For a phone assistant, separate the voice layer from backend tools: speech and interruptions belong to voice; research, calendars and bookings belong to tools. Compare Grok, GPT-Live and Gemini Live by the first useful response, task completion and the cost of the whole call, rather than a latency ranking or input price alone.
Published Skool postA field guide · 2026 edition
From prompt to agent,
a clear path.
An evolving collection of AI notes for friends overwhelmed by information and looking for a place to start: a learning path, tools, observations and pitfalls from prompt to agent.
Start the trailNo matching notes — try another word.
Field notes
Field notes · scroll sideways · share to X
Notes retain their historical context; see the model cards for current versions and prices. Where a post shows only a relative date, the synchronization date is labelled.
AI Agent Evolution Atlas · 2026
14 tracks, 87 milestones and 34 highlights on one timeline. Explore Computer Use, Voice Agent, MCP, Remote Control and Personal Agent.
Explore the atlasAI output should help people understand: use clear writing, diagrams, interactive HTML or an explanatory video to suit the task. I use diagrams for client processes, HTML for market analysis and video for teaching. Choose the format around what the audience needs to understand.
Published Skool postREA combines decompilation with observing a running app to help understand its client-side features, and can work with Claude Code, Codex and Cursor. It can inform your own implementation; the other app's server-side code and logic remain out of view.
Published Skool postThat day my GPT allowance returned to 100%; recently I had also seen resets after using about half. This was an observation from my own account, not a promise for everyone. Check the allowance and reset time shown in your own account.
Published Skool postMy setup starts with the workload: a Mac with at least 16GB of memory, considering 32GB for local video or 3D work, plus external SSD storage when needed. Add subscriptions as usage earns them. Check current official prices and allowances before choosing a plan.
Published Skool postFor flyers and listing graphics, I value editable elements: changing dates, room sizes, QR codes or a headshot separately saves repeated full-image generation. In my experience, Claude Design with templates suits revisions and batches, while I use GPT image generation more for visual ideas.
Published Skool postI use Claude for coding, design and automation, and GPT more for Live voice and WhatsApp workflows. Rewind, project memory and recorded skills reduce repeated explanations. Before changing GPT conversations, I prepare a context handoff. The division follows my own working habits.
Published Skool postI use Gemini for quick everyday questions and Gems for scheduled YouTube summaries. For the first build of a website or app, I start with Antigravity and set up the workflow and branding together. Matching tools to their strengths suits my day-to-day work.
Published Skool postAntigravity 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.
Published Skool postLive 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.
Published Skool postCodex 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.
Published Skool postFirst 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.
Published Skool postWhen Claude Code writes code, acceptance checks matter: review requirements, permissions and security as well as 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 offer more control over privacy and deployment. Whether they are capable enough depends on the model, hardware and task; test them.
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 offers about one million tokens of input context. Retrieve sections and locate clauses in long contracts, then check page references and meaning against the original.
Official sourcesMCP 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'.
A few words first
Plain-English basics · explained like you're five
Token
A small piece of text a model processes: a character, a word or part of a word. Token counts are not word counts.
LLMLarge Language Model
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.
API
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.
MCPModel Context Protocol
An open protocol connecting AI applications with external data and tools. It standardizes the connection; services and access still need configuration and authorization.
Skills
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.)
Agent
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.
loopagent loop
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."
trace
A record of model calls, tool inputs and outputs, timing and errors from a run, useful for debugging. It is not the model’s complete internal reasoning, and logs may contain sensitive data.
embedding
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.
RAGRetrieval-Augmented Generation
Retrieve relevant material before asking the model to answer. Retrieval can use vectors, keywords or both. It can reduce unsupported answers, but errors remain possible: check the sources.
prompt engineering
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.
context engineering
Manage the whole information package given to a model: which documents, in what order, how much to include and what to leave out. The relevance and quality of context affect results.
harness engineering
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.
TTSText-to-Speech
Reads written text out loud in a near-human voice. Audiobooks, turn-by-turn navigation, Siri speaking — that's it.
STTSpeech-to-Text
The reverse — turns what you say into text. Dictating to your phone, auto meeting transcripts, voice input — all of it.
GitHub
The site where programmers store code, edit together, and see every change's history. Like Google Docs + cloud storage + revision history, for code.
HTMLHyperText Markup Language
The "skeleton" language for building web pages — it says what headings, paragraphs, images, and buttons a page has. This very page is HTML.
Official sources · MCP ↗Tracing ↗Retrieval ↗
Start here · the beginner's path
The six steps · click a node to expand
Times are approximate learning plans. Adjust them to your own pace.
01
Understand what an LLM actually does
Mental model
Models generate content from context and do not guarantee factual accuracy. Search and connected tools can also retrieve current information. Distinguish generation, web retrieval and tool execution.
02
Learn to write a prompt
Prompt basics
Explain the goal, background, constraints and output format. Add examples when useful and agree how to check the result.
03
Pick one main model
Pick one
Start with one main tool and build judgment through real tasks. Compare alternatives when you encounter a clear capability or cost limit.
04
Wire it into your workflow
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 a tiny project
Ship one
Choose a small project with a result you can check, such as a public-information summary or a simple webpage. Review the result, permissions and cost.
06
Move up to agents / tooling
Agent layer
Once the first five steps are familiar, explore agents, MCP and local models for more complex work. Start with low-risk tasks, give tools minimal permissions, and retain checks and a way to stop.
Claude · GPT · Gemini · Grok · DeepSeek
Choose for the task · compare quality, speed and cost
Models and API prices checked 2026-10-17. Subscription fees, usage allowances and API billing are separate; consult the official sources for current details.
Claude Opus 5.5Arthur’s personal pick
Released September 22 · Coding, agents and knowledge work
Strengths
- Anthropic says performance approaches Fable 5.1 on most work, with faster output and clearer communication.
Weakness
- Some sensitive tasks trigger safety restrictions or routing to another model.
- Sonnet 5.5 and Haiku 5.5 are also available; choose for the task and budget.
Best for · Complex programming, long documents and tasks with multiple steps.
Standard API per million tokens: US$4 input / US$20 output. Caching and Fast mode have separate prices. Checked 2026-10-17.Official sources · Opus 5.5 ↗API pricing ↗Sonnet 5.5 ↗Haiku 5.5 ↗
GPT-6 family
Astra flagship · 6.1 Sol balances capability and cost · Luna for frequent tasks
Strengths
- OpenAI positions Astra for its hardest work, 6.1 Sol for complex programming, computer use and professional tasks, and Luna for frequent work such as summaries and extraction.
Weakness
- Compare quality, speed and total cost on your actual tasks.
- Sol, Luna and 6.1 Sol are available through Work, Codex and the API. Check current product documentation for availability in ordinary Chat.
Best for · Astra for difficult work, 6.1 Sol for complex development and professional tasks, and Luna for frequent routine work, balanced against speed and budget.
Standard API input/output per million tokens: 6.1 Sol US$2/10; Luna US$0.10/0.50. See official pricing for other models, caching and faster service tiers. Checked 2026-10-17.Official sources · GPT-6.1 Sol ↗Sol & Luna ↗Model directory ↗
Gemini 4 Argon / 3.8 Flash
Argon for complex tasks · Flash balances speed and cost
Strengths
- Argon launched October 13 for complex programming, research and agents with multiple steps.
- 3.8 Flash supports a one-million-token context for long documents and everyday development.
Weakness
- The Argon API has no free tier; 3.8 Flash has a limited free tier.
- Check the applicable data terms before uploading client information.
Best for · Google ecosystem, multimodal and long-document work; choose Argon or Flash for the task.
Standard API input/output per million tokens: Argon US$2/10 (promotion through 2026-11-30); 3.8 Flash US$0.75/3.75 (through 2026-12-31). See official full and post-promotion prices. Checked 2026-10-17.Official sources · Gemini 4 Argon ↗API pricing ↗Gemini 3.8 ↗
Grok 4.7
Released September 21 · 500,000-token context
Strengths
- Built for coding, knowledge work and agents that run for extended periods; supports tools, structured output and four reasoning-effort levels.
Weakness
- Accepts text and image input and produces text output.
- Check long-context and service-tier prices separately.
Best for · Trying extended coding and knowledge-work tasks, with costs compared against actual results.
Standard API input/output starts at US$2/6 per million tokens. Requests above 200,000 tokens have separate long-context pricing. Checked 2026-10-17; see official prices.Official sources · Grok 4.7 ↗Release notes ↗
DeepSeek V4.1 Flash / V4 Pro
Flash is now V4.1; the Pro API remains available
Strengths
- V4.1 Flash supports native vision, tool calls, and thinking and non-thinking modes.
- Both API models offer a one-million-token context; Pro does not accept visual input.
- The old deepseek-v4-flash name routes to V4.1 Flash. Use deepseek-flash for new integrations.
Weakness
- API prices vary by peak/off-peak hours and caching. Check weights, licences and hardware for each version before self-hosting.
Best for · Developers prioritizing API cost, long context and tool calling.
V4.1 Flash API off-peak per million tokens: US$0.15 input / US$0.60 output. Cache hits have separate pricing; peak prices double. Checked 2026-10-17; see official full prices.Official sources · API pricing ↗Updates ↗
AI coding tools
Codex · Claude Code · Antigravity · Grok Build
Codex (OpenAI)
agentic Desktop / CLI / IDE / CloudOpenAI’s coding agent. Read and edit code, run tests and help prepare pull requests from the desktop, terminal, IDE or cloud.
Verdict · Useful for interactive exploration and background tasks with clear boundaries. Define acceptance criteria, then review the diff and test results.
Official sources · Codex ↗
Claude Code
agentic Terminal / IDE / Desktop / WebAnthropic’s coding agent, available in the terminal, IDE, desktop and browser, with file editing, commands and tests.
Verdict · Choose the interface that fits your work. Review permissions, changes and test results for both local projects and cloud tasks.
Official sources · Claude Code ↗
Antigravity
ide App / CLI / IDE / SDKGoogle’s agent development tools offer a desktop app, CLI, IDE extension and SDK, with multiple agents and interactive or background tasks.
Verdict · Choose the entry point for the task, then inspect code, tests and action records for important results.
Official sources · Antigravity ↗
Grok Build
agentic TerminalA coding agent with plan review, parallel subagents, Skills and MCP. It can read and write code and run tests in the terminal.
Verdict · Terminal actions run locally, but model requests still go to servers. Check /privacy, retention settings and sensitive repository data. Local execution does not mean code stays on the computer.
Official sources · Grok Build ↗Settings ↗Data security ↗
Other agent tools and frameworks
Hermes Agent — An open-source agent from Nous Research that can run in your own environment with your choice of model. It supports creating and updating reusable skills from task experience. Official sources ↗
DeepSeek Harness — DeepSeek’s open-source agent runtime follows an “everything is a plugin” architecture, with replaceable models, tools and interfaces. Model + Harness connects to the harness-engineering concept above. Official sources ↗
Grok Bot — xAI’s cloud agent gives each bot a cloud computer, tool connections and the ability to continue tasks with multiple steps. Check official plans, connection permissions and data settings. Official sources ↗
Gemini Spark — Google’s cloud personal agent uses the Antigravity harness and integrates with Workspace tools. It can continue tasks in the background; check official regional and subscription availability. Official sources ↗
Git / collaboration terms
Git — A tool for recording code snapshots and changes, so you can view and restore committed versions.
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 — Create your own copy of a project and change it without affecting upstream. Use and distribution must still comply with its licence.
PR — Pull Request: the "I've made changes, please look and decide whether to merge into the main line" request.
Prompt library
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 Review this code as a senior engineer:
1. Identify up to three evidenced issues worth fixing; say if there are none
2. Suggest changes and explain the tradeoffs
3. Prioritize maintainability over perfection
4. Explain how to reproduce or verify each issue; do not claim unrun tests passed
```
{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 am considering this decision: “{DECISION}”.
Use the 10/10/10 framework:
Describe possible effects after 10 minutes, 10 months and 10 years. State assumptions, uncertainties and information I still need to confirm.
Use no more than two sentences per time horizon; avoid motivational filler. Projects and reusable examples
Practical ideas · examples are labelled
Choose-an-Agent SkillExample
Example: package interview questions, document checks and comparison criteria for Ontario buyers or sellers choosing a real-estate agent into a reusable skill. The user makes the final judgment.
Claude Skill · TRESA/RECO
Email auto-sorter AgentExample
Read an authorized mailbox, classify messages by rules and suggest labels. Confirm permission before writing changes. Email content may leave the computer when a cloud model is used.
Claude API · Python · Gmail API
Local PDF Q&A toolExample
Retrieve from PDFs with a local model and vector store and include page references. Call it offline only after confirming embeddings, inference and logs all stay local.
Ollama · Local LLM · Local vector store
Weekly-report generatorExample
Summarize authorized Git commits and Slack messages into a weekly-report draft, then verify it before sending.
Claude Code · GitHub API · Slack
Home-built MCP serverExample
Organize authorized NAS file metadata through a read-only MCP service. Limit directory access and obtain separate authorization for writes or deletions.
TypeScript · MCP SDK
Going deeper with agents
Cloud · Local LLM · Frameworks · Evals
Cloud services
- AI SDK — One interface for multiple model providers, suited to TypeScript apps
- Replicate / fal — Hosted model inference, billed by model and usage
- Modal / Beam — give your agent an execution environment in the cloud
Local LLMs
- Ollama — run a local model in one command, best for beginners
- LM Studio — a GUI, for people who dislike the terminal
- vLLM — High-throughput model inference and serving
- Hardware needs depend on model size, quantization, context and concurrency. Start with a small model. LM Studio recommends 16GB+ RAM; an 8GB Mac needs smaller models and shorter contexts.
Connection protocols and agent frameworks
- MCP — An open protocol connecting AI apps to external data and tools; services and permissions still need configuration
- LangGraph — state-machine agents, high controllability
- Mastra — TypeScript-first, great DX
Evals / monitoring
- Langfuse — open-source LLM observability
- Braintrust — eval sets + experiment tracking
- Prepare representative tests and failure cases, then iterate on the agent. Record cost, latency and errors.
Official sources · vLLM ↗LM Studio ↗
Resources
No affiliates · no fluff
Free apps worth knowing
Official sources · Cloudflare pricing ↗Gemini API terms ↗
Real-estate tasks you can hand to an agent
What a realtor can hand to an AI agent
AI can prepare research, drafts and proposed actions. The agent remains responsible for checking listing facts, client privacy, advertising and professional judgments. Sending, publishing and important system writes follow prior authorization.
Listing copy
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. Verify facts such as floor area, taxes and listing status before publication.
Marketing design
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. Check actual property features, licensed materials and required identifiers.
Bilingual comms
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. Check contracts, amounts and deadlines item by item.
Neighbourhood briefs
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. Include sources and dates; confirm school areas, boundaries and eligibility with official sources.
Data entry
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. Limit data sources, fields and write access; deduplicate and retain a change record.
Lead nurture
Lead 跟进
Draft first replies and nurture sequences, personalized by client stage.
How · CRM trigger + LLM draft, you glance before sending.
Multi-platform promotion
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.
Pricing / CMA draft
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.
Scheduling
预约协调
Read intent from email, propose slots, write to the calendar; conflicts go to you.
How · Calendar MCP + agent.
Contract review
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.
Compliance checklist
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.
Official sources · RECO advertising requirements ↗
Want to build your own AI real-estate agent? I wrote a complete hands-on tutorial from scratch — honest about where it stops, and why you still need a real person. Read the tutorial →AI is a biased, forgetful, but fluent colleague — learning to work with a colleague like that is the core skill of this era.
— Arthur · 2026.10.17
12 years full-time in real estate · Broker · FRI · SRS · ABR · MCNE · CLHMS & GUILD Elite · REAIS