, which includes documentation from several AI models.
OK, grab your publication’s style guide. Time to get started.
Choose your default style guide.
Name the guide your publication follows by default, such as AP Stylebook, and identify the edition or online version you use. Then state how your house rules relate to it: “Follow our house style first; use AP style where we have not specified an exception.”
Naming a guide does not give your agent access to it. If the guide sits behind a paywall, respect that boundary. Your subscription does not automatically authorize uploading its contents to an AI tool. Check the publisher’s terms and any applicable license before copying or connecting material. AP Stylebook’s terms restrict copying and retaining its content.
You’ll need to be very specific with your agent. Tell it that it should inform users of the default style guide the first time they ask it a question, and that if a rule in your style guide conflicts with a rule in the default, that your style guide breaks the tie.
Make important rules explicit—even when they match your default.
Your agent’s working style sheet should include your publication’s important everyday rules, not only departures from your default guide. If a rule matters to your workflow, document your publication’s decision in your own wording, with your own examples, even when that decision agrees with your default. Keep this focused on your house requirements; do not reproduce or reconstruct the paid guide.
For example, identify the specific decisions you need checked about abbreviations, numerals, job titles, and punctuation. For each, record the rule, its exceptions, and a correct and incorrect example. Verify any claimed agreement with AP against your authorized reference, and note when you checked it.
This gives the agent explicit rules to apply and gives you something concrete to test. It does not guarantee enforcement. If a question requires a rule the agent cannot access, instruct it to flag the question for an editor to check in the default guide rather than guess or claim it consulted AP.
Gather your style sources.
Choose the style guide, house style sheet, and approved terminology your agent should consult. Use material you have permission to upload to your chosen tool. Give each document a clear name and version date, and remove any outdated or superseded rules from the working collection.
Specify which source wins when rules conflict. For example, you might put your publication’s house style ahead of its general style manual. Write down that order rather than leaving the agent to infer it. Your agent should apply the rules you provide without inventing additional rules or silently filling gaps.
Don’t just instruct. Enforce.
If your agent has access to something, it may retrieve it. In addition to telling it to work from your source data only, limit its access to only that source data. Anything else it can access could cause it to give wrong answers.
Don’t rely on instructions alone to limit sources. Restrict connected files, websites, and tools to those the agent needs. Where available, disable web search and general-knowledge answers, then test whether it stays within those boundaries.
Also, remember that in addition to your source data, the agent has its training knowledge, which includes style information that may conflict with what you want. Be very explicit that it should not draw from its training knowledge to answer questions, only from the sources you provide.
Start with an icebreaker.
Decide how your agent should greet users and include this in the instructions. Most agents also have the capability to surface some sample questions when launched. This is especially good for new users who otherwise may not know how to interact with the agent. (Remember the first time you used an AI tool and saw that blank chat box?) Seeing sample questions there will clue them in to how it all works.
You can also help your users by stating any limitations at the beginning of their session. For example, if your agent does not retain context from chat to chat, it should inform users about this limitation and let them know it will not remember what they’ve talked about previously.
It should also remind them to check its work. Speaking of which…
Require your agent to show its work.
Include a requirement that your agent cite the rule from the document it used to arrive at its answer, along with a clickable link to that document so that users can double-check it. Doing this helps ensure the agent is surfacing appropriate documents only, and it makes it quick and easy to do what you should always do and check an agent’s output.
If the agent doesn’t know, make it say so.
The correct answer should always be the correct answer, or something like, “I couldn’t find a rule addressing this in the available reference material. Please check with [person or team].” Instruct the agent not to pull material from its own knowledge, make inferences based on other things that are in your data sources, or guess based on common usage. If the user asks a question your data doesn’t cover, it should not attempt to answer that question. Period.
Write the rules as simply as possible.
Limit yourself to one rule per sentence. Don’t cram in anything extraneous. Avoid negative constructions in favor of plain, declarative statements. Cleverness is wasted in this scenario, so wring all the clever out of your writing for this project. Be clear, straightforward, and humorless. (This part was difficult for me.)
Limit chatter.
Users will appreciate if it just gets to the point. Whether it should suggest a next action is up to you. Much of the time it will not accurately guess what the user wants to inquire about next, because typically a user with style guide questions has several unrelated issues to ask about. If your users aren’t very experienced with AI, however, they might not realize they can jump in and ask something else. So, your call on that one.
Write instructions you can test.
Adapt this starting point to your own sources and workflow:
Help me answer style questions using the supplied style documents. Follow this source order: [list your sources in priority order]. For each recommendation, identify the supporting rule and its document, heading, or page. Do not invent rules or citations. If you cannot find a rule, say so. If the sources conflict and the priority order does not resolve the conflict, explain it and ask me. When reviewing a passage, flag suggested changes with reasons and leave the final decision to me. Preserve the author’s meaning and voice.
These instructions describe the behavior you want. Check whether the agent actually follows them before relying on its answers.
Give the agent one job first.
Do a test run. Start with answering style questions and identifying the supporting rule. Once that works consistently in your tests, try asking it to flag possible style violations in a short passage for your review.
Decide what each of its answers should contain: the recommendation, the relevant rule, its source and location, and any exception or unresolved question.
Test rules, exceptions, and gaps.
Create a small answer key from your actual style guide. Include:
- A straightforward rule: an abbreviation, capitalization choice, or preferred spelling with a clear answer.
- A house-style override: a case where your publication departs from its general style manual.
- An exception: a rule that changes with context, such as a term used in a quotation or a proper name.
- A missing rule: a question your supplied documents do not answer. The agent should acknowledge the gap.
- An already-correct passage: text it should leave alone.
Compare its recommendations and citations with your answer key. Record failures, revise the sources or instructions, and repeat the tests after changes to the guide or agent.
Explore my AI notebook.
For further exploration, browse my collection of AI material. It covers AI more broadly; the steps above focus on a style guide agent.
About the author
Allison M. Payne
Jacksonville, Florida
B2B editor, writer, and content strategist with 26+ years in publishing. A graduate of the creative writing program at Eckerd College, I turn clients' expertise into useful content people trust, search engines understand, and AI can cite.
