prompt-distiller（即时蒸馏器）

Prompt Distiller

Transform iterative conversations into optimized one-shot prompts. Learn from multi-turn exchanges to create prompts that succeed on the first try.

Mode Selection

Determine which mode applies:

Distilling a conversation? User wants to create a one-shot prompt from a session.

• Follow: DISTILL Workflow below

Planning a new task? User wants to plan ahead with clarifying questions.

• Follow: PLAN Workflow below

Listing sessions? User wants to see available conversations.

• Follow: LIST Workflow below

DISTILL Workflow

Step 1: List Available Sessions

Run the session listing script to find conversations:

python scripts/list_sessions.py [--agent opencode|claude|gemini|codex|all] [--limit 20]

Present sessions to user with:

• Session ID/timestamp
• Initial prompt preview
• Message count
• Duration/turns

Step 2: Load and Parse Session

Once user selects a session or provides a conversation file/URL:

# Parse OpenCode/Claude/Gemini/Codex session by ID
python scripts/parse_session.py <session-id-or-path> [--agent opencode|claude|gemini|codex]

# Parse external conversation files (.md, .txt)
python scripts/parse_session.py conversation.md
python scripts/parse_session.py chat_export.txt

# Parse from URL (automatically downloads large files)
python scripts/parse_session.py https://example.com/conversation.md
python scripts/parse_session.py https://gist.githubusercontent.com/.../chat.md --save-dir ~/downloads/

The script will:

• Detect agent type if not specified (supports opencode, claude, gemini, codex, file, or url)
• Download files from URLs if too large to read online (saved to ~/.cache/prompt-distiller/downloads/)
• Parse external.md/.txt files with various conversation formats
• Extract all user messages and assistant responses
• Identify tool calls and their results
• Track file modifications and key decisions

Supported External File Formats:

• Markdown headers: ## User / ## Assistant
• Prefix format: User: / Assistant:
• Human/AI format: Human: / AI:
• Numbered format: 1. User: / 2. Assistant:

Step 3: Analyze the Conversation

Run the distillation analysis:

# Default: saves to ./distilled-prompts/
python scripts/distill_prompt.py <session-id-or-path>

# Custom output directory
python scripts/distill_prompt.py <session> --output ~/my-prompts/

# Custom output file path
python scripts/distill_prompt.py <session> --output ~/prompts/my-prompt.md

# Output to stdout
python scripts/distill_prompt.py <session> --stdout

The analyzer identifies:

1. Initial Intent - What the user originally wanted
2. Missing Context - Information that had to be clarified
3. Corrections Made - Where the assistant went wrong
4. Key Decisions - Important choices that shaped the outcome
5. Final Solution - What actually worked

Step 4: Generate One-Shot Prompt

The script outputs a markdown file containing:

# Distilled Prompt: [Task Summary]

## Original Context
- Session: [ID]
- Turns: [N]
- Duration: [Time]

## One-Shot Prompt

[The optimized prompt that includes all necessary context upfront]

## What Made This Work
- [Key insight 1]
- [Key insight 2]

## Anti-Patterns Avoided
- [What went wrong initially]

## Reusability Notes
- [When to use this prompt template]
- [Variables to customize]

Output Format Options:

# Get only the one-shot prompt (no metadata)
python scripts/distill_prompt.py <session> --prompt-only

# Get JSON output for programmatic use
python scripts/distill_prompt.py <session> --json

# Ready-to-use: display prompt + copy to clipboard
python scripts/distill_prompt.py <session> --use

The --use flag is particularly useful for immediately using the distilled prompt:

• Displays the optimized prompt
• Automatically copies to clipboard (xclip, xsel, or pbcopy)
• Provides instructions for starting a new session

Step 5: Review and Refine

Present the distilled prompt to the user. Ask:

• Does this capture the essential requirements?
• Are there domain-specific details to add?
• Should any sections be expanded or condensed?

PLAN Workflow

Step 1: Gather Initial Request

Ask the user to describe their task. Listen for:

• Vague requirements
• Assumed context
• Implicit constraints
• Missing acceptance criteria

Step 2: Generate Clarifying Questions

Run the planning question generator:

python scripts/generate_questions.py --task "<user's task description>"

The script generates questions across categories:

• Scope: What's in/out of bounds?
• Context: What existing code/patterns to follow?
• Constraints: Performance, compatibility, style requirements?
• Acceptance: How will success be measured?
• Edge Cases: What happens when things go wrong?

Step 3: Conduct Planning Interview

Present questions to the user in batches of 3-5. Prioritize:

1. Blocking questions (can't proceed without answers)
2. Scope clarifications
3. Nice-to-have details

After each batch, assess if enough context has been gathered.

Step 4: Synthesize One-Shot Prompt

Once sufficient context is gathered:

python scripts/synthesize_prompt.py --task "<task>" --answers "<qa-pairs>"

Generate a comprehensive prompt that:

• States the objective clearly
• Provides all necessary context
• Specifies constraints and preferences
• Defines acceptance criteria
• Anticipates edge cases

Step 5: Validate the Prompt

Present the synthesized prompt and ask:

• Is anything missing?
• Are priorities correct?
• Should any constraints be relaxed?

LIST Workflow

List OpenCode Sessions

python scripts/list_sessions.py --agent opencode

Default location: ~/.local/share/opencode/storage/

List Claude Code Sessions

python scripts/list_sessions.py --agent claude

Default location: ~/.claude/projects/

List Gemini CLI Sessions

python scripts/list_sessions.py --agent gemini

Default location: ~/.gemini/tmp/<project_hash>/chats/

List Codex CLI Sessions

python scripts/list_sessions.py --agent codex

Default location: ~/.codex/sessions/

List All Agents

python scripts/list_sessions.py --agent all

External Conversation Files

You can also distill prompts from exported conversations or manually created files:

# Parse and analyze a markdown conversation
python scripts/distill_prompt.py conversation.md --use

# Parse a text file export
python scripts/distill_prompt.py chat_export.txt --output ~/prompts/

Supported Formats:

• ## User / ## Assistant - Markdown headers
• User: / Assistant: - Colon-prefixed messages
• Human: / AI: - Alternative role names
• 1. User: / 2. Assistant: - Numbered messages
• >>> User >>> / --- Assistant --- - Block separators

If no format is detected, the entire file is treated as a single user message.

URL Support

Distill prompts directly from conversations hosted online. Large files are automatically downloaded locally:

# Download and distill from URL
python scripts/distill_prompt.py https://example.com/conversation.md --use

# Download from GitHub Gist
python scripts/distill_prompt.py https://gist.githubusercontent.com/user/id/raw/chat.md

# Specify custom download directory
python scripts/distill_prompt.py https://example.com/large-chat.md --save-dir ~/downloads/

# Force re-download (bypass cache)
python scripts/distill_prompt.py https://example.com/chat.md --force

# Set maximum file size (default 50MB)
python scripts/distill_prompt.py https://example.com/huge-chat.md --max-size 100

URL Features:

• Automatic file type detection from URL
• Caches downloaded files in ~/.cache/prompt-distiller/downloads/
• Respects --force flag to re-download
• Configurable max file size limit (default 50MB)
• Works with raw GitHub, Gists, Pastebin, and any direct file URL

Session Data Locations

Override with environment variables:

• OPENCODE_SESSIONS_PATH
• CLAUDE_SESSIONS_PATH
• GEMINI_SESSIONS_PATH
• CODEX_SESSIONS_PATH

Resources

scripts/

distill_prompt.py Options

parse_session.py Options

references/

• prompt-patterns.md - Effective prompt structures and templates
• analysis-criteria.md - How conversations are analyzed