萃靈 EssevinWhat deliberate AI users are reading
Model techniques and practical guides, distilled drop by drop.
A Reproducible Blind Test for High-Risk Customer Replies: Publish the 100-Point Rubric Before the Result
A reproducible protocol for comparing AI-written high-risk customer replies using one synthetic incident, identical prompts, anonymous labels, a five-dimension 100-point rubric, counted revision rounds, and published raw outputs. No model result is reported until the runs are completed and documented.
Before an Insurance Meeting: A Five-Step AI Workflow for a Needs Analysis Brief
Turn anonymised policy information and meeting notes into a traceable Needs Analysis Brief, information-gap list, follow-up questions and message draft. This workflow includes an editable prompt and clear boundaries for privacy, product advice and authorised human review.
AI→ 2026-08-28Healthcare Administration: What Can AI Handle in Rosters and Patient Guidance? Draw Three Lines First
AI can organise administrative roster data, identify missing fields, and turn approved patient guidance into a readable draft. It should not diagnose, triage, give treatment or medication advice, decide clinical rostering, or send unreviewed messages to patients.
AI→ 2026-08-26Do Not Ask AI for a Legal Answer First: Use a Risk Map to Prioritise a Long Contract
AI can map clauses, quote supplied text against an existing review lens, and organise questions for a long contract. It cannot determine legal effect or replace legal advice. This workflow keeps confidentiality, traceability, and qualified human review in the process.
AI→ 2026-08-24One Deck, Three Audiences: Build This Message Brief Before Asking AI to Rewrite
Leadership, clients, and delivery teams need different answers from the same campaign deck. Start with a traceable Message Brief, then use AI to create three source-bound outlines without inventing claims.
AI→ 2026-08-20An HR Template Factory: Turn One Job Brief into a JD, Interview Guide, and Scorecard
Start with one verified Job Brief, then use the same competency chain to draft a Job Description, structured interview questions, and an anchored scorecard without losing traceability.
AI→ 2026-08-18Helping with Primary Maths: Make AI Ask Before It Explains
Give AI the school level, full problem, and the child’s attempt, then make it ask one diagnostic question at a time, reveal hints gradually, and verify the method before offering a similar practice problem.
AI→ 2026-08-14The Role + Example Prompt Pattern That Makes AI Outputs More Consistent
“Write professionally” is too vague. Assign a useful working role, provide one or two representative examples, and add an acceptance check so an AI model has a clearer target and safer boundaries.
AI→ 2026-08-12Before Renewing Your Broadband, Use AI to Prepare the Conversation
The hardest part of a broadband renewal call is tracking the contract term, total cost, bundled services and your own limits at the same time. This guide provides a copy-ready AI prompt, a 90-second opener, four response branches and a comparison table.
AI→ 2026-08-11What Should You Ask AI First? Ten Openers That Work Immediately
A blank AI chat does not require prompt-engineering jargon. Ten copy-ready opening questions cover explanation, sorting, comparison, planning and verification, so the model can produce a useful first draft instead of guessing your intent.
AI→ 2026-08-10Why AI Sometimes Miscounts the r's in “strawberry”
“strawberry” contains three r's, yet an AI model may occasionally give a confident wrong count. The reason is not a single mysterious bug: tokenisation, character-level operations and language-model prediction are different jobs. This guide shows three ways to verify the answer.
AI→ 2026-08-06Claude Code /resume: Two Ways to Pick Up Yesterday's Work
Continue a half-finished travel itinerary, CV revision or product comparison without explaining everything again. This guide shows when to use /resume, --continue, /compact, /clear and session names, with practical everyday examples and the limits of session recovery.
AI→ 2026-08-04Five MCP Servers Ordinary People Can Actually Use
MCP is not only for programmers. These five popular GitHub projects can help with live web research, everyday browser work, mixed document formats, WhatsApp history, and Notion. This guide explains the useful hook, setup difficulty, and privacy trade-off behind each one.
AI→ 2026-07-30Stop Saying "Write Me a...": The Four Ingredients of a Prompt That Works
The same AI can produce brilliant work for one person and generic filler for the next. The difference is rarely the model — it's whether your prompt specifies a format. Four ingredients, three before/after rewrites, and a fill-in template to turn "write me a..." into an instruction the AI can actually execute.
AI→ 2026-07-29Claude Code's Plan Mode: Review the Plan Before It Touches Anything
The most expensive AI mistake isn't one wrong step — it's ten wrong steps executed in a row before you notice. Plan Mode makes Claude hand you a plan first and act only after you approve it. Here's how to turn it on and off, what your four approval options actually are, and why it's just as useful outside of coding.
AI→ 2026-07-28Meeting Minutes in Twenty Minutes: A Four-Step AI Workflow for Admin Staff
The slow part of writing minutes is not typing — it is restructuring a scattered discussion into decisions and owners. This guide gives admin and office staff a repeatable four-step workflow with copy-paste prompts, cutting a two-hour task to a ten-minute draft plus ten minutes of checking, and names the three kinds of meeting you should not hand to AI.
AI→ 2026-07-27What a Context Window Actually Is — and Why a Bigger One Doesn't Mean a Better Memory
When an AI forgets what you told it an hour ago, that is not a bug — the desk is full. This guide explains what a context window is, why token counts matter, why million-token windows still lose information in the middle, and five habits that immediately improve long conversations.
AI→ 2026-07-27Set Up Claude Code from Hong Kong or China: The Complete Four-Step Path
Anthropic does not serve Hong Kong or China, but you do not need a VPN or any workaround to run Claude Code. This guide covers the full path — install, register, set two environment variables, verify — in four steps, no coding background required.
Claude Code→ 2026-07-23The Resume Emergency Room: One CV, Three Tailored Versions in Half an Hour
Recruiters spend an average of 7.4 seconds on the first scan of a resume. Instead of mass-sending one CV to thirty jobs, tailor it to three. This guide gives you four copy-paste AI prompts — decode the job ad, map your gaps, rewrite with facts intact, then scan it through a hiring manager's eyes.
AI→ 2026-07-22Gemini 3.6 Flash: The Workhorse Gets Cheaper — and Hong Kong Can Use It Today
Google shipped three models on July 21 — Gemini 3.6 Flash cuts output pricing from $9 to $7.50 per million tokens and uses 17% fewer tokens for the same work, while the flagship 3.5 Pro is still missing. What the release actually means for users in Hong Kong and China.
AI→ 2026-07-21The July 2026 LLM Tier List: From GOAT to Cooked
GPT-5.6, Grok 4.5, Muse Spark 1.1 and Kimi K3 all shipped within a month, while Gemini 3.5 Pro missed its third deadline. We rank the mainstream frontier models of July 2026 on capability, price and real-world availability — and unpack the four most common mistakes people make when reading leaderboards.
LLM→ 2026-07-20Why AI Makes Things Up — and Six Ways to Catch It
AI hallucination is not a bug — it is the personality of a technology built to predict the next word. This guide explains why large models fabricate facts, which six question types are highest-risk, and six defenses you can apply today.
AI→ 2026-07-20An Open-Source Model Just Beat the Paid Flagship — What That Means for You
Moonshot AI's open-source Kimi K3 hit #1 on the Frontend Code Arena within hours of release, overtaking Claude Fable 5. Here is what the win actually measures, where leaderboards mislead, and how to pick models by task now that open and closed models are this close.
AI→ 2026-07-20Fable 5 Joins Max Plans at Half the Limits: Top-Tier AI Pricing Is Shifting to Pay-As-You-Go
Anthropic announced that from July 20, Claude Fable 5 will be included in Max and Team Premium plans — at 50% of standard limits, while Pro users stay on usage credits. Here is the 39-day timeline, why flat-rate plans cannot absorb frontier models, three common misreadings, and what this means for users in Hong Kong and China.
AI→ 2026-07-20Why Claude Code Gets Expensive — and the Three Commands That Fix It
Most Claude Code cost and "memory loss" problems come down to context management. Here is how the context window actually bills you, and when to use /context, /compact, and /clear.
AI→ 2026-07-16Claude Code Setup in China: Complete Guide — from Registration to First Autocompletion in 10 Minutes
No VPN needed, no code changes — just two environment variables to connect Claude Code to a stable gateway. This guide covers registration, getting your API key, configuration, and verification, plus the three most common pitfalls.
Tutorial→ 2026-07-12GPT-5.6 vs Claude Sonnet 5 for Chinese Project Code: Which Should You Choose?
We route both flagship coding models at scale through our production gateway. Here's what we actually observe about completion quality, instruction following, and token consumption — plus our real-world selection advice.
Review→ 2026-07-08How Much Can Prompt Cache Save You? Mechanics, Real-World Hit Rates, and Three Writing Techniques
Cache hit rate determines your bill. Here's how prompt caching actually works, the real hit rates we see in conversational tools, and three prompt organization techniques that can double your cache efficiency.
Practice→ 2026-07-03Drop PDFs and Excel Files Straight Into Your Model: The Right Way to Use Attachments in AI Chat
Spreadsheets, scanned PDFs, raw email, web links — which ones go straight to the model, and which need preprocessing? Here's everything about AI Chat's attachment capabilities and how to use them effectively.
Product→ 2026-06-28Seedance 2.0 Video Generation API: Quick Start Guide
From submitting your job to downloading the final video: request parameters, task polling, video retrieval, and how resolution, duration, and pricing line up.
Tutorial→ 2026-06-24Routing Multiple Upstreams with One API Key: Group-Based Routing Can Cut Costs in Half
Choose your tier per task: stable group during peak hours, economy group for batch jobs. Here's how groups, multipliers, and automatic failover work together, plus cost-optimized setups for three typical user profiles.
Practice→