Service specific prompts

Prompt chains for GitHub, YouTube, website Q&A, and similar services: context packing, output format, and failure behavior.

The repository organizes prompts, services, and FastAPI routers per domain. Prompts define the instruction templates and chains; services orchestrate ingestion, context assembly, and LLM invocation; routers expose HTTP endpoints for each service.

graph TB
subgraph "Routers"
R_GH["routers/github.py"]
R_WEB["routers/website.py"]
R_YT["routers/youtube.py"]
R_REACT["routers/react_agent.py"]
R_BROWSER["routers/browser_use.py"]
end
subgraph "Services"
S_GH["services/github_service.py"]
S_WEB["services/website_service.py"]
S_YT["services/youtube_service.py"]
S_REACT["services/react_agent_service.py"]
S_BROWSER["services/browser_use_service.py"]
end
subgraph "Prompts"
P_GH["prompts/github.py"]
P_WEB["prompts/website.py"]
P_YT["prompts/youtube.py"]
P_REACT["prompts/react.py"]
P_BROWSER["prompts/browser_use.py"]
end
R_GH --> S_GH --> P_GH
R_WEB --> S_WEB --> P_WEB
R_YT --> S_YT --> P_YT
R_REACT --> S_REACT --> P_REACT
R_BROWSER --> S_BROWSER --> P_BROWSER

• GitHub prompt system: Builds a chain that merges repository summary, file tree, and content with a user question and optional chat history. It enforces strict reliance on provided context and uses Markdown formatting.
• Website prompt system: Merges server-fetched and client-rendered contexts, prioritizing client-side content for dynamic and authenticated views. Provides guidelines for summaries, structure, links/media, metadata, and formatting.
• YouTube prompt system: Assembles a transcript-based context (with error handling) and applies a focused schema for summaries, duration, stats, themes, sentiment, and recommendations.
• React agent prompt: Supplies a tool-use instruction template for agents that rely on tool availability and structured reasoning.
• Browser automation prompt: Defines a precise JSON action plan for Chrome extension automation, including DOM manipulation and tab/window control actions, with strong constraints and examples.

The system follows a consistent pattern:

• Routers validate and parse requests, then delegate to services.
• Services ingest domain-specific data (repository, website content, YouTube transcripts), assemble context, and invoke prompt chains.
• Prompts define the instruction template and optional context assembly via Runnable chains.
• Responses are normalized and returned through routers.

sequenceDiagram
participant Client as "Client"
participant Router as "FastAPI Router"
participant Service as "Domain Service"
participant Prompt as "Prompt Chain"
participant LLM as "LLM Client"
Client->>Router : "POST /{service}"
Router->>Service : "generate_answer(...)"
Service->>Service : "Ingest domain data"
Service->>Prompt : "invoke({context, question, chat_history})"
Prompt->>LLM : "format + forward to model"
LLM-->>Prompt : "raw response"
Prompt-->>Service : "formatted answer"
Service-->>Router : "answer"
Router-->>Client : "HTTP 200 with answer"

GitHub integration prompts#

• Prompt template: System role, repository summary, file tree, relevant content, optional chat history, explicit guidelines, and formatting rules.
• Input variables: tree, summary, content, question, chat_history.
• Chain composition: RunnableParallel extracts inputs; PromptTemplate formats; LLM client executes; StrOutputParser returns text.
• Parameter handling: Router validates presence of url and question; service passes chat_history as a string; optional attached file triggers direct GenAI SDK usage.
• Response formatting: Markdown with code blocks and bullet lists; strict adherence to provided context.
• Error handling: Dedicated error messages for invalid URLs, accessibility, token limit exceeded, and general failures.
• Optimization: Truncates repository content when using GenAI SDK to fit payload limits; preserves concise summaries and trees.

flowchart TD
Start(["Invoke GitHub Chain"]) --> Build["Build LLM Chain"]
Build --> Inputs["Assemble {tree, summary, content, question, chat_history}"]
Inputs --> Run["Run PromptTemplate + LLM"]
Run --> Parse["Parse to String"]
Parse --> Return["Return Answer"]
Inputs --> |Error| Handle["Handle Exceptions<br/>Return Plain Message"]

Website analysis prompts#

• Prompt template: Dual-context guidance (server-fetched vs client-rendered), explicit preference for client context, and detailed guidelines for summaries, structure, links/media, metadata, tables, and math formatting.
• Input variables: server_context, client_context, question, chat_history.
• Chain composition: RunnableParallel builds inputs; PromptTemplate + LLM client; StrOutputParser.
• Parameter handling: Router validates url and question; service converts client HTML to markdown; chat_history assembled as a string; optional attached file bypasses chain and uses GenAI SDK.
• Response formatting: Plain markdown with bullet points, tables, and LaTeX for math.
• Error handling: Generalized error message on processing failure; client context fallback when unavailable.

sequenceDiagram
participant Router as "Website Router"
participant Service as "Website Service"
participant Tools as "Markdown Fetchers"
participant Prompt as "Website Prompt Chain"
participant LLM as "LLM Client"
Router->>Service : "generate_answer(url, question, chat_history, client_html)"
Service->>Tools : "Fetch server markdown"
Service->>Tools : "Convert client HTML to markdown"
Service->>Prompt : "invoke({server_context, client_context, question, chat_history})"
Prompt->>LLM : "format + forward"
LLM-->>Prompt : "raw response"
Prompt-->>Service : "formatted answer"
Service-->>Router : "answer"

YouTube processing prompts#

• Prompt template: Focused assistant role for YouTube video questions, guidelines for summary, duration, stats, themes, sentiment, and recommendations, and strict scope limitations.
• Input variables: context (transcript), question, chat_history.
• Chain composition: RunnableParallel with a get_context function that fetches and cleans subtitles/transcripts; PromptTemplate + LLM client; StrOutputParser.
• Parameter handling: Router validates url and question; service optionally uploads attached file via GenAI SDK; chat_history passed as string.
• Response formatting: Plain markdown with bullet points, tables, and LaTeX.
• Error handling: Known error detection for transcript retrieval; returns empty context when errors occur; service wraps exceptions and returns plain messages.

flowchart TD
Start(["YouTube Chain"]) --> GetCtx["get_context(url)"]
GetCtx --> Clean{"Transcript Available?"}
Clean --> |Yes| UseCtx["Use Cleaned Transcript"]
Clean --> |No| EmptyCtx["Empty Context"]
UseCtx --> Prompt["PromptTemplate + LLM"]
EmptyCtx --> Prompt
Prompt --> Parse["StrOutputParser"]
Parse --> Return["Return Answer"]

React agent prompt#

• Template: Instructional prompt for agents that use tools, framing the question, listing available tools, and instructing to use tools to gather information.
• Purpose: Pairs with the broader React agent orchestration (covered elsewhere) as the tool-use instruction.

Browser automation prompt#

• Template: Detailed instruction for generating JSON action plans for Chrome extension automation. Includes DOM manipulation and tab/window control actions, selector guidance, search URL construction, and strict output rules.
• Chain composition: ChatPromptTemplate piped to LLM; response parsed via StrOutputParser.
• Parameter handling: Router constructs a user prompt from goal, target URL, DOM structure, and constraints; service sanitizes and validates JSON action plan.

sequenceDiagram
participant Router as "Browser Use Router"
participant Service as "AgentService"
participant Prompt as "SCRIPT_PROMPT"
participant LLM as "LLM Client"
Router->>Service : "generate_script(goal, target_url, dom_structure, constraints)"
Service->>Prompt : "format prompt with DOM info"
Prompt->>LLM : "forward to model"
LLM-->>Prompt : "raw response"
Prompt-->>Service : "JSON action plan"
Service-->>Router : "validation + sanitized result"

• Routers depend on models and services to handle requests and responses.
• Services depend on prompts for chain composition and on domain tools for context ingestion.
• Prompts depend on the LLM client abstraction and LangChain components for formatting and parsing.

graph LR
R_GH["routers/github.py"] --> S_GH["services/github_service.py"]
S_GH --> P_GH["prompts/github.py"]
R_WEB["routers/website.py"] --> S_WEB["services/website_service.py"]
S_WEB --> P_WEB["prompts/website.py"]
R_YT["routers/youtube.py"] --> S_YT["services/youtube_service.py"]
S_YT --> P_YT["prompts/youtube.py"]
R_REACT["routers/react_agent.py"] --> S_REACT["services/react_agent_service.py"]
S_REACT --> P_REACT["prompts/react.py"]
R_BROWSER["routers/browser_use.py"] --> S_BROWSER["services/browser_use_service.py"]
S_BROWSER --> P_BROWSER["prompts/browser_use.py"]

• Context size management:
• GitHub service truncates repository content when using GenAI SDK to avoid payload limits.
• Website service assembles concise server and client contexts; client context is preferred but fallback is supported.
• YouTube service cleans transcripts and falls back to empty context on known errors.
• Token limits and retries:
• GitHub service detects token limit exceeded and suggests narrowing the scope.
• Browser automation limits DOM preview entries to reduce token usage.
• Streaming and latency:
• Chains are synchronous in current implementation; consider streaming responses at routers/services for long-running prompts.
• Model selection:
• Services support passing llm_options to prompt builders for model tuning.

• GitHub
• Invalid repository URL: Returns guidance to use the repository root.
• Access issues (404/clone): Requests public repository verification.
• Token limit exceeded: Advises focusing on specific files/directories.
• Website
• General processing error: Returns a friendly message to retry.
• Client HTML missing: Falls back to server context; client context is optional.
• YouTube
• Transcript retrieval errors: Known error messages detected and handled ; returns empty context.
• LLM invocation errors: Wrapped with a plain message.
• Browser automation
• JSON validation failures: Validation returns problems; endpoint returns structured error response.
• React agent
• Attached file upload failures: Logs and returns a plain message; otherwise normal operation.

Ground every answer in fetched context, demand the format you parse, and fail loudly when context is empty. Shared patterns beat one-off prompt files that drift.