A Claude Code plugin of high-leverage skills that improve orchestration, integration, and one-shot task completion with minimal prompt overhead.
This repo contains compact skills for agentic workflows.
These skills are not meant to add net-new intelligence. They are meant to extract more value from capabilities the agent already has — planning, execution, sub-agents, validation, and tool use.
The goal is simple: small prompts, disproportionate workflow improvement.
AI coding agents are good out of the box. They read files, write code, run tests. But left to their own defaults, they work linearly — one hypothesis at a time, one file at a time, one pass. The failure mode isn't stupidity. It's underutilization. The agent has sub-agent dispatch, parallel execution, isolated worktrees, web search, and browser automation available. It almost never uses them together without being told to.
That's the gap this plugin fills. Each skill here is a small, structured prompt that changes the shape of how the agent works — not what it knows, but how it moves. Pre-flight recon before editing. Parallel research across web and codebase simultaneously. One-prompt API integrations that chain sub-agents through a full build pipeline. Friction detection that turns recurring pain into permanent solutions.
The design constraint is intentional: high leverage per token. A skill that takes 200 tokens of prompt and saves 30 minutes of manual orchestration is worth keeping. A skill that reminds the agent to write tests is not. The qualify agent enforces this bar on every addition.
A good skill is not just a reminder. It should do one of three things:
- act as a force multiplier by unlocking more of the agent's existing capabilities
- improve one-shot completion by increasing the odds a large task is fully completed in one pass
- do both
This repo is focused on high leverage per token.
Every skill is a single SKILL.md. The description field controls when Claude auto-invokes it — see each file for the full body.
| Skill | What it does |
|---|---|
agentify |
Bootstraps a dispatchable sub-agent skill for every coding agent CLI installed on the machine (codex, aider, cursor, etc.). |
appmap |
Maps a web UI and generates a Claude skill to automate it via parallel orchestration. |
automate |
Sets up a cron/launchd job to run Claude headlessly and send results to Telegram. |
contract |
Reference convention for sub-agent I/O schemas. Loaded by orchestrator skills via /agent-workflow-amplifiers:contract and into agents via skills: field. |
forge-friction |
Surfaces recurring friction patterns detected across sessions and identifies actionable skill opportunities. |
ground-state |
Before any non-trivial implementation, dispatches a parallel recon wave (git + infrastructure + memory) that produces a 5-line ground-truth snapshot to prevent wrong-branch edits, assumed-no-CI, and missed memory context. |
integrate |
One-prompt API integration pipeline. Any API (docs URL, name, or description) → working CLI wrapper + skill + tests. |
provideme |
One-prompt provider bridge. Any coding agent CLI → local Anthropic-compatible API server (POST /v1/messages) with a {provider}-claude launcher. |
research |
Dispatches two sub-agents in parallel — one to search the web, one to inspect the local codebase — and returns a merged research brief. |
resolve |
Resolves PR code review feedback by dispatching one parallel sub-agent per issue. |
ship |
Release pipeline for already-done work: /ground-state pre-flight → tests → commit (user-approved) → push → PR with structured verification summary. --verify triggers an adversarial verifier wave on the diff. |
spec |
Transforms a loose idea into a structured, actionable spec ready for implementation. |
web |
Orchestrates parallel browser automation across independent Chrome tabs, one sub-agent per tab. |
| Agent | What it does |
|---|---|
qualify |
Gate-keeps proposed skills. Approves only real force multipliers; rejects reminders, checklists, and best-practice nudges. |
Inside a Claude Code session:
/plugin marketplace add griffinwork40/agent-framework
/plugin install agent-workflow-amplifiers@griffinwork40-agent-framework
/reload-plugins
git clone https://github.com/griffinwork40/agent-framework.git
claude --plugin-dir ./agent-frameworkRequires Claude Code v1.0.33 or later.
Examples:
/integrate stripe— read Stripe's docs, produce a CLI wrapper + skill + tests./research— before starting a task, get web + codebase context in one shot./spec "idea in one sentence"— turn an idea into a scoped plan./resolve— resolve PR review feedback in parallel, one sub-agent per issue./ship— hand off already-done work: pre-flight, tests, commit, push, PR in one command./ground-state— pre-flight recon before any non-trivial implementation./qualify— evaluate a proposed skill before adding it./unqualify— remove a qualified skill that has drifted and log the removal.
The plugin includes a friction analysis pipeline that learns from how you work.
How it works:
- Session telemetry — hooks automatically log every tool call during your Claude Code sessions
- Pattern detection — at session end, an analyzer detects friction patterns (repeated failures, edit cycling, sequential searches that should be parallel)
- Cross-session aggregation — patterns are tracked across sessions with a 28-day sliding window
- Threshold surfacing — when the same friction appears in 3+ sessions,
/forge-frictionsurfaces it with examples and a structured skill brief
The result: recurring pain points are identified and documented, ready to be turned into permanent solutions.
Data stays local. All telemetry is stored in ~/.claude/agent-framework/ — session transcripts are deleted after analysis, only aggregated pattern signatures persist.
A skill belongs in this repo only if it does at least one of:
- changes the shape of the workflow (adds phases, parallelism, sub-agent dispatch)
- unlocks more of the agent's existing capabilities
- materially improves one-shot task completion
- creates a repeatable workflow upgrade the base agent wouldn't reliably apply by default
Useful behaviors like validation, review, or TDD may still matter, but they are usually not strong enough to deserve first-class status on their own. The qualify agent enforces this bar — see agents/qualify.md for the rubric.
See CONTRIBUTING.md. New skills are expected to pass the qualify gate before submission.
Apache 2.0 — see LICENSE.