Effort Advisor
haider0072/effort-advisor
Avalia complexidade e recomenda nível de esforço antes de cada tarefa
Ver no GitHubPlugin que avalia automaticamente a complexidade de tarefas e recomenda o nível de esforço apropriado (low/medium/high/xhigh/max) antes de cada ação substantiva. Evita executar problemas difíceis com configuração inadequada de esforço, equilibrando precisão, velocidade e custo.
Instalação
npx convez-ai haider0072/effort-advisor1 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
- Otimizar token usage ajustando esforço conforme complexidade da tarefa
- Garantir qualidade em problemas difíceis executando com esforço suficiente
- Automatizar triagem de tarefas triviais versus substantivas
Veredito

Plugin pragmático que resolve um problema real do Opus 4.8: o esforço manual e propenso a erros na escolha do nível de reasoning. Baseado em sinais concretos, não em vibes. Instala-se uma vez e funciona transparente em toda sessão.
README
A tradução para o português deste README ainda está na fila.
effort-advisor
A Claude Code plugin that makes Claude assess task complexity and recommend the right effort level before every substantive task — so you never run a hard problem at low effort and get a weak result.
Claude Opus 4.8 ships an effort dial (low → medium → high → xhigh → max) that controls how deeply the model reasons before acting. More effort = more accurate but slower and pricier; less effort = fast and cheap but shallow. The catch: it's manual, and it's easy to leave a hard task running at the wrong level.
effort-advisor fixes that. Before any non-trivial task, Claude tells you how complex it thinks the task is and which effort level to use — then waits for you to decide.
What it looks like
Ask Claude to do something non-trivial, and before it starts you'll see:
Complexity: High — multi-layer root cause, touches production data, 4+ files
Recommended effort: xhigh → run /effort xhigh
Then it waits. You stay in control. Trivial work (one-liners, quick lookups, tiny edits) is skipped so it never becomes noise.
It is accuracy-biased by design: when complexity is unclear, it rounds up. Under-rating is worse than over-rating — extra tokens cost far less than a wrong result on a hard task.
How it works
A single UserPromptSubmit hook injects a short instruction into context on every prompt. No model-invocation guesswork, no settings to flip, nothing to remember — install it and it's always on. The full instruction lives in plugins/effort-advisor/hooks/effort-instruction.md and is easy to customize.
How complexity is judged
Ratings are based on concrete signals, never a vibe:
| Signal | Pushes complexity up when… |
|---|---|
| Scope | many files / modules / services affected |
| Coupling | affected parts depend on each other |
| Reasoning depth | root cause is multi-layer, not obvious |
| Risk / reversibility | DB migrations, prod changes, irreversible actions |
| Unknowns | lots of exploration needed just to understand it |
If Claude can't rate the task confidently without looking at the code first, it says so instead of guessing.
Effort scale
| Level | Use for |
|---|---|
low | classification, lookups, bulk simple edits |
medium | routine drafting, summaries, straightforward changes |
high (default) | most real coding and non-trivial debugging |
xhigh | hard problems, long-running async work |
max | the heaviest jobs where only the best result matters |
Install
1. Add the marketplace
/plugin marketplace add haider0072/effort-advisor
2. Install the plugin
/plugin install effort-advisor@haider-plugins
3. Activate
/reload-plugins
That's it. The advisor is now on for every session.
Try it locally first (without installing)
git clone https://github.com/haider0072/effort-advisor.git
claude --plugin-dir ./effort-advisor/plugins/effort-advisor
Customize
Edit plugins/effort-advisor/hooks/effort-instruction.md to change the wording, the complexity signals, or the language Claude responds in. Then bump version in both plugins/effort-advisor/.claude-plugin/plugin.json and .claude-plugin/marketplace.json so installed users pick up the update on /plugin marketplace update.
Repo layout
effort-advisor/
├── .claude-plugin/
│ └── marketplace.json # marketplace catalog (lists the plugin)
├── plugins/
│ └── effort-advisor/
│ ├── .claude-plugin/
│ │ └── plugin.json # plugin manifest
│ ├── hooks/
│ │ ├── hooks.json # UserPromptSubmit hook wiring
│ │ └── effort-instruction.md # the instruction injected each prompt
│ └── README.md # plugin-level docs
├── README.md # you are here
└── LICENSE
License
MIT © Haider Ali Khan