Maintained guides.

  1. Guide · T3 Code · checked

    Run each agent thread in its own worktree, then ship it as a pull request

    T3 Code drives the coding agents already signed in on your machine. Each thread can work in its own Git worktree, and the diff pane reviews a turn's own changes before a one-button commit, push and pull request.

    On the dossiers of T3 Code, Claude Code, Codex

  2. Guide · AI Agents for Beginners · checked

    Coordinate specialized agents: group chat, handoffs and visibility

    Lesson 8 of AI Agents for Beginners compares ways to coordinate specialized agents and shows how to log and measure their interactions.

    On the dossiers of AI Agents for Beginners

  3. Guide · Awesome LLM Apps · checked

    Answer support questions from crawled docs with a voice reply

    A Streamlit app indexes documentation collected with Firecrawl into Qdrant and answers typed questions with text and synthesized speech.

    On the dossiers of Awesome LLM Apps, Firecrawl, Qdrant, Streamlit

  4. Guide · Anthropic · checked

    Let an orchestrator split a task across worker prompts

    The orchestrator-workers notebook has a model break a task into typed subtasks at runtime and call a worker prompt for each.

    On the dossiers of Building effective agents examples

  5. Guide · Anthropic · checked

    Run a computer-use agent against a desktop in Docker

    The Computer Use Demo runs an agent loop beside a Linux desktop in one container and shows the chat and the desktop in a local web page.

    On the dossiers of Claude quickstarts

  6. Guide · Vercel Labs · checked

    Turn an issue into a reviewed draft pull request

    The Foreman template triages a GitHub, Linear or HTTP task, writes acceptance criteria, implements in a sandbox and has an independent station review the diff before a person merges.

    On the dossiers of Foreman software factory

  7. Guide · Google · checked

    Stream a research report grounded in Google Search

    A Colab notebook enables the Google Search tool on the Gemini API, streams the report and renders the search suggestions it used.

    On the dossiers of Gemini cookbook

  8. Guide · GitHub · checked

    Carry a feature from specification to implementation with a coding agent

    Spec Kit starts from project principles and moves a feature through specification, plan, tasks and implementation using a supported coding agent.

    On the dossiers of GitHub Spec Kit

  9. Guide · GPT Researcher · checked

    Plan subquestions, gather sources and write a cited report

    GPT Researcher plans research questions, sends agents to gather web and local material, and combines the source-tracked findings into a report.

    On the dossiers of GPT Researcher

  10. Guide · gstack · checked

    Give a coding agent planning, review, browser QA and release roles

    gstack installs role-based skills so a coding agent runs separate design and engineering reviews, a reporting-only browser QA pass and a release step.

    On the dossiers of gstack, Claude Code, Codex

  11. Guide · LangChain · checked

    Research with a local model and web search

    A LangGraph loop looks for gaps in its running summary, searches again for a set number of rounds and keeps the cited summary with its sources.

    On the dossiers of Local Deep Researcher, Ollama, LangGraph

  12. Guide · OpenAI · checked

    Have a lead agent call specialist agents as tools

    A portfolio manager agent keeps control, calls specialist agents and hosted tools, and writes a research memo with a trace of the run.

    On the dossiers of OpenAI Cookbook, OpenAI Agents SDK

  13. Guide · OpenClaw community · checked

    Collect, score and deliver a scheduled tech news digest

    A community recipe combines feeds, accounts, releases and search, scores items for recency and corroboration, and delivers the digest by chat or email.

    On the dossiers of OpenClaw community use cases, OpenClaw

  14. Guide · n8n · checked

    Prototype local AI workflows with n8n, Ollama and Qdrant

    Docker Compose runs n8n with PostgreSQL, Ollama for inference and Qdrant for vectors, and imports a sample workflow.

    On the dossiers of Self-hosted AI starter kit, n8n, Ollama, Qdrant

  15. Guide · STORM · checked

    Draft a cited article from a multi-perspective outline

    STORM surveys perspectives on a topic, simulates questions to a source-grounded expert, then drafts an article with citations that still needs editing.

    On the dossiers of STORM

Provider guides.

  1. Guide · OpenAI · official documentation · checked

    Prompt caching

    Put stable developer instructions, tools and shared reference material first and per-request details last, so the rendered prefix matches exactly. Track usage.input_tokens_details.cached_tokens against total input tokens to see the hit rate.

    On the dossier of OpenAI

  2. Guide · OpenAI · official documentation · checked

    Batch API

    Use Batch for evaluations and other work that can wait. It costs 50% less than synchronous calls and has a 24-hour processing window. Match results to requests by custom_id.

    On the dossier of OpenAI

  3. Guide · OpenAI · official documentation · checked

    Latency optimization

    Generate fewer tokens: ask for only the detail and structured fields the app uses, and lower verbosity or reasoning effort for routine work. Combine sequential calls when one request has what it needs, and run independent calls in parallel.

    On the dossier of OpenAI

  4. Guide · Anthropic · official documentation · checked

    Prompt caching

    Mark the end of the stable prefix (tools, then system, then messages) with cache_control. Cache reads cost 0.1 times base input, lower on some models, and five-minute writes 1.25 times; check cache_read_input_tokens and cache_creation_input_tokens in usage.

    On the dossier of Anthropic

  5. Guide · Google · official documentation · checked

    Context caching

    Implicit caching is on by default for Gemini 2.5 and newer. Put large, common content at the start of the prompt, send requests with a similar prefix close together, and read usage.total_cached_tokens to confirm hits.

    On the dossier of Google