Skills run again
Skills run at least twice in the 14 days after saving ÷ skills saved from an offer.
No baseline yet: only I have used it. In my own sessions, one prompt came back 3 times in a day.
Prototype · independent, not made by or with Perplexity
Session to Skill
Do a task 3 times, get asked once, and it becomes a skill you call with a slash command.
Problem
You only know a task is worth saving after you've repeated it. Then a skill starts from a blank page.
My testI ran one prompt 3 times, word for word, in a day. On Oct 2 to 4, 2026 the session menu had no save-as-skill option.
Solution
The skill comes out of the work: 3 runs, one offer, one command
Run 1 of 3, recorded as run 7
Make a one-page comparison of the current US monthly prices for Notion, Obsidian Sync, and Evernote personal plans, with source links.
Output: a one-page comparison
It reported Notion Plus at $10 a month.
Metrics
Saved skills get used again, and never hide a failed check
- A failed check shown without a flagnever
- Offers turned downset after baseline
- Offers accepted ÷ offers shownset after baseline
- Repeat prompts sent as a slash commandset after baseline
All proposals. The bars get set after a baseline week of real use.
Trade-offs
What I gave up, on purpose
Ask after three runs, never auto-save
Oversaving skills silently
Costfewer skills get made
Plain Markdown, not a hidden config
Overa polished settings screen
Costlooks rough to people new to AI
Keep the prompt, don't rewrite it
Overa model generalising the sessions
Costone skill per exact task, no inputs yet
Research
For anyone who wants to dig in
The needWhy a blank page is the wrong start
Skills are saved instructions an agent can reuse. Making one today starts from a blank page.
You have to know the skill before you need it.
People find out a task is worth saving after they have repeated it, not before.
Writing a skill is hard if AI is not your job.
A skill needs a prompt, inputs, rules and a way to check the result. Most people have never written one.
What I saw
I ran one prompt 3 times, word for word, in a day. On Oct 2 to 4, 2026 the session menu showed no save-as-skill option and the skill builder had no test step. I did not check one menu.
My read of the gap. I have not interviewed Perplexity users about Skills.
From 3 runs to one skillThe real sessions, and the skill they become
You did this 3 times
“Make a one-page comparison of the current US monthly prices for Notion, Obsidian Sync, and Evernote personal plans, with source links.”
- Run 7Oct 8, 2026, 2:37 AM14.29 credits
- Run 15Oct 8, 2026, 2:39 AM12.14 credits
- Run 16Oct 8, 2026, 6:35 PM21.96 credits
The answers disagreed once: Notion Plus came back at $10, $12, $12 a month. So the skill checks its own result.
The skill/app-price-comparison
App price comparison
- What it does
- Make a one-page comparison of the current US monthly prices for Notion, Obsidian Sync, and Evernote personal plans, with source links.
- What it checks before showing a result
- Covers every app asked for
- Links to each vendor's own site
- Uses only vendor sites as sources
- Fits on one page
- Says when prices were checked
- What it will cost
- 12.14 to 21.96 credits, from your 3 runs
The skill file itself, in Markdown
# Skill: App price comparison ## Prompt Make a one-page comparison of the current US monthly prices for Notion, Obsidian Sync, and Evernote personal plans, with source links. ## Inputs None. The sessions used the same words. ## Rules None yet. ## Test A run of this skill must pass every check before its result is shown as final: - Covers every app asked for - Links to each vendor's own site - Uses only vendor sites as sources - Fits on one page - Says when prices were checked ## Expected cost 12.14 to 21.96 credits, from 3 measured runs of this prompt.
What this prototype doesn't showLimits
- The skill is written by plain code from the 3 saved outputs, with no model in the loop.
- It does not hook into Perplexity's product. The offer, the build and the command are my proposal.
- Still open: where the offer appears in the product, a second example with real inputs, and corrections saved as rules.