Activity: study-room bookings

Turn a messy booking export into a dashboard your supervisor can trust. About 45 minutes, plus 10 to make it yours.

You work at a campus library. Your supervisor wants a one-page dashboard about study-room bookings for the annual report: how many bookings, when rooms are busiest, and which rooms people want. You have the booking system’s export, and you have Claude Code.

You’ll go around the agent loop once: read → plan → act → check. Then you’ll make the result stick with memory and a skill, undo a bad change, and share what you built. The data is synthetic.

NoteYour finish line

You’re done when you have all of these. A fuller list is on the start page, where you can tick items off.

  • dashboard.html opens in your browser and reports fiscal years
  • Its FY2026 total matches the official count
  • CLAUDE.md holds your reporting rule, and Claude follows it after /clear
  • /house-style gives a page the NC State look
  • You have a link to share (Pro plan or higher), or a screenshot
TipWorking as a pair?

The checker holds two things: the note from your supervisor (step 2) and the official count (step 4). Read the plan against the note before anyone approves it, and compare the dashboard’s number with the official one.

Before you start

  1. Install Claude Code and sign in: see Set up.
  2. Download study-rooms.zip and unzip it to your Desktop. On a Mac, double-click it and drag the study-rooms folder it makes to your Desktop. On Windows, right-click it, choose Extract All, then Extract, and drag the study-rooms folder to your Desktop. The folder holds README.md and room_bookings.csv.
  3. Open a terminal, go into the folder, and start Claude Code:
cd ~/Desktop/study-rooms
claude
cd $HOME\Desktop\study-rooms
claude

If PowerShell says the path doesn’t exist, your Desktop is probably in OneDrive: try cd $HOME\OneDrive\Desktop\study-rooms.

Folder not found, on either system? Type cd and a space, drag the study-rooms folder from Finder or File Explorer onto the terminal window, and press . Then type claude.

  1. Claude Code asks for a quick safety check: is this a project you trust? It is: it holds only what you just unzipped. Choose Yes, I trust this folder (use the arrow keys if it isn’t highlighted) and press . If Claude Code closes instead, that was No, exit: run claude again.
WarningWork in this folder only

Claude Code can read and change everything in the folder you start it in. Starting it in your home folder or Desktop gives it far more reach than this task needs.

Your status line

Before the loop starts, have Claude Code keep what it’s working with on screen. Paste:

/statusline show the model, this folder's name, the effort level, the percentage of the context window used, and the git branch when there is one. Read the branch from the .git folder instead of running git, and test the script before you finish.

Claude writes a small script and adds it to Claude Code’s settings, in Claude Code’s own folder rather than this one. If it asks to create or edit those files, approve. When it’s done, the line under the input box shows the model, study-rooms, the effort level and how much of the context window is used. There’s no branch: this folder isn’t a git repository.

Watch the context number as you go. It climbs with every step, and drops when you /clear in step 5.

1 Read

Ask what’s there before asking for anything else.

What's in this folder? Give me a quick summary of the data. Don't change anything yet.

Claude opens the files itself. A chatbot would have asked you to paste them in. Each step shows as a line: the files it read, and commands it ran, like Bash(wc -l room_bookings.csv). shows the details.

Read its summary closely. It may point out problems in the data. Note what it found; you’ll want it later.

Claude read both files with shell tools and reported: 10,279 rows but 9,930 real bookings (349 rows exported twice), 213 rows with old room labels like Rm 204 instead of Room 204, and some blank group sizes. It ended with “Tell me what you want built, and whether to fix those three first.”

2 Plan

Your supervisor sent you this note this morning:

From your supervisor

Quick reminder for the dashboard: everything we report uses the fiscal year, July 1 to June 30, named for the year it ends. FY2026 = July 1, 2025 to June 30, 2026.

Claude can’t see the note, and the prompt below doesn’t mention it. That’s on purpose. Predict first: will Claude ask you what “year” means, or assume?

Now ask for the dashboard in plan mode, so Claude proposes before it builds:

/plan Build a one-page dashboard, dashboard.html, for my supervisor: total bookings per year, the busiest days and hours, and the most popular rooms. Make it a single file I can open by double-clicking.

The mode line at the bottom now reads ⏸ plan mode on. Claude explores, may ask you a question or two, then shows a plan and asks how to proceed. Don’t approve it yet.

  • If it asks which kind of year: answer from the note. That’s the agent doing its job, and you having the answer. If it shows choices, use the arrow keys and , or choose the option to type your own answer and paste the sentence below.
  • If it doesn’t ask: read the plan. If it counts calendar years (2024, 2025, 2026), it made a reasonable assumption that’s wrong for your office. It couldn’t know; the rule isn’t in the data.

To fix a wrong assumption before any code exists, choose No, keep planning, then type:

We report by fiscal year: July 1 to June 30, named for the year it ends (FY2026 is July 2025 to June 2026). Use fiscal years.

When the new plan uses fiscal years, and covers any data problems from step 1 the way you want, choose Yes, manually approve edits.

In all four of our rehearsals Claude asked which year to use, but if you just said “go,” its default was calendar years every time. Don’t count on being asked. In the default auto mode, Claude is built to keep working rather than stop for questions, so plan mode is where you get to catch an assumption.

Claude asked: “‘Per year’: calendar years? Your data stops on Sep 30, 2026, so 2026 has only nine months.” Its default was calendar years, with 2026 marked “Jan–Sep”. In four rehearsal runs it asked about the year every time, and every time its default was calendar years. After the fiscal-year correction, its revised plan hatched the two partial fiscal years (FY2024 and FY2027) and compared the two complete ones.

3 Act

In manual mode, Claude asks before each change. Approve the first edit or two and look at what it’s writing.

Then switch to auto mode and let it finish. The next time Claude asks to run a command, choose Yes, and switch to auto mode. Or press , with no question on screen, until the mode line reads ⏵⏵ auto mode on (on Windows, if Shift+Tab does nothing, press ). A separate safety check still reviews each action. Press at any point to stop it and redirect.

The build can take 10 to 20 minutes: Claude writes the page, runs its code and checks its own numbers. That’s normal. Wait for it to finish before you type the next prompt: anything you type now reaches Claude mid-build.

When it’s done, open the dashboard. Ask Claude to open it, or:

! open dashboard.html
! start dashboard.html

The ! runs a command yourself without leaving Claude Code. Double-clicking the file works too.

A study-room dashboard from a rehearsal run: bookings per year, the busiest days and hours, and the most popular rooms

A dashboard from a rehearsal run. Yours will look different: same prompt, different result.

Claude wrote the page, then pasted the CSV in with cat rather than retyping it, and checked that the copy was byte-identical. With no browser to look through, it recomputed the key numbers two other ways and said plainly that it hadn’t seen the page. The page worked on the first double-click.

4 Check

A dashboard that looks right isn’t the same as one that is right. Your supervisor already knows one number.

Official FY2026 bookings4,351

Does your dashboard agree?

Whether it matches or not, make Claude prove it. /goal keeps Claude working until a condition holds, and a separate model checks after every turn:

/goal dashboard.html shows FY2026 = 4,351 bookings, the official count. If it doesn't, explain why before changing anything. Prove the final number by counting the CSV yourself, and don't hard-code it. Stop after 5 turns.

Why it might not match: the export lists some bookings twice. Did Claude mention that in step 1? Did the plan say what to do about it? The fix is to count each booking_id once.

A wrong number usually points straight at its cause:

Your dashboard says What happened
4,351 Fiscal year, each booking counted once ✓
4,513 Fiscal year, but 162 bookings counted twice (the duplicate rows)
3,127 Calendar 2026, January to September only
3,696 Calendar 2025, the last full calendar year
3,239 or 3,829 Calendar 2026 or 2025, with the duplicates

If something on the page looks wrong, show Claude instead of describing it. Take a screenshot to the clipboard ( on a Mac, on Windows), then paste it into your prompt with (Mac Terminal) or (Windows) and say what’s off.

The dashboard already said 4,351. /goal made Claude prove it: it confirmed the number wasn’t typed into the page, counted unique booking_ids in the fiscal-year date range two separate ways, and showed that counting rows instead gives 4,513 (162 repeated IDs). The goal was met in one round.

5 Remember

Claude forgets everything when a conversation ends, except what’s written down. You’ll write your rule in CLAUDE.md, a file Claude reads at the start of every session in this folder.

Add our reporting rules to CLAUDE.md: we use fiscal years (July 1 to June 30, named for the year they end), and each booking_id counts once.

Run /memory to see it listed, along with auto memory: notes Claude saves for itself as it works. CLAUDE.md is what you tell it; auto memory is what it learns. Press to close the list; you don’t need to open anything.

Now test it. Type /clear to start a fresh conversation, then ask:

Which fiscal year had the most bookings, and how much did it grow over the year before?

It should answer in fiscal years without being told again: FY2026, up about 40% from FY2025.

After /clear, a fresh session answered: “FY2026 had the most bookings: 4,351. That is up 1,242 (+39.9%, about 40%) from FY2025’s 3,109.” It followed both rules in CLAUDE.md without being told. Along the way, one of its own shell commands printed 0 because of a quoting bug. It noticed, explained why, and fixed it.

6 Make a skill

A skill is a reusable set of instructions with a name. You run it as a command, and Claude can also use it on its own when it fits. (Skills are what Claude Code used to call custom commands.) Have Claude write one:

Create a skill called house-style that gives any web page we make the NC State look: Wolfpack Red #CC0000 for highlights, dark gray text on white, Arial or Helvetica, a red header bar with the page title, and a footer that says "Synthetic data made for a workshop." Use it whenever you create or restyle a web page.

Claude writes .claude/skills/house-style/SKILL.md. This folder had no skills before, so Claude Code isn’t watching that folder yet. Tell it to look once:

/reload-skills

Then run your skill:

/house-style dashboard.html

Reload the page in your browser. If Claude Code says Unknown command: /house-style, run /reload-skills and try again. Change the skill later? Run /reload-skills again.

Now ask for something new, without mentioning the style, and watch whether Claude reaches for the skill on its own:

Make a one-page summary for the library's dean, summary.html, with the three numbers that matter most.

For the dean’s summary, Claude used the house-style skill without being asked. It also chose a third number on its own: +6% growth in the original twelve rooms, next to the +40% headline, “because the headline +40% alone would mislead”: most of the growth came from the four new rooms. See it on the finished examples page.

7 Undo

Everyone makes a change they regret. Make one on purpose:

Make dashboard.html rainbow-colored, with Comic Sans everywhere.

Look at it. Then type /rewind (or press twice on an empty prompt), pick your rainbow prompt itself, and choose Restore code. Rewind goes back to just before the prompt you pick. Reload: it’s back.

WarningWhat rewind can’t undo

/rewind restores files Claude edited with its own editing tools. It doesn’t undo what shell commands did, such as deleting (rm) or moving (mv) files. For real work, keep copies or use git.

8 Share

Publish dashboard.html as an artifact.

Claude gives you a link to a private page on claude.ai. Open it, choose Share, and you can make a public link anyone can open. This needs a Pro, Max, Team or Enterprise plan. On a Team or Enterprise plan, public links are off unless your organization turns them on, so share within your organization instead. On a Console account, share a screenshot. /artifacts lists what you’ve published.

Make it yours

You have ten minutes and an agent. Pick one:

  • Improve the dashboard: a filter by who booked, a dark mode, a button that downloads the numbers.
  • Build something you’d use: a study timer in NC State colors, a flashcard quiz from your own notes, a weekly planner. Start in a new empty folder: /exit, then cd .., mkdir my-project, cd my-project and claude, the same on Mac and Windows.
  • Try your own data, as long as it’s data you’re allowed to share with an AI service.

Give it a finish line Claude can prove from its own work, such as:

/goal every button on the page has code behind it that visibly changes the page, and the page needs no other files. Prove it from the code. Stop after 5 turns.

If something goes wrong

What happens Try
Claude asks a question instead of planning Answer it. That’s the plan step working.
The page opens blank Paste a screenshot and say “this opens blank when I double-click it.”
Your number doesn’t match and you’re stuck Ask: “List every reason the FY2026 count could differ from 4,351, and check each one.”
A window offers to install “command line developer tools” (Mac) Claude tried a tool your Mac doesn’t have. Choose Not Now; it will find another way.
Claude has gone in circles twice /clear and ask again, with what you learned in the first prompt.
Claude says “You’ve hit your Opus limit” Type /model sonnet and carry on.
You’re out of time Download the finished example and compare.

Checked against Claude Code 2.1.299 on October 14, 2026.