Turn raw occupancy data — Wi-Fi data, badge swipes, sensor data, room reservations, manual surveys — into a clean, decision-ready space-utilization report. Vendor-neutral. MIT licensed.
Built by Occuspace.
Markdown instructions that teach an AI assistant how to analyze occupancy data. The format follows Anthropic's Agent Skills spec, but the contents are plain text — Claude, ChatGPT, Gemini, and other LLMs can all use it. Set up once (see Install), then share your data and ask.
You get back:
- Headline metrics — average daily peak, typical daily peak (P90), single highest peak, Utilization vs. capacity
- Patterns — by day of week, by hour of day, plus a day-of-week × hour-of-day heatmap
- Trend detection — linear regression with R², gated at a 90-day minimum (shorter windows are dominated by weekly seasonality)
- Anomaly flags — days that look unusual versus their same-day-of-week baseline
- Source-aware caveats — Wi-Fi devices ≠ headcount, badge swipes ≠ occupancy, reservations ≠ presence
- Prioritized recommendations — specific, scoped, testable actions tied to a finding
Two output modes: inline markdown for quick questions in chat, or a self-contained HTML report when you ask for a deliverable. See examples/example-report.md (chat answer) and examples/example-report.html (deliverable) for what each looks like.
At Occuspace, we've spent 7 years analyzing occupancy and utilization data across 1000+ buildings and 40 million square feet. This skill packages what we've learned about how to frame that data — what to compare to what, when a trend is real vs. noise, what makes a recommendation actionable. The framing is the part that's portable.
- Open a new Project at claude.ai.
- Paste
SKILL.mdinto the Project's custom instructions. - Upload
references/analysis-recipes.md,references/writing-style.md, andreferences/multi-source-data.mdas project files. - Share your CSV and ask Claude to analyze it.
The optional Python helpers don't run here, but Claude's analysis tool can compute the same math when needed.
git clone https://github.com/Waitz-Inc/occupancy-insights-skill ~/.claude/skills/occupancy-insightsRestart Claude Code if it was open — the skill auto-discovers. Update later with git pull from the same directory.
Same pattern across platforms: paste SKILL.md as the assistant's instructions, attach the three references/ files, and enable code execution if you want the Python helpers to run.
- ChatGPT (Custom GPT) — instructions field + Knowledge files + Code Interpreter
- Gemini (Gem) — instructions field + attached files + code execution
- Direct API —
SKILL.mdas system prompt + references in context
Output quality scales with the model. The skill produces useful reports on any capable instruction-following LLM.
Share your data and ask. Examples:
- "Here's a CSV of Wi-Fi data for our HQ for the last 4 months. How busy was it? Any trends?"
- "This is daily badge swipe data for our research building. What's the day-of-week pattern? Any unusual days?"
- "Here's sensor data for our café over Q1. Build me a heatmap and tell me when to schedule cleaning."
The assistant figures out your column layout, asks one or two clarifying questions if needed, and produces a structured report (see examples/example-report.md).
occupancy-insights-skill/
├── SKILL.md Main skill instructions (~180 lines)
├── references/
│ ├── analysis-recipes.md Concrete recipes: peaks, DOW, heatmap, trend, anomalies
│ ├── writing-style.md How the report should read
│ └── multi-source-data.md Wi-Fi vs badge vs sensor vs reservations vs manual
├── scripts/
│ ├── compute_trend.py Stdlib-only linear regression + 90-day guard
│ └── detect_anomalies.py Stdlib-only z-score vs same-DOW baseline
├── examples/
│ ├── sample-wifi.csv 30 days of hourly Wi-Fi device counts
│ ├── sample-badge.csv 30 days of daily badge entries
│ └── example-report.md What the skill produces on the sample data
└── LICENSE MIT
The skill auto-detects columns by content, not by name. Bring a CSV with at least a timestamp and a count — it figures out the rest. The table below is just to set expectations on which kinds of data work.
| Source | Common vendors | Typical columns |
|---|---|---|
| Wi-Fi presence | Cisco Spaces, Cisco Meraki, Aruba ALE | timestamp, location, device count |
| Access control | Kastle, HID, Genetec | event time, reader/door, cardholder, event type |
| Sensor headcount | Occuspace, Density, Butlr, VergeSense, XY Sense, RZero, Freespace | timestamp, space, count, capacity |
| Reservation systems | Microsoft Graph (Outlook/Teams), Google Calendar, Kadence, Robin, Condeco, Envoy, Spaceti | start, end, room, attendees |
| Manual surveys | Spreadsheets, observation apps | date, time, space, observed count |
Date formats: ISO-8601 (2026-01-15), YYYY-MM-DDTHH:MM:SS, MM/DD/YYYY, DD/MM/YYYY, ms-epoch integers. Custom formats work via the --date-format flag on the helper scripts.
Kept lite on purpose. It won't:
- Pull data for you. Bring your own CSV.
- Predict the future. It describes what happened — it doesn't forecast.
- Detect data quality issues like sensor outages, flatlines, or coverage gaps.
- Discover operating hours automatically. You tell it the hours, or it reports 24/7 with a caveat.
- Roll up portfolios of more than a handful of spaces.
- Run in real time.
If any of those matter for your use case, that's where Occuspace comes in.
When you're ready for:
- Real-time utilization data at minute-level resolution
- Automatic operating-hours discovery from your actual usage patterns
- Portfolio rollups across hundreds of spaces and floors
- A managed dashboard + AI analyst that proactively surfaces what changed and why
→ Talk to us at occuspace.com.
Issues and pull requests welcome. The most useful feedback is misleading-framing examples — open an issue with the input shape and what you'd expect instead.
MIT — use it commercially, fork it, embed it in client deliverables. No attribution required, but appreciated.