← ClaudeAtlas

pbi-redesignlisted

Turn an extracted Power BI model into a dashboard worth acting on. Use when someone wants more than a like-for-like rebuild - a strategic view, an executive summary, "what does this data actually say". Runs as a staged workflow, analyst then strategist then designer then builder then critic, with the evidence computed deterministically and only the judgment left to the agent.
starmynd-org/powerbi-to-ai-visuals · ★ 1 · AI & Automation · score 69
Install: claude install-skill starmynd-org/powerbi-to-ai-visuals
# Design the report they should have had A faithful replica answers "what did the report say". This answers "what should somebody do on Monday". Those are different questions and they want a different page. The split that makes this work: **evidence is computed, judgment is written.** `tools/analyze_model.py` produces facts. You produce `narrative.json`. The builder merges them. Nothing on the finished page asserts anything that is not traceable to a query or to a line you wrote deliberately. ## Stage 1 - analyst ```bash python3 tools/analyze_model.py --model build/model -o build/analysis ``` Writes `findings.json` and `findings.md`: detected roles, period coverage, totals, by-year growth and margin, mix, concentration, year-on-year movers, data-quality findings, and which tables the source report never touches. Read `findings.md` before anything else, and check the roles first. They are regex guesses over column names and every one is recorded with its evidence. If the revenue column or the date table is wrong, everything downstream is wrong: fix it with `--roles roles.json` and rerun. ## Stage 2 - strategist Decide what the page is *for*, from the findings and not from habit. Ask: - What is the largest number nobody is looking at? - Which fact changes a decision, rather than describing the past? - What does the existing report structurally *cannot* show? Unreferenced tables and unused columns are the cheapest place to look; on the worked example the source repo