Quick answer
To use an AI agent for research and writing, give it a project folder with your sources, a short written brief (goal, audience, sources, output and rules), and ask it to plan before it writes. Ask for a finished artifact, such as an outline, comparison table or draft, with a source next to every fact. Then check the facts, edit it in your own voice and save the brief as a reusable template. Agent tools that were built for coding, such as Google Antigravity, Claude Code or Codex, can do this well.
Key takeaways
- Agents can read files, follow multi-step plans and save finished documents, which suits research and writing.
- A five-part brief (goal, audience, sources, output, rules) does most of the work.
- Asking for a plan first catches misunderstandings before they cost time and tokens.
- Insist on sources for every fact and a clear marker for anything unverified.
- Keep agents read-only and offline-safe unless they truly need to send or publish.
Google has just given its Antigravity agent a new interface for research, writing and creating documents, without the terminal and Git clutter. We covered it in our news round-up, Gemini Krypton Leak: Claude Fable 5.2, a Hidden Fable 6 and This Week’s AI News. This guide shows how to use the same approach in whichever agent tool you already have.
📘 Main article: Read the full news round-up for context: Gemini Krypton Leak: Claude Fable 5.2, a Hidden Fable 6 and This Week’s AI News. It covers Antigravity, Claude Fable 5.2, Gemini Krypton and more.
In this guide
- Can an AI agent help with research and writing?
- How is an agent different from a chatbot for this work?
- Step 1: Set up a project folder
- Step 2: Write a clear brief
- Step 3: Ask for a plan before any work
- Step 4: Ask for an artifact, with sources
- Step 5: Check the facts and limit permissions
- Step 6: Edit in your own voice and save the recipe
- Which tasks work best?
- Frequently asked questions
- Related posts
Can an AI agent help with research and writing?
Yes. Research and writing are mostly about gathering material, organising it and turning it into a clear document. Agents are good at exactly that kind of multi-step work, as long as you give them good sources and clear rules.
Google’s move to add a knowledge work interface to Antigravity is a sign that agent makers see writers, researchers and other professionals as users, not just developers.
How is an agent different from a chatbot for this work?
| Chatbot | Agent | |
|---|---|---|
| Works with | What you paste in | Files, folders and tools you give it |
| Steps | One reply at a time | Plans and carries out several steps |
| Output | Text in a chat window | Saved documents, tables and drafts |
| Best for | Quick questions and edits | Longer research and writing projects |
If a task takes you more than an hour of gathering and organising, it is a good candidate for an agent.
Step 1: Set up a project folder
Create one folder per project and put everything the agent should use in it.
- A
sourcesfolder with PDFs, notes, transcripts or exported web pages. - A
sources.mdfile listing any links it may read. - An examples file with a past article or report you like, so it can match the structure.
- A
brief.mdfile for your instructions (see Step 2).
💡 Tip: Remove anything private or sensitive from the folder first. The agent can only use what you give it, which is also what keeps it focused.
Step 2: Write a clear brief
A good brief has five parts: goal, audience, sources, output and rules. Here is an example you can adapt:
Example brief:
Goal: Write a 1,000-word blog post comparing three email platforms for small online shops.
Audience: UK shop owners with fewer than 1,000 subscribers, not technical.
Sources: use only the files in /sources and the official pricing pages linked in sources.md.
Output: save an outline to outline.md first, then the draft to draft.md, with a source link after every fact or price.
Rules: British English, short paragraphs, no income claims. Do not send, publish or sign up for anything. If a fact is missing, write [CHECK] instead of guessing.
The [CHECK] rule is the most useful line. It gives the agent permission to admit gaps instead of inventing an answer.
Step 3: Ask for a plan before any work
Before the agent writes anything, ask it to read the brief and sources and reply with a short plan. For example: “Read brief.md and the sources. Reply with your plan, the sections you will write and any questions. Do not start writing yet.”
Check that it has understood the audience and is using the right sources. Correcting a plan takes a minute. Correcting a finished draft takes much longer and uses more tokens.
Step 4: Ask for an artifact, with sources
Ask for a finished document, not just a chat reply. Good artifacts include:
- An outline with headings and the key point for each section.
- A comparison table with a source link in each row.
- A first draft saved as a file.
- A one-page summary of a long report.
Ask for a source link or file name after every fact, price or figure, so you can check them quickly.
Step 5: Check the facts and limit permissions
Never publish an agent’s research without checking it. Open every source link, check figures and dates, and look for anything marked [CHECK]. Our guide on How to Fact-Check an AI Research Report Before You Rely on It has a quick checklist.
⚠️ Warning: Research does not need permission to send emails, post or buy anything. Keep the agent limited to reading and saving files in the project folder. See How to Stop an AI Agent Sending Your Data Online Without Permission.
Step 6: Edit in your own voice and save the recipe
Treat the agent’s draft as raw material. Read it aloud, cut repetition, add your own examples and opinions, and make sure it sounds like you.
When a project works well, save the brief as a template, for example brief-blog-comparison.md. Next time, you only change the goal and sources.
Your next step: pick one research task you have been putting off, set up a folder and try the five-part brief this week. Then catch up on the wider news in Gemini Krypton Leak: Claude Fable 5.2, a Hidden Fable 6 and This Week’s AI News.
Which tasks work best?
| Task | Why an agent helps |
|---|---|
| Competitor or supplier comparison | Reads several pages and builds one table |
| Blog post from your notes | Organises scattered notes into a structure |
| Summary of a long report | Pulls out key points with page references |
| Customer FAQ from support emails | Groups repeated questions into clear answers |
| Course or webinar outline | Turns a topic list into a logical sequence |
Frequently asked questions
Which AI agent is best for research and writing?
Any agent that can read files and save documents can work, including Google Antigravity, Claude Code and Codex. Pick the one you already use and test it on a real task.
Do I need to know how to code?
No. You write the brief in plain English. The agent handles the files and steps.
How do I stop an agent making things up?
Give it specific sources, ask for a source after every fact, and tell it to write [CHECK] instead of guessing. Then check the facts yourself.
Is it safe to give an agent my documents?
Only give it what the task needs, remove sensitive data, and keep it limited to reading and saving files in one folder.
How long should the brief be?
Five to ten lines is usually enough: goal, audience, sources, output and rules.
Related posts
- How to Fact-Check an AI Research Report Before You Rely on It
- How to Stop an AI Agent Sending Your Data Online Without Permission
- Low, High or Max: How to Choose an AI Reasoning Effort Level
