Jonathan Guy

Nailing the handoff is just the first step to building reliable, automated AI systems. If you want to stop repeating yourself and get more out of your models, grab the rest of my tips and tricks here: pointwake.com/llm-tips

Hint: your LLM did not forget your project. You just did not leave it a handoff. If you use ChatGPT, Claude, or Gemini for real project work, you have probably felt this: You come back the next day, open the same folder, pick up the same project... and suddenly it feels like the whole conversation is gone. The context is weaker. The momentum is gone. And now you are wasting time re-explaining everything. Here is the trick I use before I log off: “Create a fully detailed handoff document from this chat so any person could step in tomorrow and pick up exactly where we left off. Include what we accomplished, the end goal, key decisions, open items, and enough context to understand everything we worked on. Format it so I can save it as a PDF.” That one step changes everything. Now you have: a clean project memory a restart point for tomorrow a document you can reuse with any LLM and a way to keep momentum instead of starting over Most people use AI like a chat. The better move is to use it like a teammate that needs a good end-of-day handoff. That is where a lot of time gets saved. #ChatGPT #ClaudeAI #GeminiAI #LLM #AIWorkflow #Productivity #AIAutomation #PromptEngineering #KnowledgeWork #FutureOfWork

@zephyr_z9 For service-biz operators unit storage doesn't matter. Budget for the layer holding SOPs and call history does. Agents that forget yesterday lose the second-visit advantage. Storage is cheap, the system that uses it isn't.

Btw, agentic AI & CPU boom also increases HDD TAM ;)

@Pirat_Nation Service-biz version is smaller and arrives sooner. One agent books the HVAC tech, another confirms availability and pricing. Trust scales with constrained transactions, not open marketplaces. The ping-pong spree skips the hard part.

Anthropic tested a marketplace for agent-on-agent commerce, giving 69 employees $100 each and let AI agents do all the buying and selling for them. The agents, powered by Claude, first asked each person what items they wanted to buy or sell, then they ran a private marketplace without any human help. They posted listings, made offers, negotiated prices, and closed 186 real deals totaling over $4,000, stronger AI models got better prices and more deals. One agent even bought 19 ping-pong balls for its owner as a self-gift. It could make life easier, but I wouldnt trust about AI models playing with my money.

@petergostev Voice deployments hit a cliff here. 5 extra seconds of reasoning on a homeowner intake call reads as the agent dying. Low slop is bad, medium latency is worse. The winning tier is whichever stays under 2s without dropping the thread.

GPT-5.5 by Reasoning Effort: I've asked it in Codex to create a physics-based visualisation of RL cycles for different sized models (70b, 1t, 10t), to demonstrate how the amount of RL you can do differs by model size. My assessment of each: - Low: weird slop - Medium: kinda cooked - High: sort of tried but ultimately incoherent - Extra High: elite - really nice idea and well executed Obviously this is just one shot, but worth trying different reasoning levels for the new models, medium seems to be pretty good for GPT-5.5 and it was really bad for many previous GPT models.

@vivoplt Operator answer differs. Service-biz teams running voice agents don't switch on token efficiency. They switch on which model holds the conversation through interruptions. Whichever lab nails interruption handling first wins that tier.

Are people switching from Claude Code to Codex just because of token efficiency, or is there more to it?

@thdxr Same status game further out. SMB owners build 12-node n8n stacks when a $20/mo off-the-shelf tool would have shipped 6 weeks ago. Work that ships moves the revenue number. Work that gets posted about usually doesn't.

you used to spend a day messing with your neovim config, feel self conscious, then get back to work now people are spending weeks on some hyper customized coding agent workflow that definitely is worse than vanilla but they can talk about it like they're ahead of the game

@Angaisb_ Cadence reads as exciting from outside. After you ship a voice intake agent, swapping models every 6 weeks is eval cost and tone drift on real customers. Service-biz operators pay for stability, not leaderboard pins.

New big models in the past 3 months: OpenAI: - GPT-5.3-Codex - GPT-5.4 (and Pro) - GPT-5.5 (and Pro) Anthropic: - Opus 4.6 - Sonnet 4.6 - Opus 4.7 Google: - Gemini 3.1 Pro (and Deep Think) It's been more than 2 months Google, please release Gemini 3.5

@ravihanda Same pattern shows up further down market. Service-biz operators aren't replacing staff with AI. They're recovering hours lost to dropped follow-ups and missed after-hours calls. Reclaimed labor, not displaced labor.

Techie in Bangalore saves 15 lakhs a month. Let me repeat - A MONTH!!! And you said AI is taking away jobs.

@VraserX For service-biz operators the model-to-agent jump only matters when the agent can connect to ServiceTitan or Jobber on day one. Faster GPT releases don't change that. The constraint is integration tax, not model capability.

Mark my words: OpenAI will release GPT-5.6 within 4 to 6 weeks. Progress is compounding now. Anyone using the current Codex can feel it. The jump from model to agent is happening in real time.

@NielsRogge Terminal-Bench is one harness shape. The harness that matters for service-biz operators is whether the agent closes a Friday 8pm intake without a human nudge. Claude Code wasn't built for that, and that leaderboard doesn't capture it.

FYI Claude Code is mostly a vibe-coded product (as they say, 100% written by Claude) It's the worst harness for Opus 4.6 among ANY harness on Terminal-Bench 2

@tpschmidt_ Native Claude in AWS doesn't move HVAC owners directly. They never touch IAM. But ServiceTitan and Jobber just got a cleaner path to ship audited agents inside the apps owners already use. Margin shifts to the integrators.

AWS just announced Claude Platform on AWS. No Bedrock required!! You get Anthropic's native Claude experience directly in your AWS account! - IAM handles access - Billing is consolidated - CloudTrail logs everything alongside your other services

What actually closes deals when AI handles it: missed-call text in <30s, lead scoring on every form fill, two-way text booking, reminder + reschedule cadence, review request after the job, voice agent for after-hours. Boring. Quiet. Compounding. None of it will trend on X.

78% of buyers go with the first business that responds. Average service-biz call-back time is 47 hours. An auto-SMS in 30 seconds isn't a tech flex. It's the highest-ROI use of AI any service business can run. The fun AI is downstream of the boring AI.

Most service businesses run AI on the wrong half of the operation. The fun half (image-gen, blog posts, AI decks) gets the spotlight. The boring half (intake, scheduling, follow-up cadence) gets nothing. Customers never see the fun half. They live in the boring half.

@Cryptic_Web3 @naval @OpenAI @AnthropicAI @xai Frontier labs are where the headlines are. Picks-and-shovels for SMBs are where the moats form. The applied layer between an HVAC owner and Claude is Jobber, ServiceTitan, GoHighLevel. Less glamour, more durable returns.

🚨BIG: @naval has launched USVC, a venture fund with a $500 minimum offering access to top AI startups. The fund opens early stage exposure to companies like @OpenAI, @AnthropicAI, and @xai, lowering the barrier to venture investing.

@EXM7777 NotebookLM shines when the operator already has the corpus. Most service-biz owners don't. The unlock is an agent that builds the SOP corpus from what already happened on calls and emails. Then NotebookLM has something to chew on.

everyone is sleeping on NotebookLM... this thing is literally the best option for: - building a custom AI trained on whatever topic you need (content creation, copywriting, automations...) - learning complex topics, studying at any level - understanding RAG, working with large files from any source plus it's powered by a model that was clearly designed for these use cases

@ZssBecker Frontier commoditization shifts margin to deployment, not labs. Local roofers don't care which lab wins. They care that an after-hours intake agent runs at 4 cents per qualified lead. Chinese pricing only accelerates that handoff.

@georgeorch Demo converts the already-shopping. Non-shoppers need a peer story with a number. Owner who hears the HVAC shop two towns over recovered 31% of dropped follow-ups in 90 days acts on it. Same agent shown live, cold, doesn't move them.

I’m slowly realizing something uncomfortable. The “just show them the demo” moment I kept hoping for after shipping a real, working agent system is extremely rare. I thought seeing what actually works would make people drop the hype and start building. I was wrong.

@StockSavvyShay Same thesis a tier down. ServiceTitan and Housecall Pro aren't getting replaced by an LLM. They become the routing and oversight surface the agents plug into. The SMB SaaS that survives is the one where the audit trail lives.

$NOW CEO Bill McDermott says the market is pricing in fear that LLMs could replace large parts of enterprise software. But in an AI economy where companies need to operationalize agentic AI safely, ServiceNow looks more like the control layer for governance, context, routing and oversight. The market is still struggling to separate software that gets disrupted by AI from software that becomes more valuable because AI needs exactly that control layer. I think ServiceNow sits much closer to the second bucket.