The biggest skills update since launch — extending the format to support the broader ecosystem while adding powerful new capabilities for Amplifier users.
Anthropic has been blending a lite version of agents and commands into their extension of the skills format in Claude Code — context: fork, agent types, hooks, model selection. We've brought those same capabilities to Amplifier while maintaining compatibility with the official Agent Skills open standard.
Full compatibility with the Agent Skills spec at agentskills.io. Skills authored to the standard work across AI tools.
Claude Code added context: fork, agent types, hooks, and model selection to their skills format. We now support these same extensions in Amplifier.
Skills authored for Claude Code's large user base work in Amplifier. Skills written in Amplifier work in Claude Code. One format, many runtimes.
Amplifier agents keep their full capabilities — custom tool composition, session resumption, deep orchestration. Skills are a complement, not a replacement.
Forked subagents, model routing, tool restrictions, lifecycle hooks — capabilities that used to require a full agent definition, now available in a Markdown file.
Same Markdown + YAML frontmatter you already know. New fields are additive — existing skills keep working unchanged.
Skills evolved from knowledge injection to a lightweight agent runtime.
contextExecution mode: fork for subagent
agentSubagent archetype to use
model_roleSemantic model routing
modelCross-platform model hint
provider_preferencesExplicit provider chain
auto-loadAlways-on at session start
user-invocableRegister as /command
disable-model-invocationHide from LLM auto-use
allowed-toolsRestrict tool access
hooksSkill-scoped lifecycle hooks
context: forkSkills run as isolated subagents. Separate context, separate model, clean return.
/code-review in chatauto-load: trueSkills with hooks activate at session start. The agent can't bypass these.
auto-load: true emit skill:loaded at session startdisable-model-invocation: true to make invisible but enforcedmodel_role + chat commandsSkills declare what they need. The routing matrix decides how to satisfy it. Same mechanism as delegate and recipes.
| Priority | Field | Example |
|---|---|---|
| 1 | provider_preferences | Explicit provider chain |
| 2 | model_role | fast, coding, reasoning |
| 3 | model | haiku → fast |
| 4 | agent | Explore → fast |
| 5 | none | Inherit parent session |
Skills with user-invocable: true register as chat commands. Same UX pattern as /modes.
Same discovery pattern as modes. /skills, /skill, dynamic shortcuts in /help.
| Command | Action |
|---|---|
/skills | List all available skills |
/skill <name> | Load a specific skill |
/code-review | Dynamic shortcut |
/mass-change <desc> | Dynamic shortcut |
/session-debug | Dynamic shortcut |
/skills-assist <q> | Dynamic shortcut |
SkillsDiscovery object written to session_state during mount — the app reads it to populate commands and /help content.
7 measures to ensure skills can't escape their sandbox.
Shell preprocessing runs with a clean env — no API keys or secrets leak into !`command` execution
Fork skills run as separate subagent sessions with their own context window boundary
Remote-source skills are marked trusted: false — shell preprocessing is blocked for untrusted sources
Subagent results wrapped to prevent prompt injection back into the parent context
Remote skill sources must use HTTPS — no HTTP, no file:// outside workspace
Skill discovery resolves symlinks and rejects paths that escape the declared skills directory
Invalid or corrupted skill cache entries are detected and cleaned up at discovery time
Share skills across Amplifier, Claude Code, and GitHub Copilot via the Agent Skills open standard at agentskills.io.
| Feature | Amplifier | Claude Code | Copilot |
|---|---|---|---|
| Basic SKILL.md | ✅ | ✅ | ✅ |
| context: fork | ✅ | ✅ | — |
| model_role | ✅ | via model | — |
| hooks | ✅ | ✅ | — |
| allowed-tools | ✅ | ✅ | — |
| $ARGUMENTS | ✅ | ✅ | ✅ |
Write once, symlink everywhere. Enhanced features activate when the runtime supports them.
amplifier-module-tool-skills absorbed into amplifier-bundle-skills. One repo, one home.
amplifier-bundle-recipesStart with a skill. Graduate to an agent when you hit a wall. Your content carries forward.
/command in chatThe convergence path: Anthropic already merged commands into skills and added subagent execution. Over time, skills will gain more agent capabilities. Start with a skill — your instructions and companion files carry forward if you graduate to an agent.
/code-review3 parallel review agents with battle-tested prompts. Reuse, quality, efficiency — then auto-fix.
Searches for existing utilities, flags duplicated logic, suggests existing functions to use instead
Flags redundant state, parameter sprawl, copy-paste patterns, leaky abstractions, dead code
Catches N+1 patterns, missed concurrency, hot-path bloat, memory leaks, TOCTOU anti-patterns
/mass-changeDescribe a change → decompose → plan → approve → parallel agents → PRs.
model_role: reasoning allowed-tools: Read Grep Glob Bash Workers auto-invoke /code-review on their own changes.
/session-debugDelegates to the session-analyst agent. Diagnoses config, provider, and tool issues.
model_role: general allowed-tools: Read Grep Glob Bash
/skills-assistThe skills expert. Context-sink "agent-as-skill" pattern. Both users and the LLM can invoke it.
A forked subagent carries heavy reference docs in its own context window. The parent session gets a concise answer back — never burdened with the full docs.
disable-model-invocation is not set — the LLM can auto-discover and delegate hereFour commands to experience the upgrade.
3 parallel review agents
Parallel work orchestration
Session diagnostics
Skills expert