Active · March 2026

Major Skills
System Upgrade

The biggest skills update since launch — extending the format to support the broader ecosystem while adding powerful new capabilities for Amplifier users.

Extending skills to support the broader ecosystem

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.

🌐

Open Standard First

Full compatibility with the Agent Skills spec at agentskills.io. Skills authored to the standard work across AI tools.

📌

Anthropic's Extensions

Claude Code added context: fork, agent types, hooks, and model selection to their skills format. We now support these same extensions in Amplifier.

🔄

Two-Way Portability

Skills authored for Claude Code's large user base work in Amplifier. Skills written in Amplifier work in Claude Code. One format, many runtimes.

🤖

Agents Remain First-Class

Amplifier agents keep their full capabilities — custom tool composition, session resumption, deep orchestration. Skills are a complement, not a replacement.

Lite Agent Capabilities in Skills

Forked subagents, model routing, tool restrictions, lifecycle hooks — capabilities that used to require a full agent definition, now available in a Markdown file.

Familiar Authoring

Same Markdown + YAML frontmatter you already know. New fields are additive — existing skills keep working unchanged.

10 new frontmatter fields

Skills evolved from knowledge injection to a lightweight agent runtime.

context

Execution mode: fork for subagent

agent

Subagent archetype to use

model_role

Semantic model routing

model

Cross-platform model hint

provider_preferences

Explicit provider chain

auto-load

Always-on at session start

user-invocable

Register as /command

disable-model-invocation

Hide from LLM auto-use

allowed-tools

Restrict tool access

hooks

Skill-scoped lifecycle hooks

context: fork

Skills run as isolated subagents. Separate context, separate model, clean return.

--- name: code-review description: "Review changed code..." context: fork model_role: critique disable-model-invocation: true user-invocable: true --- # Your skill body becomes # the subagent's task prompt

How it works

User invokes /code-review in chat
Skill body becomes subagent task prompt
Subagent runs with its own context window
Results summarized back to parent
Main context stays clean

auto-load: true

Skills with hooks activate at session start. The agent can't bypass these.

--- name: safety-gate auto-load: true hooks: Stop: - "Never deploy to production" - "without running the test suite" - "first." --- # This hook fires on every # assistant response, enforcing # the rule automatically.

Why this matters

  • Skills with auto-load: true emit skill:loaded at session start
  • Hook-bearing skills activate their hooks immediately
  • Stop hooks inject constraints the agent must follow
  • No explicit invocation needed — always active
  • Combine with disable-model-invocation: true to make invisible but enforced

model_role + chat commands

Skills declare what they need. The routing matrix decides how to satisfy it. Same mechanism as delegate and recipes.

5-level resolution precedence

PriorityFieldExample
1provider_preferencesExplicit provider chain
2model_rolefast, coding, reasoning
3modelhaikufast
4agentExplorefast
5noneInherit parent session

Skills as /commands in chat

Skills with user-invocable: true register as chat commands. Same UX pattern as /modes.

# In chat: /code-review focus on memory leaks /mass-change migrate lodash to native /session-debug bash tool failing /skills-assist how to write a fork skill

Chat commands

Same discovery pattern as modes. /skills, /skill, dynamic shortcuts in /help.

New chat commands

CommandAction
/skillsList all available skills
/skill <name>Load a specific skill
/code-reviewDynamic shortcut
/mass-change <desc>Dynamic shortcut
/session-debugDynamic shortcut
/skills-assist <q>Dynamic shortcut

Follows the modes pattern

SkillsDiscovery object written to session_state during mount — the app reads it to populate commands and /help content.

# How dispatch works: User types: /code-review memory leaks CommandProcessor checks SKILL_SHORTCUTS Constructs: load_skill("code-review") tool-skills detects context: fork Subagent runs with full prompt

Security hardening

7 measures to ensure skills can't escape their sandbox.

🧹

Sanitized Environment

Shell preprocessing runs with a clean env — no API keys or secrets leak into !`command` execution

🏠

Process Isolation

Fork skills run as separate subagent sessions with their own context window boundary

🛡

Trust Gate

Remote-source skills are marked trusted: false — shell preprocessing is blocked for untrusted sources

📦

Output Wrapping

Subagent results wrapped to prevent prompt injection back into the parent context

🔒

HTTPS-Only Sources

Remote skill sources must use HTTPS — no HTTP, no file:// outside workspace

🔗

Symlink Boundary

Skill discovery resolves symlinks and rejects paths that escape the declared skills directory

🗑

Corrupt Cache Cleanup

Invalid or corrupted skill cache entries are detected and cleaned up at discovery time

Cross-tool portability

Share skills across Amplifier, Claude Code, and GitHub Copilot via the Agent Skills open standard at agentskills.io.

One skill, many runtimes

FeatureAmplifierClaude CodeCopilot
Basic SKILL.md
context: fork
model_rolevia model
hooks
allowed-tools
$ARGUMENTS

Symlink sharing pattern

~/.skills/ # Shared home my-review-skill/ SKILL.md # Single source ~/.claude/ # Claude Code skills/ my-review-skill → symlink ~/.amplifier/ # Amplifier skills/ my-review-skill → symlink

Write once, symlink everywhere. Enhanced features activate when the runtime supports them.

Bundle consolidation

amplifier-module-tool-skills absorbed into amplifier-bundle-skills. One repo, one home.

Before & after

Before: 2 repos

amplifier-module-tool-skills/ # Standalone amplifier-bundle-skills/ # Bundle

After: 1 repo

amplifier-bundle-skills/ modules/ tool-skills/ # Module lives here behaviors/ skills/ code-review/ mass-change/ session-debug/ skills-assist/ image-vision/ context/ tests/ # 212 tests

Why consolidate?

  • Follows proven pattern from amplifier-bundle-recipes
  • Single repo for module + skills + behaviors + tests
  • No more cross-repo coordination for changes
  • 212 tests all run from one location
  • Standalone repo gets a deprecation notice, not deleted
  • Existing users continue working — no breaking changes

Skills vs agents

Start with a skill. Graduate to an agent when you hit a wall. Your content carries forward.

Start with a Skill

  • Packaged expertise or repeatable task
  • Needs to be portable (Agent Skills standard)
  • User-invokable via /command in chat
  • Lightweight — no custom tool composition needed
  • Auto-activate via hooks
  • Community-shareable across AI tools

Graduate to an Agent

  • Needs its own tool composition
  • Multi-turn session resumption across turns
  • Precise context inheritance control
  • Custom orchestrator configuration
  • Recursive sub-agent delegation
  • Deep Amplifier bundle integration

The 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-review

3 parallel review agents with battle-tested prompts. Reuse, quality, efficiency — then auto-fix.

🔄

Agent 1: Reuse

Searches for existing utilities, flags duplicated logic, suggests existing functions to use instead

Agent 2: Quality

Flags redundant state, parameter sprawl, copy-paste patterns, leaky abstractions, dead code

Agent 3: Efficiency

Catches N+1 patterns, missed concurrency, hot-path bloat, memory leaks, TOCTOU anti-patterns

# Review all recent changes /code-review # Narrow the focus /code-review focus on memory efficiency # Phase 1: git diff | Phase 2: 3 parallel agents | Phase 3: auto-fix issues

/mass-change

Describe a change → decompose → plan → approve → parallel agents → PRs.

Describe change
Research scope
Decompose 5–30 units
User approves plan
Parallel agents
PRs created
/mass-change migrate from react to vue /mass-change replace all uses of lodash with native equivalents /mass-change add type annotations to all untyped function parameters # Each worker: implement → /code-review → test → commit → gh pr create

model_role: reasoning   allowed-tools: Read Grep Glob Bash   Workers auto-invoke /code-review on their own changes.

/session-debug

Delegates to the session-analyst agent. Diagnoses config, provider, and tool issues.

# Something broken? Just ask. /session-debug # Or describe the issue /session-debug why is the bash tool failing? /session-debug provider keeps rejecting requests

What it checks

  • Bundle composition — missing modules, failed loads
  • Provider config — missing API keys, wrong endpoints
  • Tool availability — expected tools not mounted
  • Hook failures — lifecycle event errors
  • Session logs — event patterns, error chains
  • Explains findings in plain language with fix steps

model_role: general   allowed-tools: Read Grep Glob Bash

/skills-assist

The skills expert. Context-sink "agent-as-skill" pattern. Both users and the LLM can invoke it.

The context-sink pattern

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.

Companion reference files

skills-assist/ SKILL.md # Subagent instructions authoring-guide.md # How to write skills spec-reference.md # Full spec reference compatibility-matrix.md# Cross-tool compat skills-vs-agents.md # Decision guide

Key design decisions

  • disable-model-invocation is not set — the LLM can auto-discover and delegate here
  • When you ask "how do I create a skill?", the LLM matches the description and invokes it
  • Other agents can also delegate to it during their own work
  • Demonstrates that skills can replace lightweight agents
# Users invoke directly in chat: /skills-assist how do I write a fork skill? # Or the LLM invokes automatically # when it sees a skills question

Try it now

Four commands to experience the upgrade.

# Step 1: Update Amplifier amplifier update # Step 2: Review your code /code-review # Step 3: Diagnose any issues /session-debug # Step 4: Plan a big change /mass-change add error handling to all API calls # Step 5: Ask the expert /skills-assist how do I write my own skill?

Write your own in 60 seconds

# Create the directory mkdir -p .amplifier/skills/my-skill # Write the SKILL.md cat > .amplifier/skills/my-skill/SKILL.md << 'EOF' --- name: my-skill description: "My custom review" context: fork model_role: coding user-invocable: true --- Review $ARGUMENTS for issues. EOF # Use it immediately in chat /my-skill the auth module

The biggest skills update since launch

21
PRs merged
4
Repos touched
10
New frontmatter fields
4
Built-in skills
212
Tests
2
Issues closed
🔍

/code-review

3 parallel review agents

🚀

/mass-change

Parallel work orchestration

🐛

/session-debug

Session diagnostics

🧠

/skills-assist

Skills expert