Skills that make Claude Code proactively suggest its own power tools - workflows, goals, loops, hooks - at the right moment
Overview
Skills that make Claude Code proactively smarter about its own capabilities. Claude Code ships with serious force-multipliers - multi-agent workflows, session goals, recurring loops, hooks, custom skills - but you have to know to ask for them. These skills flip that: , with a concrete sketch and an honest cost, and you decide. Using OpenAI Codex instead? The same scout, retargeted at Codex's native tool set (subagent fan-out, goals, automations, lifecycle hooks), lives at codex-power-tools. The opportunity spotter. While you work on ordinary tasks, it watches for eight shapes and surfaces a proposal at the right seam: Core etiquette baked in: one proposal per opportunity, never stacks proposals, never treats enthusiasm as approval, drops it without ceremony if you decline. The payoff shape: a "how do we systematically audit these 25 endpoints?
README
Claude Power Tools
Skills that make Claude Code proactively smarter about its own capabilities.
Claude Code ships with serious force-multipliers - multi-agent workflows, session goals, recurring loops, hooks, custom skills - but you have to know to ask for them. These skills flip that: Claude notices when your task fits one of its power tools and proposes it, with a concrete sketch and an honest cost, and you decide.
Using OpenAI Codex instead? The same scout, retargeted at Codex’s native tool set (subagent fan-out, goals, automations, lifecycle hooks), lives at codex-power-tools.
Skills
suggest-power-tools
The opportunity spotter. While you work on ordinary tasks, it watches for eight shapes and surfaces a proposal at the right seam:
| It notices… | It suggests… |
|---|---|
| The same question across many files/items, audits, claims worth adversarially verifying | A multi-agent Workflow (with pipeline sketch + cost, never auto-launched) |
| Several separable, long-lived tracks that must negotiate an evolving contract, each steered in parallel | An agent team (a lead + peer teammates; writers worktree-isolated), never for tight-loop tuning on shared files |
| A multi-turn mission with a mushy finish line | A /goal with an evaluator-verifiable condition, drafted ready to fire |
| Babysitting work: waiting on CI, recurring checks, “remind me later” | /loop or /schedule |
| You just did something the hard way that will recur | Capturing it as a skill while the context is hot |
| “From now on, when X…” or a repeated manual ritual | A settings hook (the one automation memory can’t do) |
| A durable design principle crystallizing mid-conversation | A one-line decision-log entry |
| Task gravity not matching the effort applied | Effort recalibration - acts on its own levers (narrating them), proposes yours (/fast, /model, budgets) |
Core etiquette baked in: one proposal per opportunity, never stacks proposals, never treats enthusiasm as approval, drops it without ceremony if you decline.
The payoff shape: a “how do we systematically audit these 25 endpoints?” aside becomes a multi-agent workflow that produces a rerunnable consistency checker and a grounded drift report - a durable tool you’d never have thought to ask for.
Install
Copy any skill directory into your skills folder:
git clone https://github.com/Matt-Hulme/claude-power-tools.git
cp -r claude-power-tools/suggest-power-tools ~/.claude/skills/
Skills follow the standard SKILL.md convention (YAML frontmatter + progressive-disclosure references), so they also work with any agent runtime that reads Claude-style skill directories. If you rename the directory, update the name: field in the frontmatter to match.
Optional: the firing hook
A proactive skill only helps if Claude remembers to consult it, and nothing re-injects that reminder at the moment the tool choice is made - so it under-fires. The bundled UserPromptSubmit hook fixes that. It prints a one-line “consider a power tool” reminder only when your prompt carries opportunity-shaped language (systematically, across the board, one at a time, new feature, migrate, remind me, …), and stays completely silent otherwise - so it costs zero context on ordinary turns.
Wire it into ~/.claude/settings.json (merge into any existing hooks block, don’t replace it):
{
"hooks": {
"UserPromptSubmit": [
{
"hooks": [
{ "type": "command", "command": "~/.claude/skills/suggest-power-tools/hooks/power-tool-scout.sh", "timeout": 5 }
]
}
]
}
}
The script ships at suggest-power-tools/hooks/power-tool-scout.sh (so after install it lives at the path above). Make it executable (chmod +x), then open /hooks once or restart to load it. It’s precision-biased by default - loosen the pattern regex in the script if it under-fires, tighten it if it nags. Requires jq.
License
MIT
Recommended Tools
Try a different keyword or remove a filter.
Install
npx skillfish add matt-hulme/claude-power-tools