# Perplexity Brain now writes and improves its own skills

By Ivo Donker · November 20, 2026 — text: Claude (AI)
What does Perplexity Brain do with skills? Brain, the system that lets Perplexity Computer learn from past work, now creates personal skills on its own: fixed instructions for tasks a user runs again and again. It draws them from earlier sessions, has them checked and keeps them up to date. Perplexity says Brain wrote more than 80,000 skills for over 20,000 users between August 1 and November 14. The feature is on for all Max subscribers in the United States.

We read [Perplexity's technical blog post "Self-improving skills"](https://www.perplexity.ai/hub/blog/self-improving-skills) of November 18. The story spread through [a post by @perplexity_ai on X](https://x.com/perplexity_ai/status/2123449395882701224) and was in the [trending briefing of November 20](https://trending.llmnet.nl/en/briefing/2026-11-20).

## What happened

Skills are reusable instructions, sometimes with scripts, that an agent loads only when a task calls for them. Brain already kept a knowledge base about the user. It now also writes personal skills from that user's memory and session history. Perplexity uses the same approach to improve the built-in skills that every user gets, based on patterns across many sessions.

Improvement runs in four steps: find evidence in past sessions where a skill fell short, have an improvement agent propose a change, check that proposal, then activate it. Perplexity says this changes instructions and context, not the model's weights.

## The rules and the numbers

Personal skills get a daily run. The improvement agent has no internet access and no connectors. A new skill needs two separate occurrences of the same reusable behaviour; an update to an existing skill can rest on one well-grounded example. Each skill gets at most one change per run. Automated checks confirm a proposal is well formed, and a separate language model audits a sample. Each user has at most ten automatically created skills; to add one, another has to go.

Built-in skills face a higher bar. Perplexity only uses sanitized sessions that are eligible for training; credentials, secrets, attachments and media are stripped. A revision must pass every test scenario, while the old version must fail at least one. An engineer then reviews the change as a pull request.

Measurement (according to Perplexity) | Result | 

Personal skills written, Aug 1 to Nov 14 | over 80,000, for over 20,000 users | 

Automatic updates, Sep 4 to Nov 14 | over 100,000, on over 30,000 skills | 

Eight-week test: got a new skill | 72.9 percent (added one manually in the control group: 6.5 percent) | 

Personal skill loaded in a user's own conversation | 53 percent versus 13 percent | 

Failed tool calls | 11 percent fewer | 

Dissatisfied threads per user | 7.6 percent fewer | 

Cost and response time | no measurable change | 

The largest group of skills covers software engineering (11 percent), followed by research and briefings (10.1 percent). In a smaller test on SkillLearnBench, task success rose from 43.8 to 62.5 percent across 16 paired cases.

## What this means

Agents recording their own procedures is not a new idea: [Hermes Agent](https://nieuws.llmnet.nl/en/hermes-agent-0-23-2-stable) has done it for a while, and Anthropic let agents build memory with [dreaming](https://nieuws.llmnet.nl/en/anthropic-agentgeheugen-dreaming). Perplexity shows how to wrap it in checks: a fixed threshold for new skills, a cap of ten and human review for anything that reaches everyone. If you maintain your own skills, those rules are worth copying.

Read the numbers with care. They come from Perplexity itself, and the headline comparison is lopsided: 72.9 percent received a skill automatically, against 6.5 percent who made one themselves. That mostly shows automatic is easier than manual. The 11 percent fewer failed tool calls and 7.6 percent less dissatisfaction say more. We earlier covered [Perplexity Teammates](https://nieuws.llmnet.nl/en/perplexity-teammates-agents) and the [Orchestrator API](https://nieuws.llmnet.nl/en/perplexity-orchestrator-api).

## What we did not check

We did not use the feature ourselves; it is only for Max subscribers in the United States. All figures come from the blog post and were not measured independently. That Teammates and Multiagent are now in the mobile apps comes from [a post by @perplexity_ai on X](https://x.com/perplexity_ai/status/2123486281644560724); we did not verify it.

Checked on November 20, 2026.

This article was written with AI (Claude) and checked against the sources below.

## Sources

- Perplexity, Self-improving skills (November 18, 2026): [perplexity.ai](https://www.perplexity.ai/hub/blog/self-improving-skills)

- @perplexity_ai on X about Brain and skills: [x.com](https://x.com/perplexity_ai/status/2123449395882701224)

- @AravSrinivas on X about the percentages: [x.com](https://x.com/AravSrinivas/status/2123451045657690627)

- @perplexity_ai on X about Teammates on mobile: [x.com](https://x.com/perplexity_ai/status/2123486281644560724)
