- Inteligência Artificial
- Governança de IA
- Claude Code
- Produtividade
- Automação
- Transformação Digital
AI Skills by Team
Productivity Beyond the Dev Team
When people talk about AI at work, almost every example comes from the development team: generated code, written tests, reviewed pull requests. That makes sense, because developers already live in the terminal. But most of a company's repetitive work rarely sits in the code. It sits in the card copied by hand, the Friday report, the test spreadsheet built from scratch and the email rewritten three times for three audiences. In AI Governance in Practice and Skills Hub I explained why InBot chose to upskill its teams instead of shrinking them, and how skills reach each person's machine safely. Now I show what these AI skills deliver, team by team: the problem before, what the skill does and what changed. Above all, I show why the big gain does not come from a single skill, but from one skill feeding the next.
What a Skill Is, and Why the People Who Know the Process Should Write It
A skill is a folder with instructions, examples and sometimes scripts that Claude Code or Codex loads when a task calls for it. Anthropic uses a comparison I really like: building a skill is like putting together an onboarding guide for a new hire.
That comparison says a lot about who should write the guide. A language model knows a lot in general and nothing about your team's card template, the rule QA applies before opening a bug or the way sales records an opportunity. The team knows that. A skill is the team's process written in a way the AI can follow.
A technical detail makes it possible to have dozens of them installed. According to the official documentation, only each skill's name and description stay loaded all the time, about 100 tokens per skill. The body of the skill only comes in when it is triggered, and supporting files only when they are read. In practice, a catalog of ~50 skills does not weigh the assistant down. And the format is not tied to a single vendor: Codex uses the same open standard, so the same skills run in both tools.
Who builds them: it depends on the team
I would like to say that every team at InBot writes its own skills on its own. That is not true, and pretending it is would hide the most interesting part of the work.
Today, UX and Content, QA and Development build with a lot of autonomy. They bring the process, write the skill, test it and publish it through the Skills Hub, and the AI Enablement team reviews it before it goes to production. In the other teams, the AI team takes a bigger part in building. Even so, the shift has started: Sales published a skill of its own through the Hub, without anyone from the AI team writing it for them.
The barrier is the tool, not the process
What holds back the less technical teams is not a lack of knowledge about their own work. They know exactly what they do every week. What holds them back is familiarity with the tool: the terminal, Claude Code, a command to paste, a configuration file.
This is not unique to us. A BCG survey of more than 10,000 workers found that only 36% feel properly trained to use AI, and an independent Jobs for the Future survey reached the same number in 2026. The same BCG study shows that people who get five or more hours of training, ideally in person and with coaching, are much more likely to become regular users. And Gallup found that employees whose manager actively supports AI use it frequently 2.1 times more often.
Our answer had two fronts:
- Close support. The AI team works closely with people who are getting started, maps the process together with them and stays available when something does not work.
- An installation page designed for non-developers. In the Skills Hub, people choose their operating system, download the installer and paste a single command, with instructions on how to open the terminal on Mac, Linux or Windows. Anyone who prefers can ask Claude Code or Codex to run the installation. The page has an FAQ section for the most common errors, and the installer's signature is checked before it runs, so nobody needs to understand security to be protected.
AI Skills by Team: What Changed in Each One
From here on, I go through each team in the same format: the problem before, what the skills do and what changed. I will not use the skills' internal names or talk about clients. What matters is the kind of repetitive work that no longer takes up people's week.
UX and Content: FAQs, journeys and personas
I start with the heart of the product. InBot builds virtual assistants, and each one needs a persona (who the assistant is and how it talks), a question-and-answer knowledge base the AI consults and journeys, the guided service flows. All of this used to be built by hand, in spreadsheets in the platform's format, with a lot of variation from one project to the next.
Today there are skills for each piece: one creates the persona, another builds the knowledge base, another builds the journeys, another explains and debugs the markup language used in the bot's content cards, another draws the flowchart, and a last one reviews the configuration before the bot goes to production.
Two results show the care behind this:
- 800 automated tests to decide how a micro-agent should validate user input: with code or with the model alone. The "model only" version was the only source of errors across the runs, and the skill now generates the validation in code.
- The skill that explains the content card language was born for the UX team and is now distributed to 7 teams. Knowledge that used to depend on asking the right person became a lookup anyone can do.
Planning: from the meeting to the document
Planning turns meetings, briefs and transcripts into a project planning document. It used to take hours of reading and reorganizing for every new project.
The team's skills write the executive summary of the meeting, build the planning document in the official template, run a privacy check (personal data, purpose, likely legal basis and gaps) before the project starts and create standardized cards in ClickUp. Card creation passed 5 out of 5 test scenarios, and a project card is born with the team's full pre-kickoff checklist.
One lesson came from here: when we looked at what the teams were asking for, we found three requirements-gathering skills being reinvented in parallel, one in Planning, one in UX and one in Projects. They became one. Governance is also this: spotting duplication before it turns into three versions of the truth.
Projects: cards and schedules without manual copying
Projects is where the "before" is most striking:
- for every new project, someone copied a 79-card template list by hand and edited the cards one by one;
- every week, the manager read 40 to 80 cards per project to know what needed to happen;
- the Friday status report was personalized by hand, in a portfolio that reached more than 50 projects per analyst.
Now the schedule comes from the planning document, the cards are created from the template (always with a preview before saving), and the weekly summary and status report read what was created. The status report comes out in three versions: internal, client-facing and executive. An audit compares periodic snapshots of the lists and flags cards with no owner, deadlines pushed without justification and cards closed without evidence.
The team also got a communication validator, which rewrites the same email for different audiences and checks that it has an owner, a deadline and next steps. In its first version, it got 56 out of 60 checks right (93%). After the analyst's feedback, it went through 77 more scenarios before going to production.
QA: tests generated from business rules
In QA, the classic problem was the inherited test spreadsheet: you copied the one from a similar project and adapted it, carrying over scenarios and terms that had nothing to do with the new product.
The test spreadsheet skill starts from scratch, from the product's business rules, and also generates a dashboard where QA marks the execution. In a recent review, a spreadsheet of 212 cases shrank to 137 rows covering the same cases. Another skill tests the bot end to end in read-only mode: the happy path, field validation, attempts to manipulate the assistant and sensitive content.
Development: one standard where there used to be five
For the dev team, the gain was less about speed and more about standards. There were five different definitions of code review going around, and each person reviewed in their own way. They became one.
Other skills standardize branch and PR names and keep the ClickUp card in sync with the PR's status, drive deployment across environments and generate the feature's documentation, checked against the actual code. One of them drives the single sign-on setup for new clients, a process that used to require reading about 1,900 lines spread across ten documents. If you want to see what this looks like in a developer's individual workflow, I described mine in My AI Development Workflow.
Sales: CRM and routine without rework
In Sales, the skills operate the CRM with fixed rules: they search before creating, avoid duplicates, ask only for what is missing and confirm before saving. Before approval, the operation went through 250 runs in the CRM, with 250 correct results.
The team's daily dashboard and the weekly report on new companies also come from there. They only read data and leave the email as a draft for a person to review and send. And Sales itself published a skill through the Hub to organize the weekly follow-up of its strategic missions.
Marketing: one theme, every piece
Marketing had a very specific kind of rework: asking the AI for each piece separately and then copying everything into the editorial calendar. Now a theme from the calendar becomes a blog post, an Instagram post, a LinkedIn post and an email at once, and the card is created in the calendar with the publication date. Before going to production, the skill went through 80 scenarios across four rounds of testing.
The pieces every team reuses
Some skills do not belong to any team, and they may be the most important ones. One operates ClickUp from the terminal, another operates the platform's data tables, another sends email through the corporate account, another writes meeting minutes and another interviews the person, one question at a time, until the request is clear. The teams' skills do not reimplement any of this: they call these pieces.
In the same category are the skills that operate the company's systems from a request in plain Portuguese. The one that operates the platform's admin panel turned a 3-hour task into 13 minutes. With it, we also analyzed a 30-day A/B test between two language models in our chatbot, comparing tokens and response time. The platform shows one day at a time and does not export data, so this used to require daily screenshots and manual math. It took about 20 minutes.
Chaining: a Bot Project from the Meeting to Production
So far, each skill looks like a saving of a few minutes, sometimes a few hours. The leap shows up when you look at an entire delivery. Anthropic's documentation lists among the benefits of skills precisely combining several of them for multi-step tasks. At InBot, this became an architecture rule: one skill plans and another executes, and the output of one is the input of the next.
Here is the path of a bot project:
- Meeting minutes: the transcript becomes decisions, open items and owners.
- Planning document: the minutes and the brief become the project's official plan.
- Privacy check: the plan is read looking for personal data and risks before anything is executed.
- Cards and schedule: the plan becomes the schedule and the cards in ClickUp, with a preview.
- Persona, knowledge base and journeys: the bot's content is generated in the platform's format.
- Test spreadsheet and bot QA: the scenarios come from the business rules and the bot is tested end to end.
- Pre-production qualification: the configuration is reviewed before the bot goes live.
- Deployment plan: the plan reads the deployment card and points to QA's post-deployment checklist instead of reinventing the scenarios. Its rule sums up the whole philosophy: nothing without a source becomes a guess; it becomes an open item with an owner.
None of these steps is revolutionary on its own. The difference is that the quality of one guarantees the next. A note on one of the cards in this chain puts it well: it is the rigor in creating the card that makes the report come out ready later.
When one team fixes another team's skill
Chaining also works across teams. In a QA round, the testing skill found that 35% of a bot's content was out of the AI's reach: a configuration field cut the text without warning, and part of the answers was simply never found. The finding went back to UX, and the skill that builds the knowledge base now blocks generation when this problem appears. The defect was found once, and the skill keeps it from happening again in future projects.
A study with 776 Procter & Gamble professionals found something similar at scale: with AI, proposals from R&D and from Commercial people became more balanced, as if the AI brought each one the other team's perspective. Here, what brings one team's perspective to the other is the skill.
Skills that orchestrate other skills
The next step was natural. The skill that builds the deployment package for a simple bot was, at first, self-contained: it handled persona, knowledge base and journey on its own. It was rebuilt to only orchestrate, delegating each piece to the specialist skills. It became 67% smaller and produced the 7 deliverables in 11 out of 11 test scenarios. In another case, a demo bot was built with 7 skills in sequence, from the journey map to the task card.
This has a practical consequence: when the persona skill improves, every package that depends on it improves too, without anyone editing the orchestrator.
A Human in Control
None of this works if the teams do not trust the skills. And trust came from a pattern that repeats in almost all of them:
- Preview and confirmation before saving anything in ClickUp, the CRM or the platform.
- Read-only wherever possible: reports and bot tests do not change anything.
- Email as a draft: the skill writes, a person sends.
- A test suite before the release, with real cases and more than one model, as Anthropic itself recommends.
- Whoever wrote the skill is not the one who approves it: every publication goes through review.
The tests have already changed decisions. In the status report, with vague or contradictory documents, the cheapest model got it wrong and only the mid-tier one got it right. The rule became: the budget model does not run in production without human review. I explained the logic of this governance in AI Governance in Practice.
Productivity Without Hype: What Can Be Measured
"Extreme productivity" is the kind of phrase I distrust when I read it. So it is worth separating what we measured from what we did not.
External data shows that the gain is real, but uneven. In the BCG study with Harvard, consultants using AI completed 12.2% more tasks and worked 25.1% faster on tasks within what AI does well. On a task outside that frontier, they were 19% less likely to get it right. At MIT, business professionals took 40% less time on writing tasks, with 18% higher quality.
In aggregate, the number is much more modest. The St. Louis Fed estimated that people who use generative AI save on average 5.4% of their working hours. A study using official data from Denmark found no effect on wages and hours two years after ChatGPT. The average hides the concentration: the big gain shows up in specific tasks, and it does not show up when the tool is simply rolled out without changing the process. BCG says it plainly: companies cannot simply roll out AI tools and expect transformation.
That is why I believe in skills by team, not in "granting access". Gallup found in 2026 that the share of people who perceive a gain rises from 45%, among those who use AI for one or two tasks, to 90%, among those who use it for seven or more. One skill saves minutes. A set of skills that covers a team's whole workflow changes what that team can get done in a week.
And here is the honest part: we have not measured hours saved for most skills. What we have are test suites, accuracy rates, size reductions and feedback from the teams. The time figures I cite, like the 3 hours that became 13 minutes, are measured cases, not averages. The big gain is in repetitive and lookup tasks. In judgment tasks, the skill organizes the work, and the decision stays with the person.
How to Start with AI Skills in Your Team
In four months, InBot's teams got 77 initiatives off the ground and about 50 skills and automations into production. The domain merit belongs to each team that opened up its own process. The AI Enablement team comes in to bring this up to the company's standard, test it and make sure it arrives safely. If I had to sum up what we learned as a roadmap, it would be this:
- List your team's repetitive tasks. Not the hardest ones: the ones that repeat every week and follow the same format.
- Start with a small, read-only skill. A report, a summary or a check. It is the safest way to build trust.
- Write the process before the prompt. If nobody can explain the process to a new colleague, the AI will not be able to follow it either.
- Test with real cases before distributing, and with more than one model.
- Think about the next skill from the first one. Define the output of one so that the next one can read it. That is where the gain stops adding up and starts multiplying.
- Stay close to the people who are getting started. The barrier is almost never the process; it is the tool.
Conclusion
AI skills by team are not a list of shortcuts. They are each team's process written in a way the AI can follow, with a person checking what matters. One skill on its own saves minutes. Chained together, they carry a delivery across the company, from the meeting minutes to the deployment plan, with the quality of each step guaranteeing the next.
If you lead a team, my suggestion is to look at your next week and mark one task that repeats. That is your first skill. The second will show up when you notice who uses its output. And if you want to think about what this means for professions in the long run, I wrote about it in AI and the Future of Professions.
References
- Zhang, B.; Lazuka, K.; Murag, M. "Equipping agents for the real world with Agent Skills". Anthropic Engineering, 2025.
- Anthropic. "Agent Skills" (overview).
- Anthropic. "Skill authoring best practices".
- OpenAI. "Build skills" (Codex and ChatGPT).
- Dell'Acqua, F. et al. "Navigating the Jagged Technological Frontier". Harvard Business School Working Paper 24-013.
- MIT News. "Study finds ChatGPT boosts worker productivity for some writing tasks", 2023.
- Dell'Acqua, F. et al. "The Cybernetic Teammate". NBER Working Paper 33641, 2025.
- Bick, A.; Blandin, A.; Deming, D. "The Impact of Generative AI on Work Productivity". Federal Reserve Bank of St. Louis, 2025.
- Humlum, A.; Vestergaard, E. "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI". NBER Working Paper 33777.
- BCG. "Companies Must Go Beyond AI Adoption to Realize Its Full Potential" (AI at Work 2025).
- Jobs for the Future. "Worker Anxiety Over AI Is Growing, and Employers Aren't Preparing Employees for What's Next", 2026.
- Kemp, A. "Manager Support Drives Employee AI Adoption". Gallup, 2025.
- Gallup. "AI and Workplace Productivity: What the Data Show", 2026.
- Mierzwa, Matheus. "AI Governance in Practice" (Portuguese).
- Mierzwa, Matheus. "Skills Hub" (Portuguese).
- Mierzwa, Matheus. "My AI Development Workflow" (Portuguese).
- Mierzwa, Matheus. "AI and the Future of Professions" (Portuguese).