FROM THE AI FRONTIER
Frontline news — November 13: check whether an AI skill helps
A recent AI Engineer talk proposes comparing results with and without a skill. Here is a beginner’s way to check a saved instruction before adding another page of rules.
If you give an assistant a long instruction, check what it improves. Start by giving it the same task with and without that instruction. Decide in advance which mistakes you will look for.
Issue dated November 13, 2026. Today’s selection is one video from the past two weeks about evaluating skills, rather than a new product launch. I chose an idea you can try without programming.
Here, a skill means a saved set of instructions for a recurring task. The talk also covers technical agent skills; for your first exercise, you can compare ordinary text instructions.
Video: what does a skill add?
Original: [Your SKILL.md Is 70% Stuff the Model Already Knows — Nikolay Petrov, Intel](https://www.youtube.com/watch?v=yzx4IuEh_tU).
- Channel: AI Engineer.
- Publication date: November 11, 2026, according to matching Innohub and QA Bash listings. The upload time and time zone have not been confirmed at the primary source.
- Language: English.
- Duration: 14:01 in the transcript listing.
- Content read: captions on AINotes, from 00:00 to the closing remarks at 13:54. The visuals were not watched.
Nikolay Petrov proposes comparing model answers with and without a skill. He also examines whether instructions help with exact details and what happens after shortening them.
The practical value is a way to check instructions on your own tasks. The limitation is that the speaker’s observations do not establish that every skill improves results.
Where to look in the video
- 01:43 — comparing results with and without a skill.
- 09:43 — exact parameters and steps.
- 12:06 — checking again after changes.
These timestamps come from the transcript we read. Automatic captions may contain errors.
Do not remove rules based on the headline percentage
The video title is no reason to delete most of your instruction. Your work may have specific requirements for amounts, formats or conditions. I would first list where the assistant needs to follow your particular rules.
The following is an editorial adaptation for an ordinary chat. It is a proposed exercise, not a test we performed.
Example: replying to an event enquiry
Consider a fictional task: draft a reply to a guest using a short event brief. The brief includes a date, a price and a confirmation condition.
Prepare two versions:
- A: the original brief and a request to draft a reply.
- B: the same task plus your saved instruction.
Before running them, write down the criteria:
- The date and price are preserved.
- The confirmation condition remains in the reply.
- No services or promises are invented.
- The tone suits the recipient.
Use separate new chats, the same model and identical source data. Record enabled settings and any additional materials. Check that the instruction is actually available in the second version.
Compare each answer against every criterion. “I prefer B” is a useful impression, but it does not explain which rule helped.
What to do with the results
If the instruction helps preserve an important condition, check it on several new examples. If you see no difference, do not delete everything immediately: the task may be too easy, or it may never have needed the relevant rule.
Before shortening an instruction, keep the previous version. Remove one passage and compare again. Include a case where that passage was intended to help.
One pair of answers starts an evaluation; it does not establish a consistent advantage. Repeat the comparison after changing models, settings or task types. Check another assistant’s assessment against the original brief and your criteria.
Sources
- AI Engineer — Your SKILL.md Is 70% Stuff the Model Already Knows — Nikolay Petrov, Intel
- AINotes — субтитры доклада Nikolay Petrov
- QA Bash — карточка исходного видео
- Innohub — uploads AI Engineer