Bruna Prado

Tools

• Claude Coderuns the 11 Malha GTM skills in order, talking to whoever is answering
• Pythonscripts that check every quote against its source and score the rubric
• HTML graphthe navigable web that links source, fact, persona and positioning

A persona points back to its source

persona: payroll analyst pain: "redoing payroll because of a spreadsheet error" evidence: won call 02, CS round 1 confidence: high (3 sources)

Decisions

• Malha GTM has a master skill that reads the state, decides the step and picks up where it left off. Each skill also runs on its own.
• Gates: a step only opens when the previous one has enough material.
• Without a source, the claim becomes a hypothesis.

Tools

• Astrobuilds the site's 58 pages, fast and static
• CSS with tokensa single file holds colors, fonts and spacing; everything else just uses the variables
• Lucidea line-icon library, the same across the whole site
• Playwright + axeopens each page in an automated browser and checks accessibility

Decisions

• Colors only exist in the tokens file. No component has a loose hex value.
• The hero shows the live product, not a laptop mockup.
• The website copy comes from the ground-floor messaging, not from a standalone brief.

Tools

• n8nconnects the steps and holds the flow at the pause until approval
• Ollama (qwen3:8b)the local language model that researches, writes and reviews
• ComfyUI (FLUX.2 klein)opens on its own and creates each article's cover illustration
• Kokororecords the narration in Portuguese, with a natural voice
• Vela RH blogwhere the approved article is published

The node that holds everything

Wait → on webhook waits for: approved | request changes timeout: none (only a human releases it)

Decisions

• The pause is a visible node in the flow, not a hidden step.
• The writer learns rhythm from human texts and the reviewer flags what sounds like AI.
• The gaps the 4th floor finds become topics here.

Tools

• Social Media agentreads the site and writes the posts in the brand's voice
• Designer agentdefines the visual idea and format of each asset
• Reviewer agentreviews copy and artwork, and sends back what doesn't pass
• ComfyUIgenerates the images on my computer, with models licensed for commercial use
• Scheduling stepafter the OK, asks whether it can schedule: on the calendar's dates, on yours or on dates the agent picks
• Buffer API (beta)schedules the posts on the networks

Decisions

• Write like the brand: caption structure, expressions and emojis come from real posts, free of AI tics.
• No claim that isn't on the website: each post keeps the page it came from.
• One article becomes several angles, not the same text shortened.
• The schedule respects the cadence: no dumping everything on the same day.

Tools

• Nurturing agenttrained to build the strategy and write the sequences from the company's context
• Reviewer agentchecks tone, links and promises before approval
• “entrada” folder or URLwhere the agent reads the context from: website, Drive or a folder on the computer
• CRMreceives the approved sequence in the import format

Decisions

• One sequence per persona and per trigger: the payroll analyst and the owner don't get the same emails.
• Every email leads to a page on the company's own site: an article, a tool, a customer story or a demo.
• The top doesn't sell: it teaches. The product only shows up at the bottom of the funnel.

Tools

• Claude Code (skill)runs the steps and writes the storyline
• Playwrighta windowless browser that reads the site: fonts, colors, logo and screenshots
• pptxgenjsbuilds the .pptx with the 15 slide types
• fontkitmeasures every text with the real font so it fits
• PowerPointrenders the images for the visual review and the PDF
• Python (PyMuPDF, Pillow, fontTools)reads PDF, PPTX, images and fonts from the “entrada” (input) folder

Decisions

• The brand comes from the site's code, not from a screenshot: colors and fonts come out exact.
• A tested layout became 15 slide types: the skill picks the type instead of drawing from scratch.
• Proof comes with a real product screenshot, never just a sentence.
• For a real company, only public content, and the cover says it's an unofficial study. I tested it that way with a real brand, outside the portfolio.

Tools

• n8n (7 flows)schedules and chains each step of the report
• DuckDBcalculates the metrics in SQL, straight from the data
• Python + FastAPItakes the pipeline snapshot and checks the numbers
• Ollama (qwen3:8b)writes and reviews the reading of each section
• Jinja2assembles the report in HTML

The lock on every section

Writer (LLM) → Numeric check numbers match the metrics? yes → text approved no → request a fix → Writer → check again

Decisions

• The data source is registered in the workflow: the CRM API feeds a database (or a spreadsheet), and that source is what makes the report come out on its own every Monday. At Vela, the database holds fictitious data.
• The LLM never calculates: the numbers come ready from the database and it only explains them.
• Each section goes through a numeric check: a number with no origin goes back to the writer.
• Cohorts are compared at the same age: comparing a new cohort with a mature one invents a decline.
• Headline, decisions, central diagnosis and health color come from rules; the LLM writes the reading, which sits collapsed in each section.
• Everything local and free, with 100% fictitious data and planted stories to test the agent.