{
  "markdown": "# token-wise\n\nA Claude Code plugin that routes text-heavy tasks to the right Claude model — saving tokens without sacrificing quality.\n\n## Why\n\nPlanning, CLAUDE.md generation, README writing, and PRDs don't need Opus when you're just structuring content. token-wise intercepts these tasks and runs them on Sonnet or Haiku instead, switching back to your session model automatically when done.\n\n## Quick start\n\n```bash\n# 1. Install\nclaude --plugin-dir /path/to/token-wise\n\n# 2. Use model-specific commands\n/token-wise:plan-haiku      # Plan on Haiku (fastest)\n/token-wise:plan-sonnet     # Plan on Sonnet (balanced)\n/token-wise:plan-opus       # Plan on Opus (highest quality)\n\n/token-wise:readme-haiku    # README on Haiku\n/token-wise:prd-sonnet      # PRD on Sonnet\n/token-wise:claude-md-opus  # CLAUDE.md on Opus\n```\n\n## Choosing a model tier\n\n| Model | Best for | Tradeoff |\n|-------|----------|----------|\n| **Haiku** | Simple, well-defined tasks (READMEs, straightforward plans) | Fastest & cheapest; less creative |\n| **Sonnet** | Most tasks (balanced speed, quality, cost) | Good all-rounder; not specialized |\n| **Opus** | Complex specs, architectural plans, high-stakes documentation | Highest quality; slower & costlier |\n\n## Available skills\n\n| Task | Commands |\n|---|---|\n| Planning | `/token-wise:plan-haiku`, `/token-wise:plan-sonnet`, `/token-wise:plan-opus` |\n| PRD / specs | `/token-wise:prd-haiku`, `/token-wise:prd-sonnet`, `/token-wise:prd-opus` |\n| CLAUDE.md | `/token-wise:claude-md-haiku`, `/token-wise:claude-md-sonnet`, `/token-wise:claude-md-opus` |\n| README | `/token-wise:readme-haiku`, `/token-wise:readme-sonnet`, `/token-wise:readme-opus` |\n\n## Auto-detection\n\nMention a text-heavy task in your prompt (e.g., \"create a plan\", \"write a readme\") and token-wise will detect it. A blocking prompt asks you to choose a model and suggests the right command. **You must select a model to continue** — your original prompt is held until you do.\n\nAfter you choose, the skill runs on the selected model tier and automatically returns to your original session when done.\n\n### Permission approval\n\nThe first time you run any token-wise skill, you'll see a permission approval prompt. Select **\"Yes, and don't ask again\"** to skip this prompt in future runs.\n\n## Install\n\n### From the Claude Code Marketplace\n\nThe easiest way to install token-wise:\n\n```bash\n# Search and install from the Claude Code marketplace\n```\n\nIn Claude Code, open the Marketplace (Plugins menu) and search for **token-wise**.\n\n### Local development\n\n```bash\nclaude --plugin-dir /path/to/token-wise\n```\n\nOr add to persistent plugin directories:\n\n```bash\nclaude config set plugins.dirs '[\"~/path/to/token-wise\"]'\n```\n\n### From source\n\n```bash\ngit clone https://github.com/vivek-cactus/token-wise.git\ncd token-wise\nclaude --plugin-dir .\n```\n\n## How it works\n\n1. **Explicit invocation**: Type `/token-wise:skill-model` (e.g. `/token-wise:plan-sonnet`) to run immediately\n2. **Auto-detection**: Mention a text task in chat and the hook detects it, shows model options, and blocks until you choose\n3. **Background run**: The skill runs on the selected model tier\n4. **Auto-return**: Output is written to disk and your session returns to the original model\n\n## Output\n\nEvery skill writes its output to the project root:\n\n- `CLAUDE.md` — project documentation\n- `PLAN.md` — implementation or sprint plans\n- `PRD.md` — product requirements\n- `README.md` — project README\n\n## Architecture\n\ntoken-wise uses a **hook + skill system**:\n\n- **Hook** (`detect_task.py`): Monitors your prompts for task keywords and intercepts text-heavy work\n- **Skills** (`skills/`): Self-contained generators for each document type, running on the specified model tier\n- **CLI integration**: Works seamlessly with Claude Code's native command system\n\n## Contributing\n\nContributions welcome. Fork, branch, and open a PR. Skills are defined in `skills/` as self-contained SKILL.md files.\n\n## License\n\nApache 2.0 — see [LICENSE](./LICENSE)\n",
  "bytes": 3998,
  "sha": "872282a2a3661f53019692f5fe03f8e1a626e728297f34b83ca47c120f9140ec",
  "repo_slug": "vivek-cactus/token-wise",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_vivek_cactus_token_wise_token_wise_b89550e1/readme"
}