Perplexity Computer Skills - Extend Computer’s capabilities with repeatable instructions

he SKILL.md format has quietly become a lingua franca for how builders encode their best thinking into AI agents.

Claude Code popularized it.

Codex adopted it.

And now Perplexity Computer supports it which means the workflows you've spent months refining don't have to live in one tool anymore.

That's the actual news here. Not just "Perplexity added a feature." It's that your institutional knowledge the how-to-handle-this, how-to-structure-that logic baked into your skill files is now portable.

The gap this fills:

Most power users of Claude Code or Codex have accumulated a small library of SKILL.md files.

Presentation builders.

Research frameworks.

Weekly briefing templates.

These aren't throwaway prompts they're curated playbooks that encode real judgment.

Until now, those lived in one tool's ecosystem.

What Skills actually does in Perplexity Computer:

• Upload any existing SKILL.md file directly -- it works as-is if it has the right YAML frontmatter

• Describe a workflow in plain language and Computer builds the skill for you (no technical knowledge needed)

• Skills activate automatically based on query context -- you don't have to invoke them manually

• Computer can combine multiple skills mid-task (Research + Content for a blog, for example)

• Browse and manage your library from a dedicated Skills tab

The execution layer underneath is what makes this different from just "importing a prompt."

Your skill now runs on 19 specialized models: Opus 4.6 for reasoning, Gemini for deep research, specialized models for images and video.

A SKILL.md that told one model what to do now tells the best model for each sub-task what to do.

Who this is for right now:

Builders and indie founders who've invested in building agent workflows on Claude Code or Codex.

If you've built more than five SKILL.md files, this is probably worth your attention today.

If you're newer to the format, Computer can generate skills from a plain-language description, so there's an on-ramp.

One honest caveat:

Computer is currently Max-subscriber-only at $200/month.

That's not nothing.

Pro and Enterprise rollout is coming, but if budget is a constraint, that's the real friction point to weigh.

The bigger question I keep turning over: as SKILL.md portability becomes table stakes, does the agent tool that wins end up being the one with the best execution layer or the one with the best skill library ecosystem? Perplexity is clearly betting on the former. Curious whether anyone here has already migrated workflows over and what the quality delta looks like in practice.

@rohanrecommends @aravindsrinivas @Perplexity Computer Skills As a heavy user of AI workflows, this "Computer Skills" module is really quite different.

Upgrade "prompt" to "reusable workflow template

Previously, long prompts were piled up in various chat interfaces, which were difficult to reuse and could not guarantee stable output. Perplexity's Skills transform these frequently used patterns into "naming and manageable skills", and they can be triggered automatically based on descriptions, providing a more experience similar to "requesting a fixed style AI employee" rather than randomly generating a prompt.

The two creation methods cover users ranging from beginners to advanced users.

If you can't write the SKILL.md file, you can directly generate it in the Computer using the "Create with Perplexity" dialog box. This is very suitable for product, operation, and analysis students.

For teams with their own knowledge base and standards, you can manage it using a .zip + SKILL.md format. Combined with the YAML frontmatter's name and description, it can achieve precise triggering and is already very close to the feeling of "lightweight workflow orchestration".

The automatic activation driven by the description + combined use is the highlight.

The series of relay tasks mentioned in the document, such as Research, Research Report, and Slides, are very crucial:

Users only provide requirements, and Computer selects and arranges the skills based on the descriptions.

In the same task, one can first conduct research, then write a report, and finally generate a presentation. This minimizes the need for manual copy-pasting.

This is closer to the real knowledge work scenario than a "super universal chat box".

It is very friendly to team collaboration and standardized output.

The "weekly report template", "competitive product research paradigm", and "data analysis report structure" are all made into Skills. New team members can directly reuse the SKILL.md accumulated by experienced members, which not only saves training costs but also ensures a more unified output style.

The following are the several points that I will pay attention to.

Permissions and Sharing: How to manage within a team or enterprise who can create, who can modify, and who can share skills

Version Control: SKILL.md, as a text, is actually very suitable for working with Git for version management

Observability: When complex tasks involve multiple Skills and the result is unsatisfactory, can one see the "call chain" to facilitate optimization

The SKILL.md portability angle is the real story here. I maintain a library of 100+ skills for Claude Code and the biggest friction has always been vendor lock-in. Being able to reuse those workflows across tools without rewriting is a genuine time saver. Curious how it handles complex multi-step skills with tool-specific syntax.

Amazing am gonna try this out

The portability angle is what gets me. Rebuilding the same workflow logic every time you switch models is one of those invisible costs that adds up fast.

Curious how it handles context switching when the same skill file gets interpreted differently across models. Does it normalize behavior or just pass through as-is?

This looks really interesting.

Are the repeatable instructions more like automation workflows or closer to scripting for the computer?

Curious how flexible it is for developers.

The SKILL.md portability across Claude Code, Codex, and now Perplexity is a smart move — making workflows tool-agnostic rather than locking them into one ecosystem solves a real pain point for power users. Curious whether there's a versioning or diff system planned for skills, since iterating on a workflow across 19 models likely produces very different outputs than the original single-model setup.

Computer use is the big next step. Going from "answer my question" to "do this task on my computer" changes everything. I've been waiting for this since Anthropic demoed computer use last year. The repeatable instructions part is what I care about most. I have about 15 things I do every morning that could probably be automated if the AI could just watch me do it once. How reliable is it with apps that change their UI frequently? That's usually where automation breaks.

The SKILL.md integration is a game-changer for agent workflow standardization. What I find most compelling is the 19-model execution - this solves the vendor lock-in problem that's plagued AI tooling. Being able to import existing Claude Code or Codex setups without rewriting means teams can actually compound their workflow investments rather than starting fresh with each new tool.

Curious about the activation logic: does the context-based workflow triggering use semantic matching or keyword patterns? And for teams with large skill libraries, how does it handle priority conflicts when multiple workflows could match?

The portability angle is the real unlock here. Most of us have invested significant time building workflows in one environment - the idea that these become transferable assets across 19 models changes the economics of AI workflow investment entirely. No more vendor lock-in on your institutional knowledge. Curious about the context activation mechanism - how granular is the trigger detection for choosing between similar workflows?