AI agents · OpenClaw · self-hosting · automation

Quick Answer

Claude Skills vs ChatGPT Plugins vs Gemini Skills (2026)

Published:

The short answer

In 2026 the three big assistants converged on the same idea — a reusable, on-demand instruction package called a “skill” — and two of them are killing their older formats to get there. Claude Skills (a SKILL.md folder with optional scripts and files) are the reference design and run across Claude.ai, Claude Code and the API. OpenAI is retiring custom GPTs on December 11, 2026 and migrating them into plugins, where the GPT’s instructions become a skill. Google is retiring Gemini Gems starting November 17, 2026 in favor of skills you invoke with / in any chat. If you write your instructions once in the skills format, they travel; if you build in Gems or GPTs today, you are building on a deadline.

Side by side

Claude SkillsChatGPT plugins (ex custom GPTs)Gemini skills (ex Gems)
UnitFolder with SKILL.md (instructions + metadata) plus optional scripts, templates, reference docsPlugin = skills + reference files + apps (connected services), plugin.json manifestReusable custom instructions, optional reference files (≤100 MB)
Where it runsClaude.ai, Claude Code, Claude API, and any Agent Plugins clientChatGPT, Codex; and via the Agent Plugins standard Cursor, GitHub Copilot, VS Code, KiroGemini app chat (personal accounts; work/school “soon”); Gemini CLI / Vertex have separate developer Agent Skills
InvocationAutomatic: model reads name + description, loads the full skill only when relevantPlugin selects the right skill; installed per user or workspace/skill-name (soon @); also auto-applied when Gemini judges it relevant; stackable
Code executionYes — bundled Python scripts run when instructions call for themApps and MCP servers; custom GPT “actions” must be rebuiltNo
PortabilityHigh: Agent Skills is the format inside the Agent Plugins standardHigh for plugins; zero for legacy GPTs after Dec 11, 2026Low: Gemini app only; Google lists importing skills from other platforms as a FAQ but the instructions are manual copy-paste
SharingFiles in a repo; org-managed in Claude CodePlugin starts private after migration; workspace publishingGems were link-shareable; skills live in any chat, no dedicated page
Sunset—Custom GPTs retire Dec 11, 2026 (Feb 11, 2027 with Enterprise deferral)Gems → skills Nov 17, 2026 (personal), Mar 2027 (Workspace business), Jun 2027 (education)
Not supported—Custom actions, selected model, old conversations do not migrateCanvas, Deep Research, guided learning, Create video/music

Format background: Agent Plugins vs MCP vs Agent Skills and what is the OpenAI Agent Plugins standard. The April 2026 version of this comparison, written when Gems and custom GPTs were still the plan, is at Claude Skills vs custom GPTs vs Gemini Gems.

Claude Skills: the reference design

A Claude Skill is a directory. SKILL.md carries frontmatter (name, description) and the procedure; the folder can also hold scripts, templates and reference files. The design principle is progressive disclosure: Claude pre-loads only each skill’s name and description, decides relevance from that, reads the full SKILL.md only on a match, and opens bundled files only when the instructions call for them. That keeps dozens of skills resident without flooding the context window.

Skills work in Claude.ai, in Claude Code (where the bundled Claude API skill will, for example, migrate a codebase to Sonnet 5.5 with /claude-api migrate), and through the API. Because the Agent Plugins standard (August 6, 2026; OpenAI, Amazon, Cursor, Microsoft, Vercel, with Google as a core maintainer) adopted Agent Skills as its expertise layer, a well-written SKILL.md is the most portable artifact in this space. Security notes for anything that bundles executable scripts: how to secure AI coding-agent plugins and skills.

ChatGPT plugins: what happens to your custom GPT

OpenAI’s timeline, from its retirement FAQ:

  • September 11, 2026: retirement announced.
  • September 16, 2026: creating new custom GPTs restricted to Business, Enterprise and Edu workspaces.
  • December 11, 2026: custom GPTs and their GPT pages become inaccessible. Enterprise workspaces with an approved deferral get until February 11, 2027.

Migration is per GPT: My GPTs → Migrate to plugin. The latest published version migrates (drafts do not); its instructions become a skill inside the plugin, knowledge files become reference files, connected apps become apps. The plugin starts private and is not auto-installed for other users. Three things do not carry over: custom actions (rebuild as an available app or a custom MCP server), the GPT’s selected model (Enterprise defaults apply), and existing conversations. After migration the original GPT becomes read-only until retirement.

If you use someone else’s GPT, watch its notice for the creator’s replacement; permission to use a GPT does not include permission to migrate it. Plugins run in ChatGPT and Codex and, through the standard, in Cursor, Copilot, VS Code and Kiro — which is the upgrade: a GPT was ChatGPT-only, a plugin is not. Related: ChatGPT voice, plugins and GPT-6 switching.

Gemini skills: what happens to your Gems

Google’s help center and the Gems manager notice (“Gems will become skills starting Nov 17, 2026”):

  • Personal Google accounts: migration begins November 17, 2026. Workspace business, enterprise, non-profit: March 2027. Workspace education: June 2027.
  • Gems are migrated automatically, with supported files. You can recreate one earlier: open the Gem, download its knowledge files, then Settings → Skills → Create manually, paste name, description and instructions, and upload files by downloading the skill and re-uploading it with them.
  • Opal (Google Labs’ mini-app builder) and Gems by Google Labs shut down at the same time in November.

What you gain: / invocation from any chat (@ soon), automatic application when Gemini recognizes a match, and stacking several skills in one conversation — Gems required a side-panel detour and could not be combined. What you lose: Gems could set a default tool (Create image, Canvas) and be shared by link; skills do not yet work with Canvas, Deep Research, guided learning, or video and music creation, and there is no dedicated page listing the chats that used a skill. Skills are available now to signed-in adults on personal accounts (AI Pro and Ultra subscribers can also build them in the Gemini Spark tab); work and school accounts follow. Developers building agents on Gemini CLI or Vertex AI use a separate, more capable “Agent Skills” system.

How to write a skill that survives all three

  1. Lead with the trigger. The description must say when to use it — that is what Claude and Gemini read to decide. Google’s own best practice: “describe specific situations when Gemini should use the skill automatically.”
  2. Keep the body to the procedure. Unique guidelines only; no restating defaults. Put long references in bundled files (Claude) or reference files (plugins, Gemini ≤100 MB).
  3. Include examples of expected output. All three vendors recommend it.
  4. Keep integrations outside the skill. Tool connections belong in MCP servers (Claude, plugins) or connected apps (Gemini); GPT custom actions did not survive migration for exactly this reason.
  5. Store the source in a repo as a SKILL.md folder. From there: use directly in Claude, wrap in a plugin.json for ChatGPT/Codex/Cursor/Copilot, and paste the instructions into a Gemini skill.

Decision rule

  • Daily driver is Claude or Claude Code: Claude Skills, in a repo, optionally packaged as an Agent Plugin.
  • Daily driver is ChatGPT and you own custom GPTs: migrate to plugins now; rebuild custom actions as MCP before December 11, 2026.
  • Daily driver is the Gemini app: recreate important Gems as skills before November 17, 2026 rather than waiting for auto-migration, and check the missing-tools list.
  • You ship one extension to many clients: author Agent Skills, package as an Agent Plugin; treat Gemini skills as a manual export.

Last verified: September 29, 2026.

Sources