ThunderAI/claude-spec/02-prompts.md

113 lines
5.7 KiB
Markdown

# Prompts System
## Overview
Prompts are the core user-facing feature of ThunderAI. Each prompt defines an AI instruction and how it behaves. There are two kinds:
- **Built-in prompts** — defined in `js/mzta-prompts.js`
- **Custom prompts** — created by the user and stored in `browser.storage.local`
## Prompt Properties
### Base Properties (built-in only)
| Property | Type | Description |
|----------|------|-------------|
| `id` | string | Unique identifier |
| `name` | string | `__MSG_key__` i18n reference or plain text |
| `prompt` | string | The prompt template text (may contain `{%placeholder%}` tokens) |
| `type` | number | `0` = always visible, `1` = reading email only, `2` = composing only |
| `action` | number | `0` = close, `1` = reply (open compose), `2` = substitute text in-place |
| `need_selected` | number | `0` = use full message body, `1` = requires text selection |
| `need_signature` | number | `0` = no signature, `1` = include signature |
| `need_custom_text` | number | `0` = no custom input, `1` = show custom text input field |
| `define_response_lang` | number | `0` = no language hint, `1` = append response language instruction |
| `use_diff_viewer` | number | `0` = normal output, `1` = show diff viewer (ChatGPT Web only) |
### User Properties (stored per-prompt in storage)
| Property | Type | Description |
|----------|------|-------------|
| `enabled` | number | `0` = hidden, `1` = shown in popup |
| `position_display` | number | Sort order in reading view |
| `position_compose` | number | Sort order in compose view |
### Per-Prompt API Override Properties
Each prompt can override the global API connection. These mirror the keys in `integration_options_config` and `prefs_default`:
| Property | Description |
|----------|-------------|
| `connection_type` | Override API type for this prompt |
| `chatgpt_web_model` | Override ChatGPT Web model |
| `chatgpt_web_project` | Override ChatGPT Web project |
| `chatgpt_web_custom_gpt` | Override custom GPT |
| All `chatgpt_*`, `ollama_*`, `openai_comp_*`, `google_gemini_*`, `anthropic_*` keys | Override specific API settings |
## Special Prompts
Some prompts trigger additional Thunderbird actions beyond just sending text to the AI. They are identified by their `id`:
| ID | Feature |
|----|---------|
| `add_tags` | Auto-tag the email after AI response |
| `spamfilter` | Classify as spam and optionally move email |
| `summarize` | Summarize email content |
| `get_calendar_event` | Extract and create a calendar event |
| `get_task` | Extract and create a task |
These special prompts can have their own dedicated API integration settings (configured in the Options page). The list of these special prompts is in `options/mzta-options-default.js` as `special_prompts_with_integration`.
### Summarize: Dual-Mode Prompt System
The summarize feature uses two distinct prompt pathways:
**Context Menu Summarize** (right-click on messages in message list):
- Activated via the `summarize` context menu item, controlled by the `summarize` feature flag
- Uses 3 special prompts stored in `specialPrompts`:
- `prompt_summarize` — the main instruction prompt for the LLM
- `prompt_summarize_email_template` — template for formatting each email's content
- `prompt_summarize_email_separator` — separator text between multiple emails
- Supports multi-email summarization: each selected message is formatted with the email template, joined by the separator, then prepended with the instruction prompt
- All 3 prompts support placeholder autocomplete (`{%placeholder%}` syntax)
- Result is displayed via `openChatGPT()` in the standard chat output window (not inline)
- Default prompt texts are stored as i18n keys: `prompt_summarize_full_text`, `prompt_summarize_email_template_full_text`, `prompt_summarize_email_separator_full_text`
**Inline Summary on Message Display** (automatic or manual per `summarize_auto` pref):
- Uses the same 3 special prompts as webchat mode, via `taPromptUtils.buildSummaryPrompt()` in `js/mzta-utils-prompt.js`
- Does **not** support `chatgpt_web` connection type (shows error if configured)
- Result is rendered as a styled banner at the top of the message body via `mzta-compose-script.js`
- Banner includes a refresh button (↻) to regenerate the summary
- Cached per-message via `taSummaryStore` / `taStorage` (max 100 entries)
**Unified Prompt Building**`taPromptUtils.buildSummaryPrompt(messageDataArray)`:
- All summary paths (inline, webchat single, webchat multi) use this single method
- Accepts an array of `{ message, fullMessage }` entries
- Returns `{ promptText, promptInfo }` where `promptInfo` is the `prompt_summarize` prompt object
## Prompt Types Reference
```
type 0 → shown when reading AND composing
type 1 → shown only when reading an email (message display)
type 2 → shown only when composing an email
```
## Action Types Reference
```
action 0 → no output, just close (e.g. for tag/spam actions handled in background)
action 1 → open a reply compose window with AI response
action 2 → replace selected text (or insert) in compose window
```
## Adding a New Built-in Prompt
1. Add the prompt object to the `defaultPrompts` array in `js/mzta-prompts.js`
2. Add the `name` string key to `_locales/en/messages.json`
3. If the prompt text needs a localized string, add it to `_locales/en/messages.json` as well
4. Reference any needed placeholders using `{%placeholder_id%}` syntax in the `prompt` field
## Custom Prompts
Custom prompts are stored in `browser.storage.local` and managed via `pages/customprompts/`. They follow the same property structure as built-in prompts but are created/edited/deleted by the user through the UI. Custom placeholders can also be referenced in custom prompt text.