Home

Field SOP

Scenario-based tutorials with real screenshots, prompts, workflow templates and pitfalls.

Swap Your Embedding Model Without Wrecking RAG: The SOP

An embedding model upgrade SOP (basis October 9, 2026; pairs with the site's open-source and comparison pieces): switching embedding models means a full re-embed, and this SOP covers the decision and the drill in one pass. Step 0 audits your current setup: model and dims, context window, MRL usage, deployment form, prompt-prefix conventions - plus the production reminder (relayed) that a reranker often fixes retrieval misses in minutes before you re-embed anything. Five selection criteria: parameters/context/dimensions with MRL/license (HF API verified: EG2 apache-2.0 ungated but no LICENSE file, Qwen3 apache-2.0, BGE-M3 mit)/deployment form. EmbeddingGemma 2's task prompt prefixes are mandatory - queries as task: {...} | query: {...}, documents as title: {...} | text: {...}, seven task types each with a prefix; skip them and accuracy silently drops; encode(truncate_dim=512/256/128) trims vectors for up to 6x storage savings. Five execution steps: dual-run old vs new on sentence-transformers>=6.1.0 (same-prompt spot checks), canary cutover (framework only, no invented percentages), full re-embed with reconciliation and retirement, then an evaluation loop with Ragas/DeepEval linking to the site's RAG evaluation piece. Plain-text repos: github.com/UKPLab/sentence-transformers, huggingface.co/google/embeddinggemma-2.

Claude Code Subagents: A Five-Step SOP for Cheaper AI Coding

A hands-on SOP for Claude Code subagents (basis: October 8, 2026; pairs with the site's Haiku 5.5 hotspot): from the /agents command to a parallel fleet of specialized workers. File locations and priority: .claude/agents/ (project) overrides ~/.claude/agents/ (user); a markdown file with YAML frontmatter is all it takes. Fields covered one by one: name (lowercase-hyphen), description (drives auto-delegation - learn to write Use PROACTIVELY / MUST BE USED trigger phrases), tools (omitted means inherit-all; least privilege says list them explicitly), model (sonnet/opus/haiku/inherit; default sonnet), permissionMode, skills (not inherited from parent), and hooks (PreToolUse/PostToolUse/Stop; once unsupported). Model routing: search/summarize/classify/extract to haiku, multi-file changes to sonnet, architecture to opus (community-sourced per-task cost comparison $0.03/$0.24/$1.20, labeled single-source). Isolation: subagents run in their own context and return only summaries; they cannot spawn subagents (no Task tool); implementation stays with the main agent (Anthropic's own guidance). Haiku 5.5 tie-in: model: haiku picks up the new model, and Claude Code v2.1.292 adds an effort parameter on the Agent tool plus claude plugin install --marketplace (release-note caliber). Four copy-ready skeletons (code-reviewer/security-reviewer/debugger/test-writer) and four failure patterns to avoid.

Ant's Ling-3.1-flash is free for two weeks: how to use it well

A two-week free-trial SOP for Ling-3.1-flash (basis: October 7, 2026). Ant Group's Bailing team shipped it October 6: 560B total parameters, 25B active, a 1M context ceiling, 7 KDA + 1 Gated MLA layers with 512 routed experts (8 plus 1 shared); the previous Ling-3.0-flash was 124B/5.1B. Expectation-setting first: the trial serves only 256K context, well short of the advertised 1M. Three entry routes compared: chat.ant-ling.com web chat with zero setup; Ant Digital's MaaS platform (maas.antdigital.com) for API keys; Vercel AI Gateway for overseas developers (free until October 13 per the changelog), plus Novita AI. What to test in the window: long-document synthesis, office tasks producing Excel/PPT (officially highlighted on GDPval-AA V2.1 and HealthBench Professional - no published scores, none invented), and long coding tasks (the 17h compiler at 178/182 and 20h Rust port at 8.015x, self-reported). Feeding a 256K window in three steps: chunked summarization plus the final section in full, with rolling-window and headroom tips. The API uses an OpenAI-compatible endpoint pattern (base_url and parameters per the MaaS docs - nothing fabricated). Before the window closes: export your data and watch two announcement lines (1M context on the paid tier, planned open-sourcing - no dates given). Compliance: official sign-up routes only; no circumvention or account sharing.

How to Call Claude Sonnet 5.5: Three Clouds, One Price List

A hands-on SOP for calling Claude Sonnet 5.5 (basis: October 5, 2026). It opens with the compliance red line: Anthropic does not directly offer claude.ai signup or API keys to individuals in mainland China, and this guide provides no circumvention methods (no proxies, virtual cards or SMS-verification workarounds) - only three legitimate routes: direct access from officially supported regions (list per Anthropic's live page), enterprise onboarding through the three clouds where the model shipped day one (AWS Bedrock, Google Cloud Vertex AI, Microsoft Azure), and legal alternatives for mainland developers (GLM and Kimi open weights and APIs via ModelScope and official channels). Practical coverage: the model ID claude-sonnet-5-5 (all lowercase with hyphens - the easiest spelling trap), snapshot pricing $2/$10 with $0.20 cache reads (savings come from token efficiency, not price cuts), the zero-data-retention option, the official division of labor versus Opus 5.5 (daily tasks to Sonnet, complex long tasks to Opus), and same-day Claude Code integration. Five steps each with expected outcomes: pick your entry, provision and verify the key, first call, parameters and caching, and the two-model split. All pricing and capability figures are snapshots; the official page governs.

How to Use Kling AI: A 4.0 Flash Guide with 15 References

A hands-on SOP for Kling AI (basis: October 3, 2026). It opens by aligning expectations: full Kling 4.0 is not yet live, 4.0 Flash is limited to a small-scale preview for black-gold annual members, and pricing and credit costs are unpublished so none are invented. The five-step framework: entry and account (klingai.com or the app; verify membership and 4.0 Flash eligibility); run the minimal text-to-video loop (input, generate, review, download) to set your own baseline; an 8,000-token prompt structure in four parts (tone, storyboard, sound, constraints), borrowing the LTX-2 official prompting method (chronological, action-first, concrete cinematography within 200 words - explicitly labeled as LTX-2 guidance, not Kling documentation); a progressive drill for 15 multimodal references (single subject alignment, images plus subject, add a camera-move video, then push toward the 7-subject cap, with each asset's role spelled out); and a continuation workflow that slices content into 30-second blocks up to 2 minutes, paired with timeline-based continuous creation. Steps three through five are explicitly framed as workflows for when 4.0 reaches your account, not as available today.

Self-host the open-source Dots in three commands

A self-hosting SOP for the open-source Dots: install uv (Windows PowerShell and Linux routes), then one uvx command brings it up (--openrouter-key required); open 127.0.0.1:8765 for the chat-left, live-browser-right workbench. Swap models with one --model flag across anything on OpenRouter; four advanced parameters explained in depth - --seed (one identity per seed, reproducible), --profile-dir (logins and cookies persist across runs), --proxy (timezone and language follow the exit) and --help (the full option index); invisible_playwright_mcp hands the same browser to Claude Code, Codex and Gemini CLI as an MCP server; the README's own sample task (Milan-to-Lisbon daily fare check, no-ticket means say so) demonstrates the no-fabricated-numbers prompt discipline. A closing compliance section: respect target-site terms and local law, automate only your own business and accounts you are authorized to use, no fraud/scalping/bulk sign-ups. All commands and parameters come verbatim from the README.

No Ascend Hardware? Three Routes into DeepSeek's Ascend Stack

A hands-on SOP for DeepSeek's Ascend kernel stack along three routes of rising hardware demands: (1) the no-NPU reading route - clone the four repos, diff the ports against the NVIDIA-side originals, study the FlashMLA deep-dive technical report (docs/20260930-ascend-prefill-deep-dive.md), and note CUDA users can stay on the main repos since the fused kernel remains CUDA-only; (2) Ascend deployment - TileKernels installs with pip install tile-kernels (Python 3.12+, PyTorch 2.13+, TileLang 0.1.15+, Ascend 950 NPU + CANN 9.2.0+) or from source via pip install -e ".[dev]", while DeepGEMM-Ascend builds from source (git clone --recursive, ./develop.sh, pip install . --no-build-isolation on CANN 9.20); (3) DeepEP-Ascend for collective communication (git clone plus the DeepJIT submodule, ASCEND_HOME_PATH setup), honestly flagged with its missing LICENSE file and validated-stack boundary (950DT, CANN 9.2.0, torch_npu 2.13.0rc1). Includes a hardware/software requirements table, common pitfalls (scaling-factor packing, even-rank requirement, recursive submodules) and a pre-flight checklist; all commands and versions come verbatim from each repo's README.

Change One Line of base_url: 5M Free Tokens for Your Agent

A hands-on SOP for wiring LongCat-2.5-Preview into your existing agent toolchain along three routes of rising effort: (1) web-first (sign in at longcat.ai for chat, image upload and simple agent tasks); (2) OpenAI-protocol access (point base_url at https://api.longcat.ai/openai with model ID LongCat-2.5-Preview - existing OpenAI SDKs migrate with zero code changes); (3) Anthropic-protocol access for Claude Code (three env vars: ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_MODEL), with Codex, OpenClaw, OpenCode and Kilo Code following the same base_url-plus-model-name swap. Includes usage tips for the 1M context and 128K output, the 5M-free-tokens offer (relayed basis), and Preview-stage caveats (no public benchmarks, unreleased weights, interface may change); validate price, latency and success rate on a small traffic slice before production.

MiMo-V2.6 Integration SOP: Desktop, API and Local Weights

A hands-on SOP for wiring MiMo-V2.6 into your workflow along three routes of rising effort: (1) the MiMo Desktop client (install, sign in or plug in your own API key, switch MiMo-V2.6-Pro/Flash in the model list, UltraSpeed mode, screenshot-feedback iteration); (2) the MiMo open-platform API (create an app for a key, pass model name, messages, tools and multimodal inputs; for agent tasks, wire environment logs, test pass rates, screenshots and verifier feedback into an execute-check-correct loop; validate price, latency and success rate on a small traffic slice before production); (3) local weights (download from the HF collection collections/XiaomiMiMo/mimo-v26; parameter counts are unpublished, so hardware floors defer to the model cards; for RL reproduction, go through the five environment scripts in XiaomiMiMo/verl). Includes a pitfall table and a pre-launch checklist; API pricing defers to the platform documentation.

One Line Change, Half the Cost: GPT-6 Sol/Luna Migration SOP

A hands-on SOP for migrating existing GPT-5.6 calls to GPT-6 Sol/Luna: with OpenAI-compatible access, swapping the model name (gpt-6-sol / gpt-6-luna) will most likely run — but a name swap alone captures none of the migration's other half. You must restructure caching (stable prefix first, explicit breakpoints around the invariant region) to fully capture the 90% discount, and assign reasoning effort per turn (low for simple steps, high for critical ones), leaning on the official basis that mid-conversation re-tiering preserves the cache — save where you can, spend where you must. Includes a regression-comparison workload step, the availability red line that Free/Go accounts can only use Luna, and a decision method for drawing the Luna/Sol boundary by task difficulty. Request fields and effort values defer to the official documentation as the final basis.