Microsoft Skill Recorder: record a workflow, then verify the skill

Start at the official repository and its release instructions. The source-release installers use a pinned revision and build a local app. The documented workflow requires a GitHub account with Copilot access. Check the installation guide for your operating system instead of copying an old installer command from a post.

Create a separate demo folder containing a tiny CSV with invented names and values. Decide the goal before recording: “Turn these rows into a Markdown table and save it in this folder.” Close unrelated windows and notifications. Keep the demonstration short enough that you can explain every step afterward.

Begin a recording, open the fixture, perform the conversion and save the table. Narrate the reason for a decision if it would otherwise be unclear—for example, that the first CSV row is a header. Avoid demonstrating passwords, private messages or customer records.

Stop when the saved result is visible. Remove unnecessary detours from your explanation. The useful material is the intent, input, transformation and output; incidental window positions and account names should not become requirements.

Before analysis, inspect the recording and the material selected for it. Skill Recorder can capture activity, frames, clipboard previews and optional narration or terminal excerpts; the README describes those inputs.

Recording locally does not mean analysis stays local. The project documents sending timeline information, frames, narration text and selected terminal excerpts to GitHub's cloud for analysis. Review the data-flow explanation before using Analyze. If the recording contains material you cannot send there, stop and make a clean recording.

After analysis, compare the inferred intent and steps with what you actually meant. Correct accidental dependencies such as your username, a fixed download path or a one-off filename. An accurate-looking summary is still something to review.

Choose the skill output for your first trial. Review the SKILL.md preview and any bundled commands before adding or exporting it. The project also supports automation-oriented outputs, but a manually invoked skill gives you a smaller first acceptance case. See the documented creation flow.

Use this review checklist:

• The description says when to use the skill and what input it expects.
• File paths are parameters or relative to the chosen demo folder.
• The output format and overwrite behavior are explicit.
• Every script or command is understandable and necessary.
• There are no credentials, private excerpts or unexplained network calls.

For structure and portability, compare with the SKILL.md guide. Generating instructions does not train a new model or guarantee an identical visual replay.

Create a second CSV with different names, row count and filename. Invoke the reviewed skill in a supported agent with access limited to the demo folder. Ask it to produce a new Markdown file without changing either input.

Check the actual output: headers, row count, values, escaping and destination. Confirm the original inputs are unchanged. If the skill only works with the first filename or reproduces the first example's values, edit the instruction and repeat that acceptance case.

Record the failure concretely—“three of four rows preserved” is useful; “the skill seemed smart” is not. Do not schedule repeated execution until the manual result and required permissions are understood.

The repository is MIT-licensed; that does not make cloud analysis or Copilot usage unlimited. Check your GitHub Copilot entitlement and applicable usage limits before using the analysis flow. The software license, model access and your organization's data rules are separate conditions. Project requirements and license

Keep the recording and generated skill only as long as you need them under your own data policy. This page documents a proposed first workflow; it does not claim successful installation, analysis or cross-agent execution.