Home
Scenario-based tutorials with real screenshots, prompts, workflow templates and pitfalls.
A six-step SOP for cutting LLM API costs: quantify baseline, prefix caching (Qwen3.8-Max charges ¥1/M on cache hit vs ¥12 on miss, a 12x gap), model routing (expensive models for hard tasks, cheap for easy), batch API (latency for discount), context trimming with dedup, and monitoring. Each step ships with Python code, pitfalls, and FAQ.
DeepSeek's prefix cache makes long-session costs bifurcate (0.02 yuan hit vs 1 yuan miss, a 50x gap), but general agents do not optimize for DeepSeek's cache. This SOP walks through DeepSeek-Reasonix (community-built, unofficial) end-to-end: npm install, reasonix setup, reasonix.toml config sketch, cache-aware maintenance, executor+planner dual-model, and MCP plugins. All commands from the README; config fields per official docs.
An n8n workflow for knowledge base auto-update: content source trigger -> fetch content -> dedupe -> LLM summary + keywords + classification -> vectorize and write to KB -> notify. Keeps your docs library current. Includes a .json template and advanced tips (full rescan, staleness detection, permissions).
An n8n workflow for invoice processing automation: email/upload trigger -> extract attachment -> OCR/LLM field extraction -> validate -> IF pass write to finance system else human review -> archive. Includes compliance notes (sensitive financial data, recommend local OCR + private LLM, per local tax regulations). Includes a .json template download.
A three-level product management prompt pack: beginner user stories, intermediate structured PRD (9 sections), expert roadmap and requirement breakdown (Epic->Feature->Story + RICE/MoSCoW). Includes a competitor cheatsheet. Anti-fabrication constraints embedded throughout.
A three-level research/literature review prompt pack: beginner single-paper extraction (6 elements), intermediate multi-paper comparison (table + consensus/divergence), expert full review (structure + gap analysis + citation suggestions, with a hard verify-all-citations-against-originals guardrail). Includes a cheatsheet prompt and anti-fabrication guardrails.
An AI agent evaluation and benchmarking SOP: four metric categories (success rate/efficiency/safety/robustness), building an eval set (typical/edge/adversarial 60:25:15, programmatic scoring first), multi-run sampling with baseline comparison, and latency/cost logging. References the site's V4-Flash hands-on harness method.
A spec-driven development SOP: before coding, capture requirements as a spec (problem/goal/boundary/acceptance), then confirm in chunks, generate an implementation plan, break into tasks, implement with verification, and accept against the spec. Includes a spec template and github/spec-kit integration.
An n8n workflow for bug triage automation: Webhook receives bug report -> extract fields -> LLM classification (severity P0-P3 + module + repro) -> IF P0/P1 notify on-call and create high-priority issue, else route by module -> auto-reply to reporter. Includes a .json template.
An n8n workflow for AI content calendar automation: scheduled topic collection -> LLM scoring + publish-date + category -> write to calendar (Feishu/Notion) -> notify owner near publish date -> status writeback. Includes a .json template.