Day in Our Data · Saturday, October 3, 2026, 11 a.m. to 3 p.m.

From idea to demo: the workflow

AI makes writing code fast. What still takes judgment is deciding what to build, checking that it is right, and keeping track of what you learned. This is a loop you can run several times in one afternoon.

The loop

Five stages. Humans own the decisions; Claude does most of the typing. The last stage feeds the first: each loop starts smarter than the one before. The commands in each box come from the Compound Engineering plugin.

  1. Brainstorm

    Whole team · Claude asks questions

    Agree on the civic question, who it is for, and what a good result at 2:30 looks like.

    Output: A one-paragraph problem statement.

    /ce-brainstorm

  2. Plan

    Claude drafts · team edits

    Break the goal into small steps. Decide what to skip if time runs short.

    Output: A numbered plan with a must-have demo.

    /ce-plan

  3. Architecture

    Claude proposes · team decides

    Pick the data files, the tools (a map? a chart? a page?), and where it will live.

    Output: A sketch: data → cleaning → output.

    plan mode (Shift+Tab)

  4. Build

    Claude types · one person drives

    Work one step at a time. Open the result after every step. Commit when something works.

    Output: Working code, committed.

    /ce-work

  5. QA

    Everyone · plus automated tests

    Check the numbers against the raw data, try it on a phone, run the tests, get a review.

    Output: Tests passing and a list of known gaps.

    /ce-code-review

↺ Compound: write down what worked and what broke (/ce-compound), then start the next loop.

The loop on the day

The event runs 11 to 3. A realistic plan is one full loop before lunch and a second, smaller loop after.

Saturday, October 3: where each stage fits
  1. 11:00Welcome
  2. 11:15Plan
  3. 11:45Build 1
  4. 12:30QA 1
  5. 12:45Food
  6. 1:00Build 2
  7. 2:00QA + README
  8. 2:30Demos

A typical prototype architecture

Most Day in Our Data projects share one shape. Name each box when you ask Claude to plan, and it will build in that order.

  1. Raw data

    Already in the event repo

    CSV and GeoJSON files in data/, each with a fetch script that records its source.

    Output: Untouched source files.

  2. Clean

    pandas / geopandas

    Fix types, trim junk, join tables. Keep this in its own script so it can be tested.

    Output: data/clean/*.csv

  3. Analyze

    pandas, plain Python

    Answer the civic question with numbers: counts, rates, change over time.

    Output: A small summary table.

  4. Present

    folium, plotly, streamlit, HTML

    A map, a chart, or a one-page app a resident can understand in 30 seconds.

    Output: Something you can open and demo.

  5. Publish

    A zip or GitHub, then the showcase

    Write the five-heading README, then save your work (a zip from the workspace, or a push to GitHub) and enter it in the showcase.

    Output: Your entry in the showcase.

Who does what

Human and AI responsibilities at each stage
StageThe team decidesClaude doesPrompt to try
BrainstormWhich question matters to residentsAsks clarifying questions, suggests anglesInterview us about starter project 6 until you can write a one-paragraph goal.
PlanWhat is in and out of scopeDrafts steps, flags risks and unknownsPlan this in 5 steps. Mark the must-have demo and what to cut first.
ArchitectureMap or chart? Page or notebook?Proposes files, libraries, and data flowPropose the simplest architecture. No databases unless we need one.
BuildWhen a step is good enoughWrites and runs code, fixes its own errorsDo step 1 only, then tell me how to open the result.
QAWhether the numbers are believableWrites tests, reviews code, checks accessibilityWrite tests that compare your totals to the raw CSV, then run them.
CompoundWhat to remember next timeWrites notes into CLAUDE.md or docs/Add what we learned today to CLAUDE.md.

Compound Engineering in one page

Main commands from the Compound Engineering plugin
Type thisWhat it does
/ce-ideateNot sure what to build? Explore options first.
/ce-brainstormQuestion-and-answer session that ends in a written requirements plan.
/ce-planTurn requirements into a step-by-step implementation plan.
/ce-workCarry out the plan, writing and running the code.
/ce-simplify-codeTidy fresh code for clarity and reuse.
/ce-code-reviewSeveral reviewer agents check the work against the plan.
/ce-compoundRecord the lessons in docs/solutions/ so the next loop starts smarter.
/ce-debugStart from a bug report instead of a feature.
/lfgRun the whole pipeline hands-off. Fun to watch; review everything it produces.