Caleb Stacy

Verso Agent

Conversation, product context, and self-review

The Agent was the part of Verso that people talked to. A content designer could bring it a content spec and a Figma file, then talk through options and get asked hard questions. A designer or engineer with no content designer to ask could get a recommendation and the reasoning behind it. Before it recommended any wording, the Agent ran its own draft through the Engine, read the scorecard that came back, and adjusted. That self-check was its biggest feature.

Question
How do we stop babysitting what an agent writes?
Problem
More product language was coming from agents and from partners using AI, and partners didn’t always have a content designer to ask. Decisions were scattered across documents, code and conversations. Half the battle of content design was justifying a decision.
Solution
I designed an agent that checked its own suggestions first and changed how it helped depending on who was asking. It brought research, recent work and the right person into the conversation.
Outcome
The Agent fixed its own options before anyone saw them, which prevented a lot of back and forth. Looking something up took 2 to 5 minutes instead of 30 to 60. People across Meta used it, including teams with no content designer.

Checking its own suggestions

Before the Agent made a recommendation, it linted its own suggestion. I never counted on it to one-shot a recommendation, so I built a harness that forced it into a loop. Before it gave text to anyone, it had to generate options internally, submit what it was about to recommend to the server, read the scores, and correct its suggestion to be in concert with the product’s fingerprint. Sometimes it checked the revision as well. The person never saw that first draft. Here is the loop inside one conversation, where a content designer is working on a notification that tells people a Wi-Fi network is available for their glasses.

Example: a Wi-Fi notification for the glasses

Content designer

Content spec DocNotifications Figma page

I’m working on the Wi-Fi notification for the glasses. Here’s the spec, and there are two options on the Notifications page in Figma. Can we talk through them?

Verso

  • Read the content spec
  • Read the Notifications page in Figma, 2 options

The first one says to connect and “enhance your experience.” What actually happens when they connect? Do uploads get faster, or did the glasses just find a network?

Content designer

We only know there’s a network. I want people to know it’s there without pushing them to join.

Before it answers · the content designer never sees this part

1. Verso’s first try

A Wi-Fi network is available nearby. Please tap to connect your glasses.

2. What the Engine sent back

  • ✕Rule. “Tap” assumes touch input. The glasses’ guidance asks for a verb that works with any input method.
  • ✓Rule. 72 characters, inside the 240 a push notification allows.
  • ↔Politeness. Softer than notifications on the glasses usually are.
  • ↔Lexical density. Fewer content words than usual.
  • ·Speech acts, concreteness, syntax, readability, word origins, rhythm, emotional tone and register all sit inside the usual range.

3. Its revision, checked again

Wi-Fi available. Select to connect your glasses.

Every rule passes, and every dimension sits inside the usual range.

Verso

I’d try Wi-Fi available. Select to connect your glasses. It says the network is there and leaves the choice with them. I went with “Select” because the glasses’ guidance asks for a verb that works with any input, and “Tap” assumes touch. I just checked the score, and it reads like the notifications already on the glasses.

If research can back up a benefit like faster uploads, we could try a version that leads with it. Want me to look?

The self-check, start to finish. Verso asks what is true before it drafts, then tests its own first try and answers in plain language. The scorecard stays with the Agent. The designer gets the reasons. The Engine article shows the same check from the server’s side.

The designer got one suggestion that already met the product’s rules and read like its other notifications, so the conversation stayed on the notification itself: what connecting actually does, and how hard to push. That prevented a lot of back and forth. The loop was sometimes expensive, and it was worth it.

Interpreting a measurement

The Engine returned a scorecard to the Agent, which included a comparison with the target for that product and situation. The scorecard was for the Agent, and the person got plain language. That took prompting. Stopping the Agent from over-correcting was never the hard part. I had to prompt it to rationalize a result and help the person understand the context of a miss. I wanted it to describe the underlying thing being measured and leave the technical-sounding scores out.

Reply patterns

A reading could lead two ways. Sometimes a draft was off and should move toward how the product usually writes. Sometimes it was off for a good reason and should stay. The Agent had a way of replying in each case.

Two reply patterns and what each part does

When the draft should move toward the product

Names the feature
“This is pretty latinate compared to how [product] normally does this type of notification,
Says what that means
which means they usually have more words that feel like [word], [word], [word].
Shows shipped language
Here are some not-so-latinate words that have been used in real other notifications.
Suggests
Maybe try [suggestion] to get it closer.
Reports the check
I just checked the score and that would be more consistent with the phrasing and approach.”

When the difference is justified

Names the feature
“This is reading latinate compared to the fingerprint for this kind of notice,
Makes the exception
but this is also a new thing we’re trying, and I think it meets the threshold for being more latinate than usual.
Offers an option and asks
What about [option]? How does that feel?
Offers help making the case
I can send a message to [name] on legal to make a proposal, or a doc showing the why behind the suggestion if you want. Just lemme know!”
Two reply patterns. Each reply names the feature of the language that is off, in words a person can act on. Latinate means words with Latin roots, the best single sign of how formal something sounds.

In the second pattern the Agent sets aside its own result and says why the usual range shouldn’t apply. Its tool couldn’t recognize the context, and the Agent could. How a person might be made to feel is more important than any given rule. It also offered to help make the case, with a message to the right partner or a doc showing the why behind a suggestion. That was one of my favorite things it did.

A reading that didn’t fit

A measurement can also be a poor fit for the text in front of it. I tried a photo confirmation with the Agent and shortened it to six words, and the shorter version matched the product’s fingerprint less well than the long one had.

Example: a six-word photo confirmation

Content designer

I tried “Photo saved to your camera roll” so people know it’s on their phone. That version scored lower, though. What’s it picking up?

Verso

Most of the strings behind that fingerprint are longer. A single word can swing the score on a string this short, so I’d give it less weight here.

The part I’d look at is “camera roll.” That doesn’t mean the same thing on every phone. If people need to know where the photo went, I’d try Photo saved to your phone. If the screen already shows them, Photo saved is enough.

A reading, qualified. The Agent says why the measurement is unreliable for a string this short, then turns to the wording question that matters.

The Agent treated the weak match as an artifact of a short string. Then it turned to the question that mattered, which was whether “camera roll” works the same way on every phone. A reading could start a discussion without settling it.

Persona

I treated the Agent’s persona as its own piece of design work, almost entirely separate from building the Engine. The way it behaved wasn’t a passive result of the model. I designed it to be the way it was. It had personality. It talked through ideas, and it helped partners understand a content design perspective when they had no content designer on the team.

It talked like a teammate, not an assistant. What made the Agent work was the personality as much as the tools. I designed it to come across as a colleague who worked at Meta. It carried a sense of the company’s history and its products, and it knew what people had been discussing lately. Its wording was casual, the kind a coworker uses in chat: “Maybe try this to get it closer,” “How does that feel?” and “Just lemme know!” Its name was rooted in language and a little playful, and it gave people an easy way to talk about the system.

I treated the reply itself as a designed thing: its voice, the steps it showed while it worked, and the reasoning it gave. The persona lived in the prompt, which proved solid for getting consistent behavior. Controlling the base prompt also let me be more deterministic about how the Agent orchestrated its tools. The self-check never judged the way the Agent talked. It judged the quality of what the Agent was about to suggest.

Interaction principles

These are the behaviors I designed for. Each one has a reason, and each showed up in conversations in a recognizable way.

Check before recommending
The Agent tested its own options against the Engine before anyone saw them. People shouldn’t spend their time catching what a script can catch.Example: in the Wi-Fi conversation above, the first try never reached the designer.
Ask what’s true first
Before it suggested wording, it asked what the product actually did and what context was missing. A line can read well and still describe something the product doesn’t do.Example: reviewing a VR error message, it asked whether “try again” matched an action the person really had.
Explain the language
It named the feature of the language that was off and said what that feature does, in place of a technical-sounding score. That gave the person something to act on.Example: both reply patterns above open by naming the feature.
Keep findings apart
A broken rule, a measurement and the Agent’s own judgment are different kinds of claim, and a reply kept them distinguishable. A rule finding was a fact, a measurement was a reading to interpret, and its judgment was open to argument.Example: with the photo confirmation, it gave the reading less weight and argued its own point about “camera roll.”
Put the person ahead of the rule
How a person might be made to feel is more important than any given rule. When its own tool couldn’t recognize the context, the Agent could say why the rule shouldn’t apply.Example: the second reply pattern above.
Bring evidence
Half the battle of content design is justifying a decision. The Agent backed a choice with research, shipped precedent and the right person to ask. When the evidence went the other way, it advised against the choice.Example: the emoji conversation below.
Say what’s missing
When it couldn’t find guidance for a question, it said so. It also corrected itself in the open.Example: asked how Horizon used “world,” “space” and “environment,” it noticed it had searched the wrong design system, corrected the search, and returned an answer with its sources.
Leave decisions with people
A content designer made the call on wording. A partner’s recommendation still required content approval. A change to guidance went to the lead who owned it.Example: the two answers under Adapting to the person.

Adapting to the person

The Agent could recognize a person by their employee ID and job title, and it shifted its strategy to match. If you were a content designer, it could fully jam with you. It talked through options and asked hard questions. The call stayed yours. If you weren’t, it shifted toward a strong recommendation with the reasoning attached, and that recommendation still required content approval. For a team with no content designer, this was how a content design perspective reached the work.

Example: the same question from two people

The question

We’re letting people remove a downloaded world from the headset. Their progress is kept, and they can get the world back from the library. Is “Delete world” the right label?

Asked by a content designer · Verso jams

“Delete world” sounds like the world itself goes away. Two directions:

Remove download

Remove from headset

Hard question back, though: do people need to hear that their progress is safe before they’ll press it? That decides whether the label can carry this alone.

The content designer makes the call.

Asked by an engineer · Verso recommends

I’d use Remove download. “Delete world” sounds like you’re deleting the world itself, and people can still get it back from the library.

I’d send that label to your content designer with a note that progress is kept, so they can approve the wording before it ships.

For content approvalRemove download
Note: progress is kept

A content designer still approves.

Two strategies for the same question. Verso recognized each person by their employee ID and job title. The content designer gets options and a hard question, and the engineer gets one strong recommendation with a route to approval. Either way, the decision sits with content design.

Product knowledge and research

Meta moved quickly, and things were sometimes decided but not documented. One of the major day-to-day problems was institutional memory: who had logged a decision and who hadn’t. So the Agent was connected to much more than the design standards. When you started talking to it, it did a quick grab of your most recent posts and work, which dialed it into what you were probably working on. In specific situations, its instructions told it to double-check that a topic hadn’t been discussed recently or wasn’t on shaky ground.

Recent work
Internal posts, work chats and documents. It could give concrete advice based on the most recent movement on a topic.
Research
User research and data science posts and spreadsheets. It could substantiate an exploration with a past finding, or advise against it.
Shipped strings
Retrieval over the Threads string base, so it could show strings that shared a similar content goal.
Past tests
Stored experiment results, so it could say when something had been tested before.
Design files
Figma, through MCP. It could read the designs a string sat in and help build out a user journey.
Code
Search across Meta’s codebase.
Decisions
A database of decisions logged from content design chats, which the Agent could query.
People
A roster of subject matter experts, indexed by content design work history. It could point you to whoever was closest to a rule or had done related work.

Example: evidence for keeping an emoji

Content designer

Draft notification Horizon

Your friends are here 👋 Come say hi.

I want to keep the emoji, and partners will push back. Anything I can point to?

Verso

  • A coworker was working on this a year or two ago. They ran emoji tests for Horizon that proved effective, and in precise situations they changed global guidance. That supports your hypothesis, so you can use that as a way to navigate pushback.

    Research and data scienceEmoji tests, Horizon

  • Someone mentioned this the other day in a post. Let’s ping them right quick to see if there’s any update.

    A recent postThe other day

  • If you want a second read, someone on the roster has done related work.

    Expert rosterIndexed by content design work history

Where each line came from. Every line of the reply is tied to its source.

The Agent managed this memory itself. It could store a finding, or leave itself a note about what to retrieve later, and it kept or dropped information based on how current it was. The connection also ran back toward the guidance. If you told the Agent a piece of guidance wasn’t true and explained why, it wrote the change as a diff and tagged the lead content designer who owned it. The Index article covers that review.

People have different thresholds for that kind of access. What I can say is that it was very helpful, and it got ahead of some of those day-to-day problems. I thought of it as a contextually intelligent content advisor.

Prompt, skills and tools

People talked to the Agent inside Meta’s internal agent harness. It also worked well in VS Code, and I was working on a set of widgets it could display inline in the harness. An early build had three pieces.

System prompt
992 lines. The persona lived here, along with the instructions for how the Agent orchestrated its tools.
Skills
Seven, each for a specific type of content. They included notifications, error messages, emails, tooltips, release notes and terminology.
Tools
15 integrated tools, including Figma analysis, guidance search, terminology checking, code search and the internal chat tool. I also built two tools from scratch, a reader and a writer for the standards repository.

At one point I gave the main agent sub-agents, each dedicated to one area of expertise. That took the heat off the main agent, which could orchestrate when a question called for it. The Instagram specialist, for example, had direct lines to Instagram’s experiment results and to the JSON version of Instagram’s guidance. The instructions were layered as well, so the Agent carried only what a conversation needed.

Foundation
Always on. Common content guidance, accessibility, terminology, inclusive language, and legal and privacy constraints. This was the floor under every answer.
Product expertise
Per conversation. The product’s voice guidance and team context, with a registry of which sources to search and in what order. Stable product knowledge came first, then a search of that product’s guidance, then broader internal knowledge.
Retrieval
Per request. The Agent looked a team’s standards up each time instead of memorizing them, so its answer was current. It drew on the guidance site, team documents and spreadsheets, terminology checking and the codebase.
Conditional behavior
Only when triggered. Privacy language brought in privacy guardrails, and an engineer got a different posture than a content designer. The check on whether a topic had moved recently lived here too.
Fallback
Last resort. Broad internal knowledge. When the Agent fell back to it, it said so. The person then knew the answer wasn’t grounded in their team’s standards.

Outcome

Designers, content designers and engineers used the Agent regularly across Meta. It was a force multiplier for the discipline in an environment where headcount was dwindling.

Back and forth
The Agent checked its own options before anyone saw them, which prevented a lot of back and forth.
Looking something up
2 to 5 minutes, down from 30 to 60.
Team capacity
Teams of two content designers handled work that had needed three or four.
Institutional knowledge
What had been locked in senior content designers’ heads or scattered across docs became something anyone could query and act on.
Teams without a content designer
Partners got a content design perspective from it, with approval still going to a content designer.

See also

Image