Best practices

Patterns that get the most out of Multilo: give agents a way to check their work, plan before you write, set up your project, manage your chats, and avoid the common mistakes.

Multilo is an agentic writing environment. Unlike a chat window that answers and waits, Multilo’s agents can read your project, edit your document, search your sources, and work through a task step by step while you watch, steer, or get on with something else.

That changes how you work. Instead of writing every sentence and asking the AI to check it, you describe what you need and the agent works out how to get there: it reads, plans, writes, and checks. The patterns below are the ones that consistently get the best results. For how the agent loop works, see How Multilo works. For prompt-level habits, see Getting better results.

Most of these practices come from one limit: the context window fills up, and results get worse as it does.

The context window holds everything the model can see at once — your conversation, the document, the sources the agent read, and the results of every tool it ran. A long research session fills it quickly. Multilo compacts older material automatically, but a crowded window still means early instructions lose detail and mistakes become more likely. Treat context as the most important thing to manage.

Give the agent a way to check its work

Give the agent something to check against — your sources, a word limit, a list of requirements. It’s the difference between a draft you have to proofread line by line and one that arrives already checked.

An agent stops when the work looks done. Without something to check against, “looks done” is the only signal it has, and you become the checker: every mistake waits for you to notice it. Give it a clear test — the claim must be in the source, the abstract must cover four points, the section must stay under 800 words — and it can check its own output before handing it back.

Strategy	Before	After
Ground claims in your sources	“Write about survey bias.”	“Draft a paragraph on response bias in online surveys, citing only sources in my library. If a claim can’t be supported, say so.”
State what “done” means	“Make this abstract better.”	“Rewrite this abstract in under 250 words, covering background, method, results, and conclusion. Then check it against those four points and list anything missing.”
Check against the source, not memory	“Is this citation right?”	“Quote the passage from Smith (2021) that supports the claim in paragraph 3, or tell me it doesn’t.”

Once you have a check, decide how firmly it applies:

In one request: ask the agent to check its work in the same message, as in the table above.
For every request: turn on strict integrity mode so the AI never invents a citation, and write standing rules in AGENTS.md.
As a final gate: run Claim Check, which fetches every cited source and tells you whether it supports the claim.
With a second opinion: review the draft in a fresh chat, so the reviewer isn’t the one who wrote it. See Use a writer and a reviewer.

Ask for evidence rather than reassurance: the quoted passage, the figure from your data file, the list of points covered. Checking evidence is much faster than redoing the work yourself.

Explore first, then plan, then write

Separate reading and planning from writing, so you don’t end up with a polished section that argues the wrong thing.

Letting the agent start writing straight away can produce a fluent draft built on the wrong structure. For anything bigger than a paragraph, work in four phases:

Explore
Ask the agent to read and report, without changing anything: “Read the papers in my library on remote work and summarize where they agree and disagree. Don’t write anything yet.”
Plan
Ask for an outline you can edit: “Propose an outline for the Discussion that answers my research question, with the sources for each point. Don’t write it yet.” Change the outline until it says what you want.
Write
Ask the agent to write from the plan: “Draft the Discussion from that outline, about 1,200 words, citing as you go.”
Check
Ask for the gaps, then verify: “List every claim in the Discussion that doesn’t have a citation.” Finish with Claim Check.

The Full Draft Writer follows the same pattern on its own: it researches, shows you the outline first, then drafts section by section.

Planning is useful, but it adds steps

For small, clear edits — fix a sentence, change a citation style, shorten a paragraph — just ask. Plan when you’re unsure of the structure, when the change touches several sections, or when you are new to the material. If you could describe the edit in one sentence, skip the plan.

Provide specific context

The more precise the request, the fewer corrections you’ll need.

The agent can infer a lot, but it can’t read your mind. Name the section, point at the sources, and say what a good result looks like.

Strategy	Before	After
Scope the task. Say which part, and what to leave alone.	“Improve chapter 2.”	“In section 2.3, tighten the argument about sampling. Don’t touch the other sections.”
Point to sources. Name the material that holds the answer.	“How is engagement defined?”	“Using @smith-2021.pdf and @lee-2023.pdf, explain how each one defines engagement and where they differ.”
Reference your own work. Point at a model to follow.	“Write the Methods section.”	“Write the Methods in the same structure and tone as the Methods in @pilot-study.docx.”
Describe the problem. Say what’s wrong and what “fixed” looks like.	“This section is bad.”	“My supervisor says section 3 jumps between ideas. Reorder the paragraphs so each builds on the last, without changing what they say.”

An open question still has its place when you are exploring and can afford to change direction. “What would you improve in this chapter?” can surface problems you hadn’t thought to ask about.

Provide rich content

You can give the agent more than text in a message:

@-mention files and folders instead of describing where something is. See files & folders.
Attach images, audio, and video. Agents can read a chart, a scanned page, or a recorded interview. See media & images.
Keep data in the project. Drop results, tables, and figures into the project folder and agents can read them.
Import your references. Bring BibTeX, RIS, or CSL-JSON files into the library instead of pasting citations into chat.
Let the agent find what it needs. Describe the material — “the interview transcripts from March” — and the agent searches the project for it.

Set up your project

A few minutes of setup make every chat and every agent run in the project better.

Write an effective AGENTS.md

Put the instructions you’d otherwise repeat in every chat into .multilo/AGENTS.md, and keep it short.

AGENTS.md lives in your project’s .multilo/ folder and is loaded into chat and every agent. Use it for the things the AI can’t work out from your documents: your conventions, your terminology, and your rules. There is no required format — short and readable works best:

.multilo/AGENTS.md

# Style
- British spelling; APA 7 for citations.
- Call participants "respondents", never "subjects".
- Keep paragraphs under 150 words.

# Project
- Research question: how does remote work affect team trust?
- Chapter 4 is final. Don't edit it unless I ask.

For each line, ask: would removing this cause the AI to make a mistake? If not, cut it. A long file buries the rules that matter.

✅ Include	❌ Leave out
Spelling, style, and citation conventions	Things the AI can see in your documents
Terms your field or supervisor insists on	General writing advice (“write clearly”)
Your research question and scope	Summaries of every source
Sections or files that must not change	Details that change every week
Requirements from your journal or department	Long explanations and background reading

If the AI keeps ignoring a rule, the file is probably too long. If it keeps asking about something the file already answers, the wording may be unclear. Because .multilo/ is plain files, you can version it with Git and share it with co-authors.

Set your voice and integrity rules

Two more files shape everything the AI writes. voice.md describes your writing voice — Multilo can learn it from samples of your own work — so drafts sound like you. integrity.md holds your research-integrity rules, and strict mode makes the AI refuse to invent a citation. Set both from Settings, or when you create the project. See author voice & integrity.

Build your library first

Agents write from your library, cite from it, and check citations against it. With an empty library, they fall back on general knowledge. Add the papers you have, then use Sources Hub to find the references inside them and fill the gaps — before you start a big writing run.

Match the model and MODE to the task

A fast model is fine for tidy-ups and summaries; a flagship model is worth it for critique, synthesis, and long drafts. An agent’s MODE sets how deep it goes on a run. Turn on Auto to let Multilo pick the model for each message. See choosing a model.

Add skills and extensions

Skills teach the AI reusable expertise — a style guide, a review checklist, a field’s conventions — and load only when a task needs them, so they don’t crowd every request. Extensions add new agents, viewers, and integrations from the marketplace, and show the permissions they need before they run.

Communicate effectively

Ask the agent the questions you would ask a well-read colleague, and for big documents, let it ask you questions first.

Ask the questions you’d ask a colleague

When you are new to a topic or a set of papers, use the agent to learn. No special wording is needed — just ask:

What is the main argument of @smith-2021.pdf?
Which of my sources use a longitudinal design?
What does the equation in section 3 actually mean?
Why might these two studies reach different conclusions?
What would a reviewer in my field expect to see in the Methods?

Let the agent interview you

For a big piece of writing — a proposal, a thesis chapter, a review — start by having the agent ask you questions. It raises things you may not have thought about yet: scope, audience, structure, and the arguments against you.

I want to write [a grant proposal on remote-work and team trust]. Interview me in detail, one question at a time. Ask about my aims, methods, audience, constraints, and the strongest objections. Skip the obvious questions. When we've covered everything, write a one-page brief I can draft from.

Put the finished brief in a document in your project, then start a fresh chat to draft from it. The new chat begins with a clean context focused on writing, and you have the brief to check the draft against.

Manage your chats

Chats are cheap to start, and your work is safe in the project. Use both to your advantage.

Course-correct early and often

Correct the agent as soon as you see it going the wrong way.

The best results come from short feedback loops. Watch the agent as it works, and:

Stop it at any point and send a correction — “Stop. Keep the original structure and only fix the transitions.”
Answer its questions. When an agent pauses to ask you to choose, a clear answer saves a wrong guess.
Restore an earlier version from history if an edit went wrong.
Start a new chat when you move to an unrelated task.

If you have corrected the same thing more than twice, the chat is full of failed attempts. Start a fresh chat with a better first message that includes what you learned. A clean chat with a clear request almost always beats a long one full of corrections.

Keep context focused

One task per chat. A fresh chat starts with a clean context window.
Rules go in files, not messages. Instructions from early in a long chat can be condensed along the way; rules in AGENTS.md, voice.md, and integrity.md apply every time.
Point, don’t paste. @-mention the one source that matters instead of pasting whole papers into chat.
Cap long runs. Set a per-task token budget in Settings → AI; the agent pauses and asks before spending more.

Experiment freely with history

Inside a project, every write keeps the text it replaced: the last 40 versions of each document, for 14 days. So you can ask for something bold — a full restructure, a different argument — and if it doesn’t work, put the old version back. Restoring is itself saved, so a restore can be undone too. See versions & safety.

History is not a backup

History only covers files inside a Multilo project, and only the recent past. Commit the milestones you’d hate to lose to Git.

Pick up where you left off

Your work lives in the project, not in a chat. Open the project and your document, library, and rules are all there. To carry a fact forward, ask an agent to remember it: “Remember that chapter 4 is final.”

Scale up

Once one chat works well for you, get more done by running work side by side and turning repeated tasks into tools.

Write while an agent works

Several sessions can work on the same document at once without overwriting each other. Draft the introduction yourself while an agent writes a later section, then review its work when you reach it. See versions & safety.

Use a writer and a reviewer

A fresh chat makes a better reviewer, because it isn’t attached to text it just wrote. Use two chats:

Chat A (writer)	Chat B (reviewer)
“Draft the Related Work section from my library, about 800 words.”
	“Review the Related Work section for missing perspectives, unsupported claims, and anything that misrepresents a source. List problems, not fixes.”
“Here is the review: [paste]. Address each point.”

Turn repeated work into agents

If you run the same kind of check on every chapter, save it as your own agent and run it in one click:

/createagent a "Chapter checker" that checks a thesis chapter for a clear aim, a link back to the research question, and a citation for every factual claim

It appears in the Agents Library with a YOURS chip. See custom agents.

Add a final review step

The longer an agent works on its own, the more an independent check matters before you call the work done. Run Claim Check for the citations, then review the draft against your brief in a fresh chat:

Check this draft against my brief in [@brief.docx]. Confirm every aim is addressed, every factual claim is cited, and nothing outside the brief's scope was added. Report gaps, not style preferences.

A reviewer asked to find problems will usually find some, even in good work. Fixing every one leads to over-cautious, padded writing. Act on the gaps that affect your argument or your requirements, and treat the rest as optional.

Avoid common failure patterns

These mistakes are common. Spotting them early saves time:

Pattern	What happens	Fix
The everything chat	One chat drifts from the introduction to citations to an unrelated email. The context fills with things that no longer matter.	Start a new chat for each unrelated task.
Correcting over and over	Each correction adds another failed attempt to the context.	After two failed corrections, start fresh with a better first message.
The overstuffed AGENTS.md	The important rules get lost among the obvious ones.	Cut every line the AI already follows without being told.
The plausible, unchecked draft	The text reads well, but a citation doesn’t say what the sentence claims.	Run Claim Check and ask for quoted evidence. Don’t submit what you haven’t checked.
The empty library	With nothing to cite from, the agent falls back on general knowledge.	Add sources first, and use Sources Hub to fill the gaps.
The whole-thesis request	“Write my thesis” in one message gives generic text with no structure you chose.	Explore, plan, and write one section at a time — or let the Full Draft Writer plan first.

Develop your intuition

These patterns are starting points, not rules. Sometimes a long chat is exactly right, because you are deep in one hard argument and the history matters. Sometimes you should skip the plan, because you want to see what the agent makes of a loose idea. Sometimes a vague prompt is the point.

Notice what works. When you get a great result, look at what you did: the request, the context, the sources, the model. When a result is poor, ask why. Was the context too crowded? The request too vague? The task too big for one go? Over time you’ll know when to be specific and when to leave room, when to plan and when to explore, and when to start fresh.

How Multilo worksThe agent loop, tools, and context.Common workflowsStep-by-step recipes for everyday tasks.Prompt libraryCopy-paste prompts by task and role.Getting better resultsPrompt habits for better drafts and edits.