Home
Scenario-based tutorials with real screenshots, prompts, workflow templates and pitfalls.
A hands-on SOP for hooking DeepSeek-V4-Pro into Claude Code: two core env vars (ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN) + the Cline custom-base-URL route. Five steps: install, get key, set vars, run, verify. Includes official-verified commands, 5 pitfalls (AUTH_TOKEN ≠ API_KEY, compat ≠ equivalence, thinking billed, price-hike warning, Windows session vars), 5 FAQs.
An AI music generation workflow SOP with two routes: Suno/ElevenLabs (fast SaaS) + ACE-Step (free local). Five steps: requirements and lyrics, choose tool, generate, post-mixing, integrate (video/podcast/content). Includes runnable commands, 5 pitfalls (commercial rights, lyrics copyright, quality, vocals, format), 5 FAQs.
An AI image-to-3D workflow SOP with a two-route decision: img2threejs code-based (open-source free, rule-based objects) vs Meshy/Tripo mesh (SaaS, complex organic shapes). Five steps: prepare assets, choose route, generate, QA and fix, export and integrate. Includes runnable commands, 5 pitfalls, 5 FAQs.
An open-source video editing SOP: a full Shotcut multi-track + AI-assisted (subtitles, smart transitions, grading) workflow in five steps -- organize footage, rough multi-track cut, AI subtitles/transitions, color grading and keyframes, export and distribute. Includes runnable commands, 5 pitfalls, and 5 FAQs; a zero-cost Premiere replacement.
A context-engineering SOP for AI coding agents (Claude Code/ZCode/Crush): structure context in five steps -- map project context, write rule files (CLAUDE.md/AGENTS.md), manage the context window, persist memory, and verify/iterate. Includes a real rule-file example, 5 pitfalls, and 5 FAQs.
A hands-on SOP for building a Computer Use Agent: sandbox setup, wiring the Anthropic Computer Use API, the main screenshot-decide-act loop, safety guardrails, and tuning -- five steps. Includes runnable Python (ast.parse-verified), 5 pitfalls, and 5 FAQs. Since "can operate a computer" equals "can cause harm", it emphasizes sandbox isolation, action allowlists, and human confirmation for sensitive operations.
Stack three layers to build a zero-cost multimodal pipeline in 2026: Agnes's free OpenAI-compatible API (one key for text/image/video), Chinese pay-as-you-go models (DeepSeek/Qwen/Jimeng/Kling), and local open-weight models (Wan2.1/Hunyuan/CogVideo). Includes a five-step SOP, runnable Python with async video polling and local fallback, plus five pitfalls.
A five-step red teaming SOP using free open-source tools (PyRIT/promptfoo/DeepTeam), covering six agentic attack surfaces (prompt injection, jailbreak chains, data leakage, tool misuse, privilege escalation, persistence) with runnable YAML and Python examples and a reusable system prompt guardrail template.
Function calling / tool use is the core capability of AI agents. This SOP gives developers a copyable tool-calling implementation template: define tool schema -> model decides which tool to call -> app executes -> feed result back into the next turn. Covers OpenAI (function calling / Responses API vs Chat Completions, strict mode) and Anthropic tool use per current official APIs, with five pitfalls (loose schema / parameter hallucination / unhandled call failures / no timeout / security sandbox) and five FAQs. APIs per official docs.
Hands-on setup for mem0 (62K stars) and Letta (24K stars, formerly MemGPT): from pip install to a working memory store, with an extraction prompt template, a mem0 vs Letta vs pure RAG comparison table, a when-to-use long-term memory vs RAG vs context window decision framework, and 5 real pitfalls. mem0's new algorithm lifted its LongMemEval score from 67.8 to 94.4.