A research agent can bring back more work than evidence

A research agent can return an impressive number of links before the original question has become particularly clear. That's a strange kind of productivity: the task began as “find something worth posting” and now somebody has to read a small report about the entire industry.

The useful part of a tool such as Manus is the possibility of delegating bounded research. Its documentation describes a feature called Wide Research. That explains an available approach, not a guarantee that every returned source will support the claim you wanted to make. The source still needs opening. Annoying, yeah. Still the job.

Suppose the task is a November carousel about a question customers keep asking. “Research November trends” is a broad invitation to collect whatever looks current. “Find official explanations and specific examples relevant to this supplied question, for this audience and region” gives the research a much more defensible purpose.

Even then, recent publication and relevant evidence are different things. A page updated last week might repeat an old example. A community comment might be useful language for understanding a frustration without establishing how common that frustration is. An attractive screenshot from a report doesn't explain the population behind its percentage.

Give the search somewhere to stop

A short working brief can name the question, the audience, the period, and what counts as a usable source. It can also say what is outside the task. Without that last bit, every adjacent subject becomes another apparently necessary section.

For an illustrative research task, the instruction could be:

Research only the supplied question for the stated audience.
Return a small set of relevant sources with dates and links.
For each proposed fact, show the passage supporting it.
Separate official guidance, reported observations, and anecdotes.
If evidence is missing, write UNKNOWN. Do not draft the post yet.

That final sentence is useful. Research and writing can happen in the same interface, but they should not quietly become the same decision. Otherwise a weak source becomes a strong sentence simply because the model has moved on to making the article sound finished.

A brief has to survive opening the links

Read the underlying source before adopting its conclusion. Does the linked page actually contain the claimed number? Is it about the same feature and region? Is the example still available? If the answer is no, remove the claim or narrow it to what the evidence supports.

A report from the company selling a tool can contain useful information. It should still be described as that company's report. The same applies to an enthusiastic post from a community member: potentially interesting, not magically representative of everybody using the product.

The finished research brief might be much shorter than the agent's first output. That's fine. A few grounded observations and one useful example give a writer more to work with than a catalogue of unrelated trends. Keeping rejected claims in working notes also stops them from wandering back into the next draft with a nicer adjective.

Once the material is sound, ViralityWand can help prepare a carousel from it. The formatting stage cannot rescue a source that said something else. A beautiful slide with a citation attached is still wrong when nobody checked where the citation goes.

Further reading: Wide Research · Instagram trends this week · TikTok trends this week · ViralityWand editor · caption generator · hashtag generator.