Setting up a company’s social media presence is the kind of task that usually gets decided on intuition: “everyone’s on Instagram, so let’s go to Instagram”. We did the opposite. We built Anjos & Brum’s presence with AI at every stage — and there was a single rule: no decision without a criterion underneath it — either a piece of data or a house rule. This is the honest account of what got produced, what the data decided and, the part that paid off most, what was aborted halfway through.
The workbench: which tools we work with
Today we use Claude for internal needs, whether for text or for code. For generating and improving images, we use ChatGPT.
Our process includes a few in-house skills. A skill is a reusable package of instructions, examples and scripts that the AI loads before it starts working — in practice, a house procedures manual: instead of repeating the brief and the brand rules with every request, the skill makes sure the output always lands within the standard. That is how we standardise the work and gain performance.
What got produced
The LinkedIn page came out complete: tagline, “About” section, slogan and banner. The starting point wasn’t a blank prompt — it was our official design system. The brand manual (palette, typography, gradient rules, logo variations) went in as the prompt’s contract: the AI executes the system, it doesn’t invent a style. That’s the difference between asking for “a nice banner” and asking for “a banner within these rules”.
Along the way a new skill was born — one that generates the visuals for our posts (carousels and cards, always in the brand identity), with automated content validation and a visual inspection before delivery. And we closed a publishing workflow tested end to end for the first time: blog article, text adapted for the network, visual generated by the skill, publishing checklist. It wasn’t a lab test — the post actually went live. During the inspection, the machine also caught a duplicated URL on the card before publishing. A small detail — and it is precisely because they are small that details like these so often escape human review.
What the data decided
Before designing anything at all, there was always research and context-building first. Three examples of what the research changed during planning and execution:
-
A custom banner generates roughly 2.1 times more views than the default blue banner, according to the 2025 LinkedIn Profile Engagement Report, cited in industry guides (Zeliq, 2026). Conclusion: yes to a banner — designed carefully and within the standard.
-
A company page’s banner displays at 1128×191 pixels and, once you discount the edge cropping and the logo overlapping the bottom-left corner, little more than 150 usable pixels of height remain (specifications compiled by ConnectSafely, 2026). Conclusion: a single tagline, and no logo inside the banner — LinkedIn already shows it underneath, so duplicating it wastes the most visible space on the page.
-
Posts from personal profiles perform better than those from company pages — in Metricool’s 2026 study, 2.60% engagement (interactions per post) against 1.60% (Metricool, 2026) — and the mechanics explain why: the most common reason for a post to reach 2nd-degree connections is an interaction from someone in the 1st degree, meaning the algorithm tests the content first on the network closest to whoever published it (algorithm guide, DSMN8, 2026). Conclusion: the page is the shop window; distribution will happen through the partners’ profiles.
What was aborted
This is where AI saved us most — not for what it did, but for what it did not do.
At the start, the plan was to create and feed both the LinkedIn and the Instagram profiles. After the research, the Instagram calendar died before it was born. The initial plan called for four to five posts a week; the numbers said otherwise: in follower-based measurements, the average engagement per post for a company account on Instagram sits between 0.3% and 0.5% and keeps falling by double digits year on year (Socialinsider and Rival IQ data, compiled in the 2026 benchmarks), whereas LinkedIn converts visitors into leads (business contacts) on the order of 277% better than entertainment networks — the original study is HubSpot’s, from 2012, and it remains the most cited one in current benchmarks; the age of that figure is worth keeping in plain sight. The practical decision: we reserved the @anjosbrum handle, left the profile presentable, and stopped there. Zero calendar. The energy goes where the data points.
And not everything came from market research — there was also the brake applied by our own manual. One of the generated banners came out with the gradient inverted, blue flowing into purple instead of the other way round. It would have gone unnoticed by almost everyone; validating the published page against the design system caught the error before it became the standard. It’s the same principle as the duplicated URL: systematic verification doesn’t exist for the big errors, which anyone can see — it exists for the small ones, which nobody sees.
What doesn’t exist yet
Engineer’s honesty: the workflow doesn’t run on its own yet. Automated orchestration — the layer that fires everything off without someone pressing a button — is the next step, not the current one. What is proven is the core: from idea to published post, with the right brand and review built in. And the figures cited above are market benchmarks, not our own measurements; that is exactly why they were useful — to decide where not to spend energy, before spending it.
The lesson
The value of AI over those two days wasn’t speed. It was that it disagreed — with me, and with the very plan I had drawn up the day before — when the data said so. Strictly speaking, there was no will of its own involved: the criteria were mine, and the difference is that the machine applied them without my temptation to work around them. That’s what I mean by “disagree”. A tool that only produces is a fast intern. One that puts the brakes on with good reason is a decision-making partner.
This article is, in fact, the first piece of the very workflow it describes: from it comes the LinkedIn version, the visual from the skill, and the publication. Behind the scenes, working in public.
And if you’re setting up your own company’s presence, the next step isn’t opening accounts — it’s writing down, first, the criterion that will make you close them.