The "Analysis instructions" field in AI processing options lets you do more than tell the AI what the job is. You can also write the kind of SOP you want. In the "Advanced use of analysis instructions" articles, for each situation where you want the result to differ from the default analysis, we show the text of "Analysis instructions" and the SOP actually created from that text. For the basic way to tell the AI what the job is, see Improving AI analysis accuracy (subtitles, step splits, body text).
In this article, from a roleplay training video shot in one continuous take that includes inappropriate responses (NG), on-the-spot corrections, and retakes, we create training material rather than a work SOP. The material teaches by comparing NG with the correct way. The example is a roleplay training for customer service in a store. You can view the example SOP without logging in (Store customer service roleplay training).
"Analysis instructions" is available on all plans. The feature to register the text as a template is available on the Pro and Enterprise plans.
With the default analysis, NG moments and retakes are cut
If you analyze without writing anything, the AI reads the video on the premise of creating an SOP that lets the viewer do the work correctly. So it cuts sections it judges unnecessary for doing the work correctly. For example, sections like these:
- Sections where no one is working (waiting time, small talk, greetings, etc.)
- Corrected speech and retakes (the wrong first attempt is cut, and the corrected retake is kept)
This works fine for a work SOP, but it is the opposite for roleplay training videos. NG responses, corrections, and retakes are exactly what you want to teach, and the greetings and the time spent listening to the customer are also part of the material.
When we actually analyzed the example video without writing anything, the correction "That's wrong" was read as a cue for a speech retake, all NG scenes were cut, and the SOP showed only the correct way. Also, general cautions not said in the video (such as the bowing angle) were added to the Tips and Caution fields, and all the questions became ones answered with "Yes / No." By writing in "Analysis instructions" what kind of material you want, you can change these decisions.
What you can change with "Analysis instructions"
Instructions written in "Analysis instructions" are reflected not only in the subtitles and the step text, but also in the following places. For the example, we gave the following instructions for each.
| What you can change | Instruction in the example |
|---|---|
| How steps are split | Make one correction into one step |
| What is cut and what is kept | Do not cut NG responses, corrections, or retakes |
| How work content, tips, and cautions are used | Write the correct way in the work content, and write the NG responses in the Caution field as "Common mistake example" |
| How subtitles are written | Even in NG scenes, do not write subtitles that instruct the wrong way |
| How questions are created | Show a scene and have the learner choose the correct response, with an NG response that actually appeared in the video as one of the wrong choices |
| On-screen annotations | Add an annotation starting with "NG" at the start of an NG response, and one starting with "OK" at the start of a retake |
Because the AI judges and creates the SOP based on your instructions, not everything will necessarily follow them exactly. Always check the finished SOP, and fix it in the video editor if needed.
Settings for the example
AI processing options
Apart from "Analysis instructions," we set the options as follows. For the meaning of each item, see How to choose each AI processing option. The AI processing options open when you upload, and the same screen also opens from "Check AI analysis results" → "Re-run AI" on an analyzed material (its title is "AI analysis settings").
| Item | Setting | Reason |
|---|---|---|
| Split granularity | Step | Because the granularity is close to one group per correction |
| Text style | Polite | To keep the text in polite form, since it is read as training material |
| Subtitle style | Match text | To keep the subtitles in polite form too |
| Videos without narration (analyzed only by video) | Leave unchecked | To use the spoken corrections as clues |
| Question type (advanced settings) | Comprehension test | To mix in wrong choices and check whether the content was remembered |
| Target step count (advanced settings) | Auto | Because "Analysis instructions" says to make the number of steps the same as the number of corrections |
| Aggressively cut sections without speech (advanced settings) | OFF | Because scenes without speech, such as greetings and the time spent listening to the customer, are also part of what to teach |
If you know the number of corrections in advance, you may also specify the number of steps with "Target step count."
Analysis instructions text
This is the text used in the example. It does not include content specific to that video, such as correction items or timestamps, so it can be used as-is for videos shot the same way. Including line breaks, the Japanese original is 991 characters and fits within the template body limit (1,000 characters). The translated text is longer, so shorten it to 1,000 characters or less before you paste it into a template.
[About this video]
This is footage of roleplay training for customer service and similar work. When an inappropriate response occurs, the trainer (a supervisor or the customer role) calls out a word such as "Stop" or "That's wrong" and shows the correct way. The staff member then redoes the same scene before moving on to the next one.[Desired SOP format]
Make one correction one step. Make the number of steps the same as the number of corrections, and do not decide it from the length of the video.
One step runs from the point where an inappropriate response begins to the point where the correction and the retake end.
Even if the correction words are not in the transcript, if the same scene is redone, treat it as a correction, and write the correct way based on how it differs from the first attempt.
Do not make a scene without a correction into its own step; include it at the start of the next correction's step. Include anything after the last correction in the last step.
If a retake also serves as the inappropriate response for the next correction, start the next step from the statement or action that is the target of that next correction.
Write the title and body so that the correct way shown in the correction can be reproduced exactly as it is.
Write inappropriate responses in the Caution field as "Common mistake example," and do not write them in the body.
Do not add general advice that is not covered in the video to any field or to the subtitles.[Subtitles]
Write the correct way that can be practiced as it is in the subtitles.
In scenes with an inappropriate response, do not write the subtitle as if instructing that response; write it as a "Do not ..." caution.
In scenes with a correction, do not describe the training progress, such as "receives a correction" or "the trainer tells"; write the correct way that was pointed out.[What not to cut]
"Stop," "That's wrong," and similar words are corrections, not slips of the tongue or re-recorded takes. Inappropriate responses, corrections, and retakes are the content to teach, so do not cut them. Scenes without hand movement, such as greetings, eye contact, time spent listening, and movement, are also content to teach.[How questions are created]
Make the questions multiple choice. Show a scene and have the learner choose the correct response. The correct answer is the correct way shown in the correction, and one of the wrong choices is an inappropriate response that actually appeared in the video. Do not make questions about whether a correction occurred or who gave it.[On-screen annotations]
Add a short annotation starting with "NG" at the moment an inappropriate response begins, and one starting with "OK" at the moment a retake begins.
What each paragraph does
| Paragraph | Role |
|---|---|
| [About this video] | Tells what the video is and the cue for a correction ("Stop," "That's wrong"). Without this, the correction words are read as a cue for a speech retake |
| [Desired SOP format] | Sets how steps are split, the scope of one step, and how work content and cautions are used. The sentence "Even if the correction words are not in the transcript" is a safeguard in case a correction was said too quietly to remain in the transcript. The last sentence prevents general advice not covered in the video from being added |
| [Subtitles] | Subtitles are read while watching the video, so this paragraph sets that subtitles in NG scenes must not instruct the wrong way. Without it, subtitles in NG scenes sometimes became sentences that directly instructed the inappropriate response, such as "Please greet and bow at the same time" |
| [What not to cut] | Specifies that, for this video, sections that are cut by default (corrected speech, retakes, and scenes without hand movement) are kept |
| [How questions are created] | Makes questions that ask learners to recall the scenes they saw in the training. Questions about the progress of the training, such as whether a correction was given, are not created |
| [On-screen annotations] | Shows the switch between NG and retakes on the video |
When using this for a different type of training video, it is easier to build the instructions by keeping this paragraph structure as is and rewriting only the content.
Analysis results of the example
This SOP is the example that was analyzed with this text and these settings, then published after small manual edits. The video, about 6 minutes from entering the store to seeing the customer off, became the following 8 steps.
- Separating greeting and bowing
- Supporting the customer to sit down
- Posture and eye contact when facing the customer
- Managing personal information when leaving the seat
- Listening to the customer
- Pointing with the palm
- Checking for questions after an explanation
- Seeing the customer off to the exit
Every step has an annotation starting with "NG: ..." at the start of an NG response and one starting with "OK: ..." at the start of a retake, and the Caution field contains a "Common mistake example." For example, the first step is as follows.
| Field | Content |
|---|---|
| Title | Separating greeting and bowing |
| Work content | - When welcoming a customer, first greet them with "Welcome." - After finishing the greeting, bow politely. |
| Tips | Be mindful not to greet and bow at the same time |
| Caution | Common mistake example: bowing at the same time as the greeting |
| On-screen annotation | NG: Greeting and bowing at the same time OK: Bowing after the greeting |
Five of the 8 steps became questions that show a scene and ask the learner to choose the correct response, and the wrong choices included responses that actually appeared in the video. For example, the question for the third step is as follows.
When talking face to face with a customer, which posture and gaze are correct?
- Face the customer with your body and talk while looking at their eyes (correct answer)
- Talk while staring at the tablet screen
- Talk while looking only at the materials in your hands
The remaining steps were 2 questions answered with "Yes / No" and 1 step with no question. Because some places differ from the instructions like this, check the questions and fix them as needed (Step-level question settings).
In the example, so that the voices giving corrections are heard as they are, subtitle read-aloud is not used and the original audio of the video is kept (Changing read-aloud and Mute Video Audio settings).
Things to keep in mind when filming
With this use, how you film greatly affects the results. Please follow these four points when filming.
- Say corrections out loud. The step text is created from the content of the corrections, so speak at a volume that can be heard clearly
- After a correction, carry out the correct way all the way through. This becomes the work content
- Point out only one thing per correction. If you combine two, one step will contain two items
- Film from the NG response through to the retake without stopping partway
The person giving the correction can be a supervisor or the customer role. The video can be analyzed whether you stop once and redo it, or fix it within the flow. Since it is easy to fit the staff member, the customer role, and your hands in the same frame, we recommend filming in landscape orientation.
Registering the text as a template
If you analyze videos shot the same way repeatedly, register the text as a template, and you can insert it from the icon at the top right of the "Analysis instructions" field.
- Open "Dictionary / Templates" in the sidebar, and on the "Templates" tab, select "Analysis instructions"
- Press "Create template," enter a name and body, and save. For the name, use something that shows which video it is for when you choose it in "Analysis instructions" (example: "Roleplay training: keep NG and retakes"). Paste the text above into the body. The body can be up to 1,000 characters
- In the AI processing options that open when you upload, select the registered template from the icon at the top right of "Analysis instructions"
Templates are available on the Pro and Enterprise plans. Registering and editing can be done by Team Administrator or higher roles.
Frequently asked questions
Q. NG scenes are being cut.
A. In "Analysis instructions," write the correction cue words (such as "Stop" and "That's wrong") and that NG responses, corrections, and retakes must not be cut, then re-analyze. Cut sections can also be restored with "Undo" in the video editor.
Q. Corrections were given, but the steps are not split.
A. If a correction is said quietly, it may not remain in the transcript and may not be treated as a correction. The sentence in the example text, "Even if the correction words are not in the transcript, if the same scene is redone, treat it as a correction," is a safeguard for this. When filming, speak at a volume at which corrections can be heard. Steps can also be split in the video editor (Adjusting step splits).
Q. I want to change the text and compare the results.
A. Copy the material, change "Analysis instructions," and re-analyze, and you can compare the results while keeping the original result (Trying different analysis settings (copy the material and re-analyze)).
Q. I have registered content from a different industry in "About this site" in the Terminology Glossary.
A. The content of "About this site" and the Terminology Glossary is passed to the AI as a premise for every analysis. When you analyze a type of video that differs from the registered industry, clearly state in "Analysis instructions" what the video is, as in [About this video] in the example (Terminology Glossary (terms and multilingual translations)).