Skill Radar
Roni-quant/skill-radar
Invocação inteligente de skills com zero falsos positivos
Ver no GitHubSistema de triagem orientado por sinais para invocar skills no Claude Code apenas quando necessário. Reduz falsos positivos ao usar uma taxonomia explícita de gatilhos, evitando narração forçada e mantendo execução direta para tarefas simples.
Instalação
npx convez-ai Roni-quant/skill-radar1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automação de fluxos de desenvolvimento sem overhead de skills desnecessárias
- Edição de arquivos simples sem narração ou preamble forçado
- Resolução de bugs com roteamento automático para debugging sistemático
Veredito

Skill sólida que resolve um problema real: invoke skills apenas quando faz sentido, não por padrão. Implementação clara com triagem baseada em gatilhos explícitos, ideal para devs que querem eficiência sem ruído.
README
A tradução para o português deste README ainda está na fila.
skill-radar
Trigger-driven skill invocation for Claude Code. Fewer false positives, no forced narration, subagent-safe.
A lean alternative to
using-superpowers— fire skills on signal, not on every keystroke.
What you get
- Explicit Trigger Taxonomy — skills fire on real signals, not "1% chance → invoke."
- No-skill list — single-file edits, lookups, and trivial fixes stay direct.
- Mode-aware announce — silent in terse/caveman modes; one-line for heavyweight skills only.
- Subagent priority injection — re-states the user > skills > default hierarchy in dispatched prompts (closes obra/superpowers#237).
Why
- Triage, not paranoia. Upstream
using-superpowersenforces "1% chance → invoke." That's a false-positive engine — see "The Controllability Problem" for the full case.skill-radarswaps it for an explicit Trigger Taxonomy. - Silent on noise. Single-file edits, lookups, status checks: no skill, no narration, no preamble.
- Mode-aware. Terse / caveman / quiet workflows skip the forced
Announce: Using Xline. - Subagent-safe. Re-states the priority hierarchy in dispatched prompts to close the obra/superpowers#237 injection gap.
Install
Plugin marketplace (recommended)
/plugin marketplace add Roni-quant/skill-radar
/plugin install skill-radar@skill-radar-marketplace
Manual (Claude Code)
git clone https://github.com/Roni-quant/skill-radar ~/.claude/skills-src/skill-radar
ln -s ~/.claude/skills-src/skill-radar/skills/skill-radar ~/.claude/skills/skill-radar
Codex / Gemini install paths are tracked for v0.2 — see Roadmap.
Try it
Once installed, send any of these to Claude Code and watch what skill-radar decides:
| You say | skill-radar does |
|---|---|
Build a CSV export feature for the trades table. | Matches "build" → invokes brainstorming → writing-plans. |
Fix the off-by-one in line 42. | Matches "fix" → invokes systematic-debugging. |
Rename decision_id to trade_id in trades.py. | No match → direct edit. No skill, no narration. |
That's the whole product. Three lines of input → three different routing decisions, no over-firing, no forced preamble.
How it works
Internally, every user message goes through this loop:
- User sends a message.
skill-radarmatches the message against its Trigger Taxonomy.- Match → invoke the indicated skill, then act.
- No match → act direct. No skill load, no announcement.
- If an invoked skill turns out to be a wrong fit mid-task → say so once, abort, continue direct.
- Heavyweight rigid skills (TDD, debugging) get a one-line announce; everything else stays silent in terse modes.
- Subagents inherit a re-stated priority hierarchy in their dispatch prompt.
For full transcripts of each routing decision in real sessions, see Examples.
Comparison
| Dimension | using-superpowers | skill-radar |
|---|---|---|
| Trigger rule | "1% chance → invoke" | Explicit taxonomy match |
| No-skill cases | Implicit / discouraged | Explicit list |
| Mode awareness | Forced narration | Respects terse/caveman |
| SKILL.md size | ~117 lines | ~50 lines |
| Subagent injection | Documented gap (#237) | Re-states priority |
| Description style | Workflow summary risk | Triggers only |
Philosophy
- Triage > paranoia. Most messages don't need a workflow.
- Skills earn invocation. Loading a skill is a context cost.
- Description bloat = drift. Triggers in description; steps in body.
- Subagents need explicit priority. Re-state the hierarchy on dispatch.
Full discussion in docs/philosophy.md.
Examples
examples/transcript-feature.md— build chain.examples/transcript-bugfix.md— debugging chain.examples/transcript-trivial.md— no-skill case.
Compatibility
Coexistence:
skill-radaris a drop-in alternative to upstreamusing-superpowers, not a replacement that breaks it. Both can be installed simultaneously — pick one as the active triage layer in your globalCLAUDE.md.
To prefer skill-radar globally, add to your ~/.claude/CLAUDE.md:
## Skill triage
Use `skill-radar` for skill invocation decisions. Do not invoke `using-superpowers` simultaneously.
Roadmap
See ROADMAP.md for v0.2, v0.3, v1.0 plans and out-of-scope notes.
Support
Questions, bug reports, taxonomy gaps: open an issue.
Contributing
See CONTRIBUTING.md. Short version: edits to Trigger Taxonomy require a new pressure test.
Credits
- obra/superpowers — built the foundation we lean on.
- obra/superpowers-marketplace — marketplace.json shape.
- anthropics/skills — single-repo plugin+marketplace pattern.
- obra/superpowers#237 — subagent injection gap.
superpowers:writing-skills— description-drift fix, TDD-for-skills methodology.
License
MIT — see LICENSE.