Your best explanation of a task may be in a recording. A new colleague needs the steps, the exceptions, and somewhere to check an unfamiliar term.

OGAD (Off Grid AI Desktop) can transcribe saved training audio locally and help you turn the text into an onboarding guide. You review the instructions, add missing visual details, and keep the finished guide in a searchable project. The saved-file workflow uses the free core app.

Download OGAD

Desktop releases

What would you like to do with Off Grid AI?

Have a feature or use case you would like us to support? Tell us what you want to do and which device you use.

Write to support@offgridmobileai.co, join our Slack community, or talk to us on Reddit.

This is useful for a small training company, agency, or operations team with repeated explanations scattered across recordings. Start with a process you already teach. You do not need to create a complete training library before seeing whether the method helps.

What makes a recording useful for onboarding?

A useful recording explains what the learner must achieve, what they need before starting, and how to know the task is complete. It also names the cases that require a different action. AI can help organise those details when they are present in the source.

Suppose your operations lead records a walkthrough of preparing a client workshop. The session covers confirming attendance, creating the participant pack, checking access, and sending the final instructions.

During the explanation, they mention two exceptions: an external trainer needs different access, and a participant who joins late needs a separate message. Those details belong in the guide because they change what the new colleague does.

A useful result has five parts:

Guide section	Question it answers
Purpose	What should I finish?
Before you start	What information and access do I need?
Steps	What do I do, in what order?
Exceptions	When should I stop or use a different route?
Completion check	How do I know I did it correctly?

Keep the original recording available. It is the source you use when the draft leaves something unclear.

What do you need on your computer?

Use OGAD with a downloaded speech model and a local text model. Complete the app, runtime, and model downloads before working without internet. If you want to query the finished documents later, complete the local project indexing setup too.

The model-setting steps below use an Apple Silicon Mac. The core saved-audio and document workflow is also available on Windows, where speech runtime setup can differ. Release 0.0.54-beta.108 also provides Linux packages as a beta. Saved-file transcription and document projects do not require Pro meeting recording.

Prepare a short training audio file, preferably covering one process. MP3, WAV, M4A, and FLAC are among the supported audio formats. Clear speech is more useful than a long recording with several unrelated tasks.

If you have a screen recording, export its audio first. A speech transcript cannot explain an unspoken click or read a silent slide. Keep the video available for your visual checks and add the missing details yourself.

How do you create the first draft?

Transcribe the audio, check it, then ask for a guide with a specific structure. Separate speech recognition from writing so that you can correct names, menu labels, and numbers before they become instructions.

1. In Models, choose local transcription and text models that fit your computer.
2. In model settings, open Transcription and set Current model and Spoken language. Use a multilingual model when needed.
3. Open chat, select + > Attach files, and choose the audio file.
4. Wait for processing. Open the attachment’s text preview and compare key passages with the recording.
5. Copy the transcript into a local text editor, make necessary corrections, and save a checked version. Attach the checked file to the drafting request.

The desktop file-processing code routes audio through transcription and handles video through a separate visual path. Attaching the video alone is not a substitute for extracting its spoken instructions.

Use a request such as:

Turn the attached checked transcript into an onboarding guide for someone performing this task for the first time. Include purpose, prerequisites, numbered steps, exceptions, and a completion checklist. Use only the transcript. Mark missing information as “Trainer to confirm.” Preserve exact amounts and menu labels. Do not invent a step because it seems likely.

Read the draft as if you were the new colleague. When it says “open the usual folder,” it needs a real folder name or a note for the trainer to supply one.

How do you turn a summary into usable instructions?

Check each step for an action, an object, and an expected result. “Prepare the workshop” is a summary. “Create a participant list using the confirmed registrations” gives the learner something to do.

For the workshop example, ask the model to improve only the wording supported by the transcript:

Review this draft for vague instructions. List steps that lack a named document, location, owner, or completion condition. Suggest a clearer sentence only where the transcript supplies the missing detail. Otherwise leave a question for the trainer.

Then watch the parts of the video where the trainer points or clicks without explaining. Add screenshots or written details to your own guide. Label these as reviewed additions so they are not confused with the transcript.

Have someone familiar with the process check the result. For a first trial, ask a colleague to complete one task using the guide and record where they stop to ask a question. Those gaps tell you what to clarify next.

How do you make the guide easy to query later?

Save the approved guide as a text or Markdown file. Open Projects > New project, enter a name, and press Enter. Open Knowledge & settings > Knowledge base > Add files. Wait for indexing and leave the file enabled for retrieval. Open Chats > New chat inside the project to ask your question.

These controls are part of the project knowledge base. Give files clear names and versions, such as workshop-preparation-approved-2026-09.md.

Try a real new-starter question:

Search the onboarding documents. What changes when an external trainer leads the workshop? Cite the document that explains it. If the guide does not answer, say what is missing.

Check the source beside the response. An answer is useful when it points to the approved instruction you intended the learner to use.

Adding a document on your computer does not automatically create a shared training portal for colleagues. You can share your reviewed guide through your existing process. Shared access and distribution are separate decisions.

How should you handle long recordings and updates?

Split a long session at task boundaries. Ask for one guide per process, then use a short contents page to connect them. This keeps the task clear and reduces the chance that a limited model context loses an exception near the end.

When a process changes, update the approved guide and its date. Remove or disable the outdated document in the project before relying on new answers. Keep the raw recording separate from the approved instructions so an old explanation is not treated as current policy.

Problem	What to do
The transcript has wrong menu names	Check the recording and correct the text before drafting
The guide invents a missing step	Remove it and ask the trainer for the actual instruction
The answer uses an old process	Check which project files are enabled
The model misses an exception	Ask about that exception directly and inspect its source
The source is mostly silent demonstration	Add reviewed visual instructions; speech alone is insufficient

Build one guide your next colleague can use

Download OGAD and choose one short training recording. Produce a checked transcript, a clear task guide, and three questions a new colleague should be able to answer from it.

Keep speech and text processing local for this workflow. Selecting a remote model changes where that stage runs. Downloads and later sharing also require their own connection; the article does not depend on uploading recordings to a cloud transcription provider.