✓ Free · MIT License Manufacturing Healthcare

OEE from CSV

Calculate Availability × Performance × Quality = OEE and a Six Big Losses Pareto from any MES, SCADA, historian, or manual downtime export. Formulas shown with actual numbers. Never a bare OEE percentage without the loss breakdown.

⬇ Download Skill How it works ↓

OEE = Availability × Performance × Quality

Run Time ÷ Planned Time

Ideal Cycle Time × Total Count ÷ Run Time

Good Count ÷ Total Count

OEE

Often-quoted target: 85%

The Problem

A bare OEE number tells you nothing to act on

Most teams can tell you their OEE. Almost none can tell you which loss category is driving it, or whether they're even calculating it correctly. The six traps below are almost universal, and this skill calls them out rather than letting them hide in the math.

⚠️

Planned downtime counted as loss

Scheduled breaks and other schedule loss, such as no orders or no staffing, come out of planned time. Planned changeovers stay in Availability, because they are a main target for SMED. Decide how you treat planned maintenance and use the same rule every time. If your export doesn't separate them, the skill says so instead of silently inflating the loss.

📉

Ideal cycle time set to "average"

Using average run rate instead of the best-demonstrated rate flatters Performance and hides speed loss. The skill always asks where that number came from before trusting it.

🔁

Changeover hidden inside run time

If changeover isn't logged as downtime, it shows as Performance loss instead of Availability loss, which misdirects the countermeasure to the wrong category entirely.

📊

No Six Big Losses breakdown

"OEE is 61%" gives no team anything to act on. The breakdown, with a Pareto, shows where those 39 lost points actually went and which categories to attack first.

🕳️

Messy exports silently dropped

Most downtime exports have bad header rows, blank columns, and inconsistent reason codes. The skill auto-detects delimiters, tolerates the mess, and lists exactly what it found vs. what's still missing.

📋

Unmapped reason codes buried

A large "Unmapped" bucket in the loss table is itself a finding. It means the reason code taxonomy at the source system needs work. The skill surfaces it instead of hiding it in rounding.

Six Big Losses

Every downtime reason maps to a loss category

The skill keyword-maps downtime reason text into the six categories automatically. Anything that doesn't match goes into Unmapped. It is shown, not hidden.

Availability Loss 1

Breakdowns

Keywords: failure, breakdown, fail, fault

Availability Loss 2

Setup & Adjustments

Keywords: setup, changeover, adjust, c/o

Performance Loss 3

Idling & Minor Stops

Keywords: jam, minor stop, idle, blocked, starved

Performance Loss 4

Reduced Speed

Keywords: slow, reduced speed, speed loss

Quality Loss 5

Process Defects

Keywords: scrap, defect, rework

Quality Loss 6

Reduced Yield

Keywords: startup, warm-up, yield

How It Works

Draft-first. Formulas shown. Gaps flagged.

The skill runs the calculation immediately from whatever you paste or attach, then flags anything that needs your judgment before you use the number.

Paste or attach your downtime export

Any CSV from your MES, SCADA historian, or a manual downtime log. Auto-detects comma vs. tab delimiter, tolerates a messy header row and blank columns.

The calculator runs

Availability, Performance, Quality, and OEE are computed using oee.py. If columns are missing, you get a clear list of what was found vs. what's still needed. Not a crash, not a guess.

Report with formulas and Pareto

Every formula shows the actual numbers substituted, not just the result. Six Big Losses table, downtime Pareto, and per-day/shift/line breakdown if those columns are present.

Traps are flagged, not hidden

If there's no planned-vs-unplanned separation, if Performance goes over 100% (usually an ideal cycle time that is too slow, or counts logged outside run time), or if a loss can't be mapped, it's called out explicitly.

Export to Word

One command converts the report to a clean .docx ready to circulate: python3 .claude/skills/oee-from-csv/scripts/export_docx.py report.md

Examples

Manufacturing and healthcare both supported

🏭 Stamping Line

🏥 MRI Utilization

A stamping line downtime log for one 480-minute shift. The period totals sit on the first row, and each row after that is a downtime event. The script auto-detects the delimiter and tolerates a messy header row.

planned_time_min,downtime_min,reason_code,ideal_cycle_time_min,total_count,good_count
480,42,Changeover,0.06,6800,6724
,18,Mechanical Failure,,,
,,Scheduled Break,,,

What the skill flags

Result from oee.py: Availability 87.5%, Performance 97.1%, Quality 98.9%, OEE 84.0%. The export has no planned/unplanned column, so all downtime is treated as unplanned, and the report says so. The Scheduled Break row has no minutes; if it were logged, it would be planned downtime and should not count as an Availability loss. The skill also asks whether 3.6 s per unit is the best-demonstrated rate and not an average, because an average flatters Performance.

MRI scanner scheduled-hours and scan count export. OEE terms map imperfectly onto healthcare, so the skill uses your vocabulary (scheduled hours, scans, turnaround) and includes an honest caveat.

scheduled_hours,downtime_hours,reason,scans_scheduled,scans_completed
8,1.2,Unplanned Maintenance,18,16
8,0.5,Patient No-Show,18,17
8,0,Routine Cleaning,18,18

What the skill flags

"Patient No-Show" is an external factor, not an equipment loss. The skill notes this may distort the OEE interpretation and recommends a separate no-show metric alongside OEE. The script also needs an ideal cycle time and a total count, which this export lacks, so the skill asks for them (for example, the ideal minutes per scan) before it calculates.

What you get

A real run on the sample data

Unedited output. The sample message at the top went to Claude Sonnet 4.6 (an earlier-generation model, so a newer one may word things differently) with this skill installed in Claude Code, which ran its calculator script, on September 29, 2026. Scroll inside the frame to read the whole reply.

Open the full image

Installation

Get it running in two minutes

Why Claude Code: this skill does its arithmetic with a small Python script so no number is ever guessed, and a claude.ai chat can't run the script. Install Claude Code from code.claude.com/docs. It needs Python 3 on your computer.

Download and unzip oee-from-csv.zip.

Put it where Claude Code looks: in the folder you'll work in, run mkdir -p .claude/skills and then move the unzipped oee-from-csv folder into .claude/skills/.

Run it: open a terminal in that folder, run claude, and say "Run the oee-from-csv skill on my_export.csv." Try the sample file first, line1_downtime.csv, which is in resources/examples/.

Requires Claude Code

This skill works with Claude Code (Anthropic's CLI). It is not a browser extension or SaaS tool. It's a folder you add to your project. Get Claude Code →

Related Skills

When OEE leads somewhere else

A high Six Big Losses bucket in Breakdowns suggests root-cause work. A large gap between scheduled and run time warrants a VSM to see the full picture.