A field guide · 2026 edition
Two years of notes on using AI, gathered for friends drowning in information who haven't found the door yet — one no-detour path, plus the tools, thoughts, and pitfalls along the way.
Start the trailNo matching notes — try another word.
Field notes · scroll sideways · share to X
Antigravity finally stopped being a chore. /teamwork-preview has a coordinator split the job across specialist agents, /boost checks everything before it acts, and Pro/Ultra quotas went up (check your own account for the numbers). It's crossed from toy to something that can carry light daily work.
Live voice can stay on while you work, but it's a separate model (GPT-Live-1), not the main one answering. I use it to hand out tasks and keep a running work log, not to finish hard jobs in one shot. And Desktop Work voice draws on its own quota — don't open it unless it's actually driving the computer.
Codex sessions can hand work to each other now: one passed its deploy plan to another, which picked it up and paused its own release to avoid a collision. Running several tasks at once, I'm hoping to stop being the one relaying messages between them.
First lesson from GPT-6: watch the tokens. Ultra and multi-agent runs aren't my default — a few agents going at once burns through budget fast. Start on a lower reasoning tier and step up only when the task earns it. Size the effort to the job.
Two weeks into Claude Code, the biggest realization: people who write tests instantly pulled ahead of people who don't. Once the agent finishes, you have exactly one job left — verify. Verification runs on tests.
Prompt engineering isn't about magic words, it's about stating the task clearly. People who can write a good prompt for Claude get better at product specs, emails, and technical docs too.
Local models aren't about saving money — they're about privacy and determinism. Ollama plus a 70B Q4 model handles 60% of daily questions; only the other 40% needs the cloud.
The beginner's biggest mistake: treating AI as a search engine. It isn't. It's a biased, forgetful, but fluent colleague. Learning to work with a colleague like that is the core skill of this era.
Gemini 2.5 Pro's 2M context — I dropped a whole 800-page contract in and asked, and it genuinely remembered a clause on page 30. Five years ago this was science fiction.
MCP is underrated as a protocol. Once agents can call any tool in a standardized way, the moat at the model layer gets hollowed out further — what's left is the moat of 'tools + data'.
Plain-English basics · explained like you're five
The smallest unit of text an AI reads and writes. Like snipping a sentence into little pieces (a word, half a word) — the AI "thinks" one piece at a time, and each piece is a token.
A program that has read enormous amounts of text and learned to "guess the next word." You start it off, it keeps writing. GPT, Claude, and Gemini are all LLMs.
The "socket" that lets two pieces of software talk. Your program sends a request in an agreed format; the other side sends the result back in kind. Plugging into an AI means calling its API.
A "universal plug" standard for connecting AI to tools. With it, an AI can reach your calendar, files, and databases the same way — no separate connector built for each one.
A pre-written "how to do this one thing" manual + materials for the AI. When the task comes up it opens the matching skill and knows the steps and standards. (This page was built by an AI following a skill.)
An AI that breaks work into steps, calls tools, and carries a whole task through on its own — not just replying, but actually getting the job done.
The rhythm of AI at work: think a step → reach for a tool → look at the result → think the next step, circling until the job's done. That loop is an agent's "heartbeat."
The full running record of what an AI did step by step — what it thought, which tool it called, what it got back. When something breaks, you retrace it, like a dashcam.
Turns a piece of text into a string of number-coordinates; the closer in meaning, the closer the coordinates. AI uses it to judge how alike two passages are, and to pull the most relevant bits out of a big pile of material.
Before answering, the AI pulls relevant passages from your own library and answers holding those — so it speaks from your material instead of making things up from memory (uses the embedding above to find them).
Figuring out how to say things to an AI clearly, so it gives you what you want. Same task, different phrasing, wildly different results — that "how to ask" craft is this.
One step past "how to ask": managing the whole packet of information you feed the AI each time — what to include, in what order, how much, what to cut. Most of whether it answers well comes down to this.
Building a whole "shell" around the model — tools, the loop above, permissions, how context is managed — turning a model that only chats into an agent that actually does the work. That shell is the harness.
Reads written text out loud in a near-human voice. Audiobooks, turn-by-turn navigation, Siri speaking — that's it.
The reverse — turns what you say into text. Dictating to your phone, auto meeting transcripts, voice input — all of it.
The site where programmers store code, edit together, and see every change's history. Like Google Docs + cloud storage + revision history, for code.
The "skeleton" language for building web pages — it says what headings, paragraphs, images, and buttons a page has. This very page is HTML.
The six steps · click a node to expand
01
Mental model
It isn't looking up answers on Google for you. It's more like someone who's very good at continuing a sentence — you start it off, it guesses "what word most likely comes next," and keeps going. Get this, and the rest stops being confusing.
02
Prompt basics
Four things are enough: ① the context ② who it should act as ③ the format you want ④ an example. Nail those four before chasing tricks.
03
Pick one
Don't shop around. Pick one (GPT or Claude or Gemini), use it honestly for three straight months, then talk about switching.
04
Integrate
Put it on real work: writing email, editing docs, reading PDFs, reviewing code. Make it a tool you reach for daily, not an occasional toy.
05
Ship one
Build something very small — under 100 lines is fine. Finishing one with your own hands jumps your understanding a whole level.
06
Agent layer
Once the first five feel easy, go advanced: let AI do the work itself (agents), connect tools (MCP), run models on your own machine. Don't rush this floor before the foundation is solid.
Pick one. Use it for three months. Then reassess.
Claude Opus 5.5My pick
Out Sept 22 · my main pick · Fable 5.1-level work at 40% less than Opus 5 to run
Best for · Writing, code, long documents, agents — make it the daily driver, and save Fable for the problem that's genuinely hard
API $4 / $20 per million tokens (in/out), 20% below Opus 5 · on Pro, Max, Team and Enterprise plansGPT-6 family
Astra on top · Sol the workhorse · Luna the fast, cheap one · and 6.1 Sol landed Sept 29
Best for · Astra for the hardest coding, science and computer-use work; Sol for daily agents and code; Luna for bulk routine jobs
API: Sol $2 / $10, Luna $0.10 / $0.50 per million tokens · Plus and up get them in Work and Codex; Luna is on the desktop app for Free and GoGemini 3.8 Flash
Built for speed · coding and agents are the focus · Google ecosystem
Best for · Speed, giant documents, Google users, devs who want AI Studio for free
Free (incl. AI Studio) + paid tiersGrok 4.7
Out Sept 21 · xAI's new flagship · 500K context · same price as 4.6
Best for · Long-running agent work, coding and knowledge tasks where you also want it cheaper than peers
API $2 / $6 per million under 200K prompt tokens, $4 / $12 above (in/out) · SuperGrok from $30/moDeepSeek V4
Open-weight · the app itself is free · general + reasoning merged into one
Best for · Free everyday chat, or self-hosting an open model to cut costs
App/web totally free · API off-peak $0.007/$0.22/$0.66 per million tokens, 5M free tokens for new accountsCodex · Claude Code · Antigravity · Grok Build
OpenAI's official coding agent. Runs in the cloud — give it a repo and a task, it opens a PR itself.
Verdict · Good for: batched, well-defined changes, writing tests. Not for: exploratory work where you think as you go.
Anthropic's terminal coding agent. Runs in your local shell — reads files, runs commands, runs tests.
Verdict · The strongest local coding agent right now — if you're willing to work in the terminal.
Google's agent-native IDE — multiple agents work tasks in parallel, with a built-in verification flow.
Verdict · New, worth watching. The idea is "IDE for agents," not "IDE with AI."
xAI's coding agent. Runs in the terminal, up to 8 parallel agents, working a plan → search → build loop.
Verdict · Local-first — your code never touches xAI's servers. Arena Mode auto-scores and ranks candidate solutions before you even look.
Other agent frameworks
Hermes Agent — An open-source autonomous agent from Nous Research — self-hosted, runs on your own server 24/7 against whatever model you point it at. Finishes a hard task, writes itself a reusable skill, and gets sharper the more you use it.
DeepSeek Harness — DeepSeek's open-source agent runtime, built "everything is a plugin" — model, tools, and interface are all swappable. Its own definition of an agent is Model + Harness — the exact split behind the "harness engineering" entry above.
Grok Bot — xAI's always-on AI teammates — each bot gets its own cloud computer, signs into your existing tools, and keeps grinding through multi-step work even after you've stepped away.
Gemini Spark — Google's 24/7 personal agent. Runs on Gemini 3.5 and is built on the same Antigravity harness listed under Coding tools above, wired straight into Gmail / Docs / Slides.
Git / collaboration terms
Git — A tool that snapshots your code and records every change, so you can jump back to any old version anytime.
repo — Repository: a folder holding all of a project's code + its history.
commit — Saving a checkpoint: recording what you changed this time with a one-line note. Like a save point in a game.
branch — A side road off the main line to change things, merged back once it works.
fork — Copy someone's whole project under your own name — change it freely without touching the original.
PR — Pull Request: the "I've made changes, please look and decide whether to merge into the main line" request.
Copy, fork, replace placeholders
You're my writing partner. I have a rough idea: "{IDEA}".
Please:
1. Distill its core thesis (one sentence)
2. Give three possible angles, each argued in a paragraph
3. Once I pick an angle, give a 5-part outline Below is code I just wrote. Play a senior engineer:
1. Point out the 3 most worthwhile issues (no more than 3)
2. Give fixes, noting the trade-offs
3. Don't chase "perfect" — chase "maintainable"
```
{CODE}
``` Explain "{TOPIC}" to me Feynman-style:
1. First in a way my 12-year-old could follow
2. Then deepen it with an everyday analogy
3. Finally, point out one common misconception I'm stuck on a decision: "{DECISION}".
Analyze it with the 10/10/10 framework:
1. Will I regret it 10 minutes from now
2. 10 months from now
3. 10 years from now
Two sentences max each. No pep talk. Small things, shipped — not planned
A vendor-neutral Claude skill: teaches Ontario buyers/sellers how to judge, interview, and background-check a real-estate agent — and never recommends a specific one (me included). Open source, cloneable.
Claude Skill · TRESA/RECO · MIT
A small local agent that scans my inbox daily, sorts mail into 5 buckets, and labels it. Saved me about 4 hours in two weeks.
Claude API · Python · Gmail API
Feeds all my contract PDFs into a local vector store for offline Q&A. No internet — a lawyer friend gave it 5/5.
Ollama · Llama 3.1 70B · LangChain
Reads git commits and Slack channels, auto-drafts a report every Friday. I only edit 20%.
Claude Code · GitHub API · Slack
Exposes my home Synology NAS to Claude so it can read my family-photo metadata and organize albums.
TypeScript · MCP SDK
Cloud · Local LLM · Frameworks · Evals
No affiliates · no fluff
Free apps worth knowing
What a realtor can hand to an AI agent
Full auto = the agent runs it, you look at the result; AI drafts · you review = the agent writes, you go line by line — it goes out under your name; Prep automated · you execute = the agent assembles everything, but the signature, the price and the send button stay with you. Not because the machine can't reach them — because that step shouldn't be handed over.
房源文案
Give it the specs and selling points, get listing descriptions, social captions, and email blasts (EN + ZH).
How · Claude / GPT on your copy template, run in batches.
营销设计
Listing posters, social graphics, just-sold cards and video covers, rendered in bulk from your brand template — several versions to pick from.
How · An image model plus a locked brand template (colours, fonts, credentials line), run in batches.
双语沟通
Translate and polish client messages, emails, and documents between English and Chinese, keeping tone and professionalism.
How · Any large model; use Gemini / Claude long-context for big files.
社区资料
Compile schools, amenities, commute, and sale history into a one-page client brief.
How · A web-search agent on a uniform template.
CRM 录入
Structure contact info from business cards, forms, and emails into your CRM, and tag it.
How · Agent + CRM API / Zapier, on a schedule.
Lead 跟进
Draft first replies and nurture sequences, personalized by client stage.
How · CRM trigger + LLM draft, you glance before sending.
多平台宣传
One piece of content reworked for each platform — Xiaohongshu, WeChat, Instagram, YouTube — with the posting schedule laid out.
How · The agent drafts and schedules; you give it a look and confirm the send.
定价草稿
Pull comparables into a first-pass CMA draft. You must check the numbers and the comp choices.
How · Data feed + LLM summary, human sets the price.
预约协调
Read intent from email, propose slots, write to the calendar; conflicts go to you.
How · Calendar MCP + agent.
合同条款初审
Key clauses pulled out, watch-points listed, every clause lined up against the template — all assembled by the agent. Whether to change or sign is your call and your lawyer's.
How · Long-context model plus a clause library builds the comparison; a human decides clause by clause.
合规核对
The checklist and supporting documents for the RECO / OREA / FINTRAC process, assembled with gaps flagged up front. The final tick is yours.
How · The agent prepares the list and the paperwork; you tick and you sign.
AI is a biased, forgetful, but fluent colleague — learning to work with a colleague like that is the core skill of this era.
— Arthur · 2026.09.29
12 years full-time in real estate · Broker · FRI · SRS · ABR · MCNE · CLHMS & GUILD Elite · REAIS