Latest in AI
Perplexity Brain Skills: Review What Computer Learns Before Reusing It
Check access to self-improving skills, inspect Brain-authored procedures, disable unwanted skills and separate skill controls from memory settings.
An assistant remembering your preferred report format is useful. An assistant repeatedly applying a rule you only wanted once is not. Self-improving skills make that distinction worth checking before you trust the next recurring task.
Perplexity's November 18 explanation of Brain describes automatic skill creation and revision. The practical question is what Computer has learned, when it should apply that procedure, and how you stop using it when the procedure is wrong.
I would start with one recurring, reviewable deliverable: a weekly client update. Keep sending outside the experiment. The goal is to reduce editing, not to automate a mistaken promise to a customer.
Check access without confusing it with Brain's original launch
The self-improving skills announcement says personal skill improvement is available to US Max subscribers and enabled by default for eligible users. This is not Brain's first appearance: the article reports automatic skill creation dating back to August. Treat November 18 as the public explanation and current availability statement, not the first day the system learned anything.
Personalization changes instructions and context, not model weights. Personal skills use your history; built-in improvements draw on patterns across users and pass replay evaluation plus internal human review. Those are different processes.
For readers outside the US, having Brain or a Max subscription does not establish access to this particular automatic-skill feature. Verify what your account actually exposes before changing a reporting process around it.
Find the skill and check when it should run
Perplexity's current skill-management guide places installed skills under Plugins, then Installed, with the Skills view selected. Open a skill to inspect its name and description. The description helps Computer decide when to load it. The announcement says Brain-authored skills are identified in Plugins.
For a weekly client update, I would check four things: the intended client, the reporting period, the required evidence and the audience. “Write a weekly update” is too broad if one client expects an internal operational report and another expects a short external email.
Give a correction that states the boundary: “Use this structure for Acme's weekly report, not for all client messages.” That is much more useful than repeatedly saying the output feels wrong.
Keep a copy of the intended procedure somewhere you control. It becomes your reference when judging whether a later result improved or simply changed.
Disable a bad skill without treating it as deletion
The management guide documents an Enabled switch on Installed that disables a skill without uninstalling it. Start there when you want to stop using an unwanted procedure while investigating what went wrong.
Do not combine three separate intentions into “turn it off”: stopping a skill from loading, stopping automatic improvement, and clearing saved memory are different requests. Automatic improvement has a separate control. Do not assume the Enabled switch also prevents future skill creation; check the current account controls for that separate choice.
For the client-report example, disable the suspect skill, state the required structure explicitly in a new task and compare the result. Check whether a different skill or instruction is producing the same unwanted behavior before blaming memory alone.
Review memory separately from the procedure
The Brain help page explains that memory entries link to their source and can be edited or deleted. It documents Settings → Memory settings → Brain as the control that stops new knowledge being written. It also says incognito sessions are excluded and proactive features such as improving skills are controlled separately.
That distinction matters when the error is a fact rather than a workflow. A report could follow the right structure while using an obsolete launch date. Correct the underlying fact and its scope instead of rewriting the whole reporting procedure.
Turning off new memory writing should not be described as deleting existing memories, uninstalling skills or reversing work already completed. Pick the control that matches the outcome you want, then inspect the next result.
Test reuse against a small acceptance checklist
Before relying on the learned procedure, give it a fresh reporting period and check the result against the source material. My checklist would include accurate dates, unchanged numbers, clear separation of completed and planned work, and no invented commitment to the client.
Add an exception case too. If the client has no update this week, does Computer say so, or manufacture a polished paragraph to fill the format? A reusable procedure should handle missing information without turning it into a claim.
Measure how much correction the report needs over several runs. More saved skills is not the success metric. Fewer repeated mistakes, while preserving the ability to handle exceptions, is the useful outcome.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI teams explain product behavior and practical workflows through clear technical writing and storytelling.