Better Prompt for OpenCode

Automatically fixes and improves user prompts before the model sees them. This is the OpenCode build (@objctp/opencode-better-prompt). For the Claude Code version or a project overview, see the GitHub repo.

Install

Add the package to your OpenCode config. OpenCode installs npm plugins automatically at startup (via Bun), so just save the file and restart.

// Project scope: opencode.json (in your project root)
{ "plugin": ["@objctp/opencode-better-prompt"] }

// Global scope: ~/.config/opencode/opencode.json
{ "plugin": ["@objctp/opencode-better-prompt"] }

On first run, the plugin copies a default config to ~/.config/opencode/better-prompt.local.md if none exists.

Requirements

Just OpenCode with at least one provider connected and a model available.

Settings

Edit ~/.config/opencode/better-prompt.local.md. The options are the same as the project overview; the frontmatter looks like this:

---
enabled: true
correction: true
correction_model: haiku
translation: false
translation_model: haiku
enhancement: false
enhancement_model: sonnet
audit: true
verbose: false
---

You can also edit it live with /better-prompt:config — changes take effect on the next prompt.

Models

There are three ways to set a stage's model:

• Built-in aliases — haiku, sonnet, opus. Work everywhere.
• Dynamic aliases — fast, capable, powerful. OpenCode resolves these from the providers you've connected, picking the cheapest model with tool support in each tier.
• Explicit IDs — a full provider/model string, e.g. opencode-go/deepseek-v4-pro.

When a stage's model matches its default, Better Prompt sends no override and the agent just inherits the model your session is already using.

Picking models in the TUI

Inside /better-prompt:config:

• Enter on a model field — cycle the tier: fast → capable → powerful
• Space on a model field — cycle individual models within the tier (sorted by cost)
• Enter on a boolean field — toggle it on or off
• Esc — close

What you see while it runs

Two things appear automatically:

Toasts — a short notice for each prompt: Processing prompt... while it works, then Prompt modified or No changes when it's done (or an error message if a stage failed).

Sidebar panel — a live view of the pipeline, refreshed a few times a second:

Better Prompt
 │ ◆ correction   correcting
 │ ◇ translation  done · 120ms · (en)
 │ ◇ context      summarising
 │ ◇ enhancement  done · 340ms
   $0.0123 · 1.2k→450t  (sess: 5.2k→2.1kt)

Each stage shows a status symbol — ◆ running, ◇ done, ○ skipped, ▲ failed — with how long it took and, for correction, how many mistakes it caught. The footer adds the cost and token use for the current prompt, plus a running total for the session.

Turn verbose on and the correction stage also lists up to five of the mistakes it fixed, inline.

Commands

All three are available as slash commands and from the command palette:

• /better-prompt:toggle — flip stages on or off (palette: BP: Toggle Stage)
• /better-prompt:config — edit settings live (palette: BP: Show Config)
• /better-prompt:audit — show the audit trail (palette: BP: Audit Trail)

How it works

The plugin hooks OpenCode's chat.message event. When you send a message, OpenCode hands the parts to the plugin before the model sees them. The plugin runs the pipeline and writes the improved text straight back into the message — the model only ever receives the result.

The pipeline itself:

your prompt → correction → translation → context → enhancement → result

• Correction — grammar and spelling.
• Translation — non-English to English; skipped if the prompt is already English.
• Context — a running summary of the conversation, fed to enhancement for continuity. Refreshed in full every 10 prompts and updated incrementally in between.
• Enhancement — rewrites for clarity and structure, using that summary as context.

To shield code, commands, or literal text from every stage, wrap it in a code fence (```). The plugin is also instructed to leave content inside block delimiters (:::) and similar structural wrappers alone, though code fences are the most reliable.

A couple of shortcuts keep it cheap: when enhancement is on without translation, correction is folded into enhancement; when translation is on without correction, the correction agent runs only to detect language.

The plugin ignores its own sub-agent calls (correction, translation, enhancement, summarisation), so it never tries to enhance its own work. While it runs it writes a state file to ~/.local/state/opencode/better-prompt/state.json, which the sidebar reads.

Audit log

With audit on, each prompt is appended to <project>/.opencode/better-prompt/audit.json as one JSON object per line:

{
  "date": "2026-03-12T10:30:00Z",
  "prompt": "original prompt text",
  "language": "en",
  "corrected": "prompt after correction (null if correction is off)",
  "enhanced": "final prompt (null if enhancement is off)",
  "mistake-nature": ["grammar", "spelling"],
  "mistakes": [
    { "type": "grammar", "original": "incorrect phrase", "correction": "corrected phrase" }
  ],
  "models": {
    "correction": "haiku",
    "translation": null,
    "enhancement": null,
    "context": null
  },
  "usage": {
    "cost": 0.0032,
    "inputTokens": 1234,
    "outputTokens": 567,
    "cacheWriteTokens": 120,
    "cacheReadTokens": 0
  }
}

• language is what the correction stage detected.
• corrected is the text after correction, before enhancement.
• enhanced is the final prompt after every stage.
• A model field is null when its stage was off; models.context is the model used for summarisation.
• usage is the token cost across all stages for that one prompt.

The file grows without bound, so rotate or clear it with /better-prompt:audit if it gets large.

Troubleshooting

Plugin isn't loading

• Confirm the package is in your opencode.json and you've restarted OpenCode.
• OpenCode installs npm plugins via Bun at startup — check the OpenCode log if install failed.

Dynamic models (fast / capable / powerful) don't resolve

• These come from your connected providers. Connect at least one provider with a model in OpenCode's settings, or set an explicit provider/model ID instead.

Settings changes aren't picked up

• Check you're editing ~/.config/opencode/better-prompt.local.md. Changes via /better-prompt:config apply immediately; manual edits apply on the next prompt.