skills

Amp Hosted Skills

Start with one instruction-only skill such as Next Steps. A hosted copy makes the instructions available across Amp environments; it does not supply local tools, credentials, transcript stores, or a new provider adapter. This guide is based on Amp's Skills and Global Plugins & Skills documentation checked on 2026-10-04. Hosted publication and live skill execution have not been verified for this collection.

Choose the scope

ScopeDestination and ownerUse it for
Hosted personalYour Amp personal skills Git repositoryA reviewed skill you want across Amp environments
Hosted workspaceThe workspace skills Git repository, managed by workspace adminsA tested skill shared with all workspace members
ProjectThe project's .agents/skills/Skills and resources that should travel with that project
Machine-localamp skill add <source> --global uses ~/.config/agents/skills/Skills needed only on that machine

The first-party --agent amp installer selects project .agents/skills/ or user ~/.agents/skills/ placement. Its user scope is also machine-local; neither installer publishes to hosted personal or workspace storage. Amp's use of “global” for hosted repositories and the CLI's --global flag refer to different scopes.

Select a compatible payload

Copy from the generated skills/<name>/ directory in a reviewed upstream revision. Preserve its complete files and its metadata.version. The canonical owner remains src/skills/<name>/ in tkstang/skills; the hosted directory is a distribution copy. Never import src/skills/, the entire repository, or plugin skill directories as a substitute for a standalone payload.

Hosted repositories require immediate children such as next-steps/SKILL.md, with directory and frontmatter name matching. A nested skills/next-steps/ layout is not the hosted layout. Amp's recursive local discovery does not establish hosted nesting or symlink support. Copy regular files and directories.

Amp documents a per-repository maximum of 200 skills and 25 MiB, plus 200 files per skill, 10 MiB per file, and 25 MiB per skill; hosted files must be text. Include the repository's existing contents when checking capacity. Above 200 skills, a Git push can succeed while only the first 200 alphabetical directories are loaded. Keep large or non-text resources in a project skill rather than removing required resources to fit the hosted limit.

As of pull request #131 on 2026-10-12, the 19 standalone payloads contain 104 regular text files totaling 5,632,083 bytes. The largest skill has 19 files and 2,180,634 bytes; the largest file is 516,172 bytes. No symlinks or name/directory mismatches were found. These are static packaging measurements, not live Amp acceptance. Do not bulk-import the collection: choose by its runtime requirements and existing skill names.

SelectionStatic portability boundary
next-steps, explain-options, align, must-we, debriefInstruction-only starting points; no bundled runtime
sitrepInstruction-only; reads the tracking artifacts the session already uses and runs read-only Git when the work is in a repository
author-skill, complexity-reviewInstruction-only; filesystem/repository access is needed for file work, and Complexity Review uses Git for repository history
session-handoffCore handoff can use conversation and repository evidence; observer/export integrations are optional and do not gain Amp transcript support
babysit-pr, land-prNeed authenticated GitHub access, Git and repository checks; Land PR also requires Babysit PR and merge authority
phone-a-friendNeeds Node.js 22+ and an installed, authenticated supported peer CLI; hosted storage does not provide these prerequisites
consensus-reviewRequires Node.js 22+ and a supported actual host/provider; its current --host accepts Claude, Codex or Cursor, not Amp
session-export-transcript, session-search, session-fork-to-destination, session-observer, session-observer-collab, session-retro, reorientDepend on supported local provider stores, identities or exporter/observer workflows; no Amp transcript adapter is included
agent-messagingLocal collaboration state and exact supported native identities; its current runtime pins exclude Amp
seshtk-cliExecution needs Node.js 22+, the external Session Toolkit CLI, and supported local Claude Code, Codex or Cursor stores; no Amp transcript adapter is included. Explanations need no CLI.

Six plugin-only Consensus workflows (create, decide, evaluate, panel, plan, refine) use a shared helper outside their skill directory. Extracting those directories loses the complete plugin boundary. Use the supported plugin form in an appropriate host; the collection's Claude/Codex/Cursor plugin manifests are not Amp plugin registrations.

Prepare a hosted personal copy

Amp can manage hosted repositories through a thread. For a skill already published with an Amp share URL, ask it to import that shared skill into your personal skills. Amp records import origin and can prepare later updates. A GitHub directory URL is not established here as an Amp share URL.

For this upstream Git distribution, use the documented repository workflow:

  1. Choose a reviewed upstream commit or release checkout containing the generated payload. Run pnpm run build:check in that developer checkout to verify generation freshness; installation itself needs no build. Note the full commit SHA, payload path, skill version and upstream license.
  2. Run amp skills repositories to identify your repository and write access, then amp clone user-skills to obtain its checkout. Inspect existing skills and capacity before adding the selected payload.
  3. In that hosted checkout, copy the complete reviewed skills/next-steps/ payload to a new immediate child named next-steps/. Refuse an existing destination, including a symlink. Do not overwrite a same-name skill or copy the surrounding upstream skills/ directory.
  4. Compare the copied file inventory and bytes with the upstream payload. Review the Git diff, preserve upstream attribution/license notices, and record tkstang/skills, the full source SHA, skills/next-steps/, and its metadata.version in the hosted commit message. Keep authored fixes upstream.
  5. Review and commit the copy. Push only when you intend to publish to that account. Publication is a separate action from preparing the local copy.
  6. Start a fresh Amp thread or ask Amp to reload skills. Ask which next-steps skill was loaded and where it came from, then try the bounded request in Getting Started. Record actual discovery, invocation and permission results before claiming live compatibility.

Local and built-in skills take precedence over hosted skills; hosted personal skills also mask same-name workspace skills. Inspect origin, not only the name. amp skills list in another shell does not reload an existing thread. If a required tool or compatible identity is missing, stop that dependent workflow. Do not relabel Amp as Codex or install a provider CLI merely because the payload was accepted.

Update and share deliberately

A copied upstream payload has no continuous synchronization contract. Select a new reviewed upstream revision, compare it with the recorded source revision, and prepare replacement of only the previously imported skill directory in the hosted checkout. Review local modifications before replacement, include added and removed payload files, and compare the resulting inventory and bytes with the new generated payload. Preserve its version and record the new source SHA. Review, commit and push to publish; then reload or start a fresh thread and repeat bounded acceptance. Bring portable fixes back to the canonical upstream owner instead of maintaining a second authored skill.

For an Amp shared import, Amp documents asking it to check imported skills for updates or running amp skill update <name> from the hosted repository checkout. Review, commit and push the prepared update. This updates the recorded shared origin; it does not establish automatic tracking of tkstang/skills for a manual Git copy, nor does it silently overwrite your changes.

After personal acceptance, share through Personal Settings → Skills → Share. A workspace admin can copy a tested shared skill into the workspace repository, review and publish it. Sharing a URL and publishing for every workspace member are distinct actions. Recheck the destination's capacity and name conflicts, then verify discovery and bounded behavior in that scope.

On this page