Home

Field SOP

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

One Missing BOS, Every Output Breaks: A 280B Base Setup SOP

Take ERNIE 5.5 Flash Base from pre-download compliance checks to a first successful run: start with three questions from LICENSE v1.0 (territory excluding the EU, UK and South Korea, the 50 million dollar related annual revenue threshold, the Powered by ERNIE notice and one-time notice to Baidu_ERNIE@baidu.com, separate prior written authorization above the threshold, plus the distillation ban and withheld training data); choose between the FP8 and BF16 checkpoints; run the official Transformers quick start with add_special_tokens=True so the BOS token is present; serve the base model through the completions endpoint with the official SGLang or vLLM examples (--tp 4, 131072, up to 1048576) and verify with curl using prompt rather than messages; then find the continued-training entry point in NVIDIA-NeMo/Megatron-Bridge under examples/models/ernie55/. Three traps: a missing BOS, treating a Base checkpoint as a chat model, and misreading territory or revenue conditions.

17 Seconds on a CPU: Prove Which Model Answered You

Verify the cryptographic birth certificate of an AI reply on a plain CPU with the open-source FlashProve from Anthropic: prepare Linux (g++ 13 or later, non-root, under 2GB RAM; WSL2 as the common Windows route), clone anthropics/FlashProve (0.16.0), build with verifier/build.sh in one to two minutes, run the stored example under examples/accepted-proof (expect a final line YES, exit code 0, under 2 seconds), read results via the exit-code table (0 accepted, 1 rejected, 5 unchecked, 14 invalid call), then the real-world five-argument call and the rule that the commitment ID must come from a trusted third party. Honest limits throughout: it proves provenance not correctness, a 1024-token context, officially not maintained, and no major provider ships proofs yet.

Weights Downloaded, Now What? A VRAM Planning Guide

Between "weights downloaded" and "model running" sits a VRAM bill: weight memory equals parameters times bytes per parameter (BF16/FP8/INT4), and per-token KV cache equals 2 x layers x KV heads x head_dim x bytes. Worked examples span Qwen3-32B (66GB down to 17GB across precisions) and K3.1 Pro (5.6TB BF16), plus three misjudgments, precision-by-GPU guidance (24GB/96GB/Blackwell), vLLM and Ollama toolchains, and a five-step landing checklist.

Wire Hy4 In With Three Params: Five Steps and Three Traps

A five-step guide to wiring the final Tencent Hunyuan Hy4 API into your own workflow: create a console key and enable the model (a 402 means it is not enabled), change only base URL, API key and model (two regional endpoints, three protocol paths), keep the platform name hy4 distinct from the OpenRouter name tencent/hy4, feed images with the documented image_url and base64 pattern, then plug into CodeBuddy, Claude Code and similar clients and verify. Includes runnable Python, the 401002/402xxx/429xxx error codes, SSE mid-stream error frames, thinking mode, and three separate pricing regimes.

Build a Game Without Code: A 49-Agent AI Studio

Turn a single Claude Code session into a 49-agent game studio with Claude Code Game Studios (github.com/Donchitos/Claude-Code-Game-Studios, MIT, about 33,837 stars): three-tier directors/leads/specialists, 72 slash commands, and 12 hooks that catch hardcoded values and accidental pushes, with a protocol that keeps every decision in your hands. Includes a five-minute setup, a demo-trimmed workflow, third-party API options, and a fit checklist.

Solo AI Content Pipeline: Seven AI Employees Run the Whole Line

This breaks a solo content company into seven stages - discover, screen, test, document SOP, generate content, productize, and sales - and marks five as automatable by agents and two as must-do-yourself. It names agents like MarketRadar, WorkflowArchitect, and BrandGuardian, maps seven QA AI roles, and lands on a five-step build (write a spec, set a human checkpoint, ship one before scaling), with the reminder that an agent is not a co-founder and the overseas reference cost is about 38 dollars a month.

Still Writing Weekly Reports at 8 PM Friday?

A five-step workflow that turns weekly fragments into a report worth submitting: ten minutes of Friday triage into three piles (progress, risks, support needed), structured input, generation, manual number-checking, and polish. Two copy-paste prompt templates are the core deliverables: one turns a running log into a draft (role setup, four-part structure, the anti-hallucination line "never write what I did not mention, mark uncertainty as to-be-confirmed," an AI-flavor blacklist, and a hard 500-word cap); the other polishes into executive register (harsh-reviewer self-check, exemplar alignment, and a "do not add or alter any numbers" guard). Anti-hallucination is not paranoia: user reports describe unconstrained models adding growth figures never in the source and computing their own summary values, convincingly enough to fool the author. The data-red-line section is the hardest edge: Nantong police's "three don'ts, three dos," a Chaoyang court ruling (the dismissal was ruled unlawful and the employee won — not a license to upload), and Samsung's 1,024-byte leak precedent, landing on weekly-report specifics: redact client names, amounts, and internal system names before pasting. Methodology: conclusion first, results up front, data speaks; leaders read three things — progress, risks, support needed; surfacing problems earns credit. Tools (a weekly-report generator, Feishu's AI personal weekly report, DingTalk, WPS AI) get one paragraph, not a comparison.

Invoice OCR to Excel: Batch Recognition Without Manual Entry

A two-route SOP for turning a stack of invoices into one Excel sheet (division of labor with the site's deepseek-ocr local-deploy piece: that one gets the model running, this one gets batches structured and exported). Route A, Baidu AI Cloud's VAT invoice recognition: the vat_invoice endpoint accepts pdf_file directly, with free quotas of 1,000 calls/month for individuals and 2,000 for enterprises issued monthly, failed calls included in the count, and 2 QPS free / 10 QPS paid (official figures; the community-circulated 5/3 QPS is not adopted). Route B, local PaddleOCR (v3.7.0 with PP-OCRv6, 91,929 stars, Apache-2.0, CPU-runnable), with the key caveat in bold: the open-source edition has no invoice-specific model, so field mapping means writing your own regexes — that engineering gap is the core selection argument between the routes. The full pipeline runs as written: pathlib/glob batch collection, PyMuPDF rendering at 300dpi (with an AGPL/commercial dual-license warning and alternatives), per-file recognition into dicts, and styled xlsx export via pandas/openpyxl. Error playbook covers blur, stamp occlusion, skew, wrong ticket types and QPS throttling; authenticity checks go only through the national VAT invoice verification platform (5 checks per invoice per day, invoices from the last 5 years). Tencent/Alibaba comparisons and a large-model route (GLM-OCR) section both carry explicit sourcing limits, with no marketing accuracy numbers.

Transcribe 13-Minute Audio in 59 Seconds, Fully Local

A five-step local speech-to-text SOP (faster-whisper as the main engine, repo github.com/SYSTRAN/faster-whisper): pick by VRAM (24GB card large-v3+float16+beam 5; 8GB card int8_float16, step down to medium if tight; CPU-only small/base+int8), install (GPU needs CUDA 12 cuBLAS plus cuDNN 9, older stacks pin ctranslate2 3.24.0/4.4.0), a minimal transcription script, three speed-and-VRAM levers (fp16 to int8 saves about 1.6GB, batch 8 costs about 1.5GB for roughly 3.7x speedup, beam comes third), and advanced steps (Silero VAD, word timestamps, hand-rolled SRT, WhisperX diarization). One benchmark framing is stated everywhere: large-v2, beam_size=5, 13 minutes of audio on an RTX 3070 Ti 8GB — the original implementation takes 2m23s/4708MB, faster-whisper int8 59s/2926MB, batched 16s/4500MB; on CPU the piece honestly notes whisper.cpp fp32 (2m05s/1049MB) beats faster-whisper fp32, with no sweeping claims. The pitfall list includes never using pure int8 on GPU (CTranslate2 docs: compute capability 7.0/6.1+ required, older cards silently fall back to float32). A single closing paragraph points to NVIDIA's Nemotron 3.5 Transcribe as the new-engine option, leaving the benchmarks to the open-source column piece.