Getting better results

The difference between a mediocre AI result and a great one is rarely the model — it’s the context and the ask. A model can only reason about what it’s shown, and it answers the question you actually wrote. These habits get you consistently better drafts, edits, and answers. If you want the ideas underneath them first, read How AI models work.

Be specific

Say what you want, how long it should be, and any constraint that matters. A vague ask forces the model to guess; a specific one points it straight at the result.

Give the right context, not all of it

The model only sees what’s in the context window, so pointing it at the right material matters more than giving it everything. In Multilo you do this by:

• naming the section you mean, or selecting it before you ask, so the edit lands in the right place;
• @-mentioning the specific files or sources the answer should draw on;
• letting an agent pull from your library rather than pasting whole papers into chat.

Break big asks into steps

A model does its best work on one clear task at a time. Instead of “write my literature review,” sequence it: gather sources, outline the themes, draft one section, then revise. Each step has focused context and a result you can check before moving on, so errors don’t compound. For whole-document jobs, the Full Draft Writer does this staging for you — plan, then draft section by section.

Start small and refine

Treat the first reply as a starting point, not a verdict. Ask for one change at a time — “make the second paragraph more concrete,” “add a counter-argument here” — and build on what’s already there. Because results aren’t identical every run, re-running a step is also a legitimate way to get a better version.

Start a fresh chat for a fresh task

When you move to an unrelated task, start a new chat. A long conversation carries all its earlier turns in the context window, which both crowds out room for the new work and can pull the model toward the old topic. A fresh session is a clean slate — and a genuinely independent second look when you want the AI to review something it just helped write.

Set your rules once, reuse them

Anything you find yourself repeating in every prompt belongs in a rule that applies automatically:

• voice.mdteaches Multilo how you write, so you stop asking for “my voice” every time;
• integrity.md enforces your research-integrity rules — most importantly, never fabricating a citation;
• AGENTS.md holds per-project instructions every agent reads.

Treat these as living documents: start minimal, and add a rule only when you notice the AI repeating a mistake. A short, current set of rules works better than a long, stale one.

Match the model and MODE to the task

A quick cleanup doesn’t need your most capable model, and a hard synthesis shouldn’t run on the fastest one. Pick the model tierthat fits the difficulty, and raise an agent’s MODE when you want a deeper pass. Matching both to the stakes gets better results and spends fewer credits.

Verify — don’t trust

Models can state something fluently and still be wrong, and they have a knowledge cutoff. For research that’s not optional. Run Claim Check to confirm each citation is actually supported by its source, and use your libraryas the ground truth rather than the model’s memory. The goal isn’t to out-prompt the model — it’s to keep a human check on what gets into your work.