Records a Windows procedure click by click and writes it up as a step-by-step guide with annotated screenshots.
It hooks your mouse, screenshots each click, asks UI Automation what you clicked,
draws a highlight around it, and produces a single self-contained guide.html you
can email, drop on a file share, or paste into a SharePoint page.
Ships two ways to use it:
Scribeling.exe— a standalone recorder with a GUI. No Python needed.- A Claude Skill — installed into Claude Code on Windows, so you can say "document how to create an Autopilot profile" and have the recording and the write-up handled in one pass.
The skill route exists because of a division of labour worth being explicit
about: the recorder is good at capture and bad at writing. It sees one click at a
time and names controls by their accessibility label, which yields captions like
Click **Convert all targeted devices to Autopilot** — accurate, and useless as
documentation. A language model is the reverse: good at writing, blind to your
screen. So the recorder captures, and Claude rewrites the captions from the
screenshots, groups steps into stages, and adds prerequisites.
Download Scribeling.exe from the
latest release, or build it yourself:
git clone https://github.com/thediscovery-cmd/scribeling.git
cd scribeling
.\build.ps1Output is dist\Scribeling.exe. Double-click it, fill in a title and
description, press Start recording.
| Key | Button | Does |
|---|---|---|
| F7 | Pause | stop and resume capture |
| F8 | Finish | end the recording, go to review |
| F9 | Undo | drop the last step |
| F10 | Note | insert a note step |
A small always-on-top HUD lists steps as you make them. Clicks on the HUD are ignored, so its own buttons never end up in your guide.
Then the step editor opens. For each step you can:
- edit the Step caption, add a Detail line, and set a Stage
- Delete it, or tick several and Delete selected to clear out misclicks
- move it ↑ / ↓
- Merge up into the step above (captions join, the later screenshot wins)
- swap the screenshot with Image…, or drop it with No image
- click the screenshot to Redact (solid box), Blur (pixelate) or
Crop it. Edits are saved to a new file, so Reset to original can
undo them later; only the edited version goes into
guide.html - add a blank + Step below for a note or a step the recorder missed
Undo (or Ctrl+Z outside a text box) reverses any of these, deletes
included. Save writes steps.json; Export guide also builds the HTML.
To change a guide later, choose Edit an existing recording… on the start
screen, or run Scribeling.exe edit "<session folder>".
Output lands in Documents\scribeling\<timestamp>\:
<session>/
├── steps.json the source of truth - edit this
├── shots/ step-001.png, step-002.png, ...
└── guide.html generated output, self-contained
steps.json is also editable by hand. Change captions, add a detail
sentence, group steps with phase, add a prerequisites list, set
"hidden": true on anything you want gone, then:
Scribeling.exe rebuild "C:\Users\you\Documents\scribeling\2026-08-15_143002"Captions accept **bold** and `code`. Top-level author and tags fill
the byline and the product badges; a step with "action": "navigate" and a
url renders as a clickable link.
guide.html is one file with everything inline, so it works from an email
attachment or a file share with no network access. It has:
- A stage sidebar built from your
phasenames, with a "2 of 3" counter that follows you as you scroll, and up and down buttons to jump between stages. - Step cards with numbered badges you can link to directly
(
guide.html#step-7). - Click-to-zoom screenshots. Click again to see the image at full size, and press Esc to close.
- A highlight box and a pointing hand on each screenshot. The box shows which control was clicked, and the hand (with a soft glow) shows exactly where.
- Light and dark themes. The guide follows the reader's system setting, and the button in the corner overrides it.
- A byline with the author, step count, how long the procedure took, and the capture date. The author defaults to your Windows display name.
- Copy-on-click for anything in
`code`, such as a value to type or a group tag. - Print-friendly output: printing drops the sidebar and keeps steps whole.
See examples/example-guide.html.
Download scribeling-guides.skill from the latest release
and either click Save skill on the file card in Claude, or unzip it by hand:
Copy-Item .\scribeling-guides.skill "$env:TEMP\scribeling-guides.zip" -Force
Expand-Archive "$env:TEMP\scribeling-guides.zip" -DestinationPath "$env:USERPROFILE\.claude\skills" -ForceStart a fresh Claude Code session — skills load at startup — then describe what you want documented. Claude sets up its own virtual environment on first use, so there is nothing else to install.
To package it from source: powershell -File tools/pack-skill.ps1
Every step is named from the accessibility name of what you clicked. Control
types that tell a reader nothing (Group, Pane, Text, Custom, Document)
are never spoken aloud; only Edit, CheckBox, RadioButton, ComboBox,
TabItem, MenuItem, Slider and Spinner get named, because those nouns
carry information — "the Name field", "the Yes option".
The context line above a run of steps appears only when the window changes, so a guide that stays in one portal reads as one continuous list instead of repeating the same tab title thirty times. Browser names are stripped from window titles.
Repeat clicks on one target inside 1.6 seconds collapse into a single step, fingerprinted by element name, control type and bounding box. When UI Automation returns nothing the fingerprint falls back to the click position, so blind clicks at different places stay distinct.
A medium-integrity process cannot read the UI Automation tree of an elevated window. Record anything running as administrator from a non-elevated recorder and you get screenshots with no element names, and captions reading "Click the highlighted control".
build.ps1 therefore manifests the exe as requireAdministrator by default — a
high-integrity process can read both high and medium integrity windows, so it
covers everything. Use .\build.ps1 -NoAdmin if you would rather not, in which
case the setup screen warns you and offers to restart itself elevated. For the
skill, launch the Claude Code terminal as administrator.
The UAC consent dialog itself runs on an isolated desktop that nothing can screenshot or hook, at any integrity level. scribeling polls for it and writes an "Approve the User Account Control prompt" step with no image, so the guide has an instruction rather than an unexplained jump.
- Electron, Java, and legacy Win32 apps with no accessibility layer give empty element names. Screenshots and coordinate-based highlights still work.
- Fields flagged
IsPasswordare never transcribed, but a screenshot of the dialog is still taken. Check your screenshots before sharing a guide — they come from a live desktop and may hold a tenant name, a serial number, or an open inbox behind the window you meant to capture. - Screenshots are taken on mouse-down, so a UI that repaints instantly can be a frame behind.
- The exe is unsigned. SmartScreen will warn on first run, and a PyInstaller
onefile build that installs a global keyboard hook matches a fair number of
infostealer heuristics — if your EDR quarantines it, swap
--onefilefor--onedirinbuild.ps1, which trips far fewer detections.
scribeling/
├── build.ps1 builds Scribeling.exe
├── requirements.txt
├── examples/example-guide.html what the output looks like
├── skill/scribeling-guides/ the Claude Skill
│ ├── SKILL.md
│ ├── references/caption-style.md how captions get rewritten
│ └── scripts/
│ ├── scribeling.py the recorder (canonical copy)
│ ├── setup.ps1 venv bootstrap
│ ├── launch.ps1 detached launch with verification
│ └── inspect_session.py compact session summary
└── tools/pack-skill.ps1 builds the .skill archive
skill/scribeling-guides/scripts/scribeling.py is the only copy of the recorder;
build.ps1 and the skill both point at it.
An independent implementation, written from scratch against the public Windows
UI Automation and screen capture APIs. It is not derived from, affiliated with,
or endorsed by any commercial documentation product, and contains no third-party
code beyond the dependencies listed in requirements.txt.
MIT