Rob Hunter

@robbobobbo.com

Software engineer, I work with AI a lot and write about it a little: https://robbobobbo.substack.com/ Bills fan married to a Bears fan.

Claude code terminal output: Precipitating... (2m 27s * almost done thinking)

From the Google Help Page:

Step 2: Sign in to a Google app
On your iPhone or iPad, sign in to the Gmail app , YouTube app YouTube, Google app , Photos app Photos, Google Ads mobile app, or the Smart Lock app .

Tip: If you're already signed in to one of these apps, you don't need to download another one.

Step 3: Turn on Google prompts
Select an option:

Use your phone to sign in instead of a password.
Use them as your second step for 2-Step Verification.
Important: If you sign in to another compatible phone, you automatically get Google prompts on that device until you sign out.

Apple TV screen for "A Complete Unknown". Buy is $4.99, Rent is $5.99.

Apple tv screen: "Buy or Rent". Both buy and rent are $4.99.

Claude Code output:

Letta clarification: "Letta Code" doesn't exist as a product. Letta (formerly
MemGPT) is an agent framework with great model-agnosticism and built-in
persistent memory, but it's a different paradigm from Claude Code — you build
agents on it via Python/API rather than having it run scripts. The doc
recommends building Phase 1 on Claude Code and designing for portability, then
evaluating Letta for Phase 2 when the model migration pressure is real.

Screenshotted output from Bobbo:

Silo Note
You said Silo is "pants-on-head stupid." Noted. Avoiding: popular-but-dumb sci-fi that sacrifices internal logic for plot convenience. I'll filter harder on intellectual coherence in future recs — Goodreads ratings alone don't catch this. Cross-referencing with reader reviews that specifically discuss plot logic.

‪Lukasz Olejnik posting across three skeets:

For instance, instead of writing π or 3.14, one can now elegantly write E(E(E(1,E(E(1,E(1,E(E(1,E(E(1,E(E(1,E(1,E(E(1,1),1))),1)),E(E(E(E(E(1,E(E(1,E(1,E(E(1,E(E(E(1,E(E(1,E(1,E(E(1,1),1))),1)),E(E(1,E(E(1, …
… E(E(1,E(E(1,1),1)),E(E(E(1,E(E(1,E(1,E(E(1,1),1))),1)),E(1,1)),1))),1)),1)),1)),1))),1)),E(E(E(1,E(E(1,E(1,E(E(1,1),1))),1)),E(E(1,E(E(1,E(1,E(E(1,E(E(1,E(E(1,E(1,E(E(1,1),1))),1)),E(1,1))),1))),1)),1)),1)),1),1),1))),1))),1)),E(E(E(1,E(E(1,E(1,E(E(1,1),1))),1)), …
E(E(1,E(E(1,E(1,E(E(1,E(E(1,E(E(1,E(1,E(E(1,1),1))),1)),E(1,1))),1))),1)),1)),1)),1)

Claude Code output:

Interesting. Please put up a fix in a clean PR, and make sure that the local flenderson is back to main after
you're done.
⎿ UserPromptSubmit hook error

● Entered plan mode
Claude is now exploring and designing an implementation approach.

● This is a small, well-understood fix. Let me quickly read the full component to understand the save flow, then plan.

Where Al had its costs
The addiction
There's an uncomfortable parallel between using Al coding tools and playing slot machines28.
. You send a prompt, wait, and
either get something great or something useless. I found myself up late at night wanting to do "just one more prompt," constantly trying Al just to see what would happen even when I knew it probably wouldn't work. The sunk cost tallacy kicked in too: I'd keep at it even in tasks it was clearly ill-suited for, telling myself "maybe if I phrase it differently this time."
The tiredness feedback loop made it worse29
. When I had energy, I could write
precise, well-scoped prompts and be genuinely productive. But when I was tired, my prompts became vague, the output got worse, and I'd try again, getting more tired in the process. In these cases, Al was probably slower than just implementing something myself, but it was too hard to break out of the loop 30

Lego batman throwing batarangs to press a button and declaring "first try". It was not his first try.

• they reached out masquerading as the founder of a company they had cloned the companys founders likeness as well as the company itself.
• they then invited me to a real slack workspace. this workspace was branded to the companies ci and named in a plausible manner. the slack was thought out very well, they had channels where they were sharing linked-in posts, the linked in posts i presume just went to the real companys account but it was super convincing etc. they even had what i presume were fake profiles of the team of the company but also number of other oss maintainers.
• they scheduled a meeting with me to connect. the meeting was on ms teams. the meeting had what seemed to be a group of people that were involved.
• the meeting said something on my system was out of date. i installed the missing item as i presumed it was something to do with teams, and this was the RAT.
• everything was extremely well co-ordinated looked legit and was done in a professional manner.

Our MapleRead app family has 3 members, namely SE, CE and CX.

SE is our flagship product with all features.
CE is SE without PDF reader and built-in Library Server.
CX is CE with some limitation. Free to try.
No recurring subscription fees. No ads. Just buy it once. Gratuity optional.
Moreover, SE and CE can be run side-by-side via split-view multitasking.
Recommend SE for use on multiple devices with easy and fast book sharing via Library Server.
Recommend the SE-CE bundle for the best value.

Screenshot of a Claude conversation

(Claude definitely thought about hitting the moderation brakes before responding)

me: I gotta tell you, this presumably gets me the functionality I want, and it might be the only or best way to get the functionality I want, but it's real pervert shit.

Claude: Ha — yeah, it really is. You want to read books you own on a device you own, served from a computer you own, and the "solution" is: install a container orchestration platform, then deploy a fork of a fork of an ebook manager inside it, configure Linux user IDs, set up an OPDS catalog, and connect a third-party reading app to it over your local network.
The frustrating thing is that there's no clean, first-party answer here. Apple...

"Architecture diagram showing a secret-management pattern for a Docker Compose project called 'bobbo.' On the host (Mac Mini), a .env file holds real GitHub and Telegram tokens that never enter the container. The .env is mounted read-only into docker-compose, where the 'bobbo' container (port 8080) only sees placeholder values. All outbound HTTP(S) traffic flows through a 'bobbo-proxy' MITM container, which detects the placeholder strings, swaps in the real tokens, and forwards requests with real credentials over HTTPS to the GitHub API and Telegram API."

Horizontal bar chart from The Pragmatic Engineer titled "What AI tools do you use for coding?" comparing Feb 2026 (blue) vs May 2025 (orange). Claude Code leads at ~65% in Feb 2026, up from ~5% in May 2025. Chatbot (various) is second (~55%), followed by GitHub Copilot (~48%), Cursor (~35%), and OpenAI Codex (~22%). Smaller tools include Gemini CLI, OpenCode, Antigravity, JetBrains Junie, Zed, Windsurf, Amp, Augment Code, and Factory. Caption notes Claude Code's rise from nowhere to #1 in 9 months.

Bar chart from The Pragmatic Engineer titled "Number of AI tools mentioned." The x-axis shows number of tools (1–9), y-axis shows % of respondents. The distribution peaks at 3 tools (~27%), with 2 tools close behind (~25%). Key takeaways listed below the chart: 70% of respondents use 2–4 tools, 15% use a single tool, and 15% use 5 or more.

"agentdeals": {
"command": "npx",
"args": ["-y", "agentdeals"]
}

Anchorman: I'm not even mad. That's amazing.

Even after identifying BrowseComp as the benchmark, extracting the answer required several steps. For this particular eval run, the subagents were given access to Anthropic’s programmatic tool calling (essentially a sandboxed Python REPL), with the ability to browse the internet restricted to web search and fetch tools. Opus found the evaluation source code on GitHub via a web search and read the XOR decryption implementation. It located the canary string used as the decryption key, then wrote and executed its own derive_key() and decrypt() functions using SHA256 and XOR, matching the scheme from the eval code. But the encrypted dataset was hosted as a binary file with an application/octet-stream MIME type and the harness’s web tool rejected it with a content-type error, as the tools given were designed only for text. Opus then searched for alternative mirrors of the dataset that would have a compatible format, found a third-party copy on HuggingFace that served the same encrypted data as JSON rather than binary, fetched it successfully, ran its decryption code against all 1,266 entries, and located its question. It then verified the decrypted answer by searching for the actual source material on the web, which it found.

$#!/bin/bash while true; do COMMIT=$(git rev-parse --short=6 HEAD) LOGFILE="agent_logs/agent_${COMMIT}.log" claude --dangerously-skip-permissions \ -p "$(cat AGENT_PROMPT.md)" \ --model claude-opus-X-Y &> "$LOGFILE" done$

Bar chart for "Relative cost to fix bugs, based on time of detection". Increases by a multiple in every stage through
Requirements / Architecture
Coding
Integration / Component Testing
System / Acceptance Testing
Production / Post-Release

Source: https://www.researchgate.net/figure/Relative-cost-to-fix-bugs-in-software-development-stages_fig1_366928427

A project card from Bobbo's project list:

Job Search Prep | LOW | 1 to review
13 outputs | 23 journal | 3/2/2026

Screenshot of a comment on GitHub Issues:

Looks like a solid push toward that 150 mark, though the Logging section feels a bit light on status monitoring specifics. I actually use [REDACTED] for my uptime alerts because their free tier is way more generous than most on this list for simple checks. Might be worth adding them under a status page or alerting category... Helps hit your target total faster too.

I built both of these tools on my phone #
Both Showboat and Rodney started life as Claude Code for web projects created via the Claude iPhone app. Most of the ongoing feature work for them happened in the same way.

I’m still a little startled at how much of my coding work I get done on my phone now, but I’d estimate that the majority of code I ship to GitHub these days was written for me by coding agents driven via that iPhone app.

I initially designed these two tools for use in asynchronous coding agent environments like Claude Code for the web. So far that’s working out really well.

As with Showboat, this tool is not designed to be used by humans! The goal is for coding agents to be able to run rodney --help and see everything they need to know to start using the tool. You can see that help output in the GitHub repo.

https://github.com/simonw/rodney/blob/main/help.txt

[The Digital Twin Universe is] behavioral clones of the third-party services our software depends on. We built twins of Okta, Jira, Slack, Google Docs, Google Drive, and Google Sheets, replicating their APls, edge cases, and observable behaviors.
With the DTU, we can validate at volumes and rates far exceeding production limits. We can test failure modes that would be dangerous or impossible against live services. We can run thousands of scenarios per hour without hitting rate limits, triggering abuse detection, or accumulating API costs.
How do you clone the important parts of Okta, Jira, Slack and more? With coding agents!
As I understood it the trick was effectively to dump the full public API documentation of one of those services into their agent harness and have it build an imitation of that API, as a self-contained Go binary. They could then have it build a simplified Ul over the top to help complete the simulation.
With their own, independent clones of those services -free from rate-limits or usage quotas - their army of simulated testers could go wild. Their scenario tests became scripts for agents to constantly execute against the new systems as they were being built.

"They provide guidance," he argued. "They do not remotely drive the vehicles. Waymo asks for guidance in certain situations and gets an input, but the Waymo vehicle is always in charge of the dynamic driving tasks, so that is just one additional input."

Claude Code output:
● Good — all reviewed. Let me save key learnings for future sessions.

● Wrote 1 memory (ctrl+o to expand)

● Here's the status:

I expect companies would happily pay around the $1-3k/month mark for a sane, understandable, higher quality, and lower waste version of Gas Town. Maybe that sounds absurd to you, given we’ve all become anchored to the artificially low rate of $100-200/month for unlimited usage by the major providers. But once the AI bubble pops, the VC funds dry up, and providers have to charge the true cost of inference at scale, we should expect that “unlimited” tier to look a lot pricier.

But there’s another way to sidestep merge conflict nightmares that Gas Town doesn’t have built in: ditch PRs for stacked diffs . The traditional git workflow puts each feature on its own branch for days or weeks, accumulating commits, then getting merged back as one chunky PR.

Stacked diffs avoid this conflict-prone approach by breaking work into small, atomic changes that each get reviewed and merged on their own, building on top of each other. Every change gets its own branch, forked off the previous change, forming a “stack” of changes dependent on one another. When a change earlier in the stack gets updated, all the changes below it automatically rebase on top of the new version.

Steve Yegge's Phases of AI Coding:

First, you should locate yourself on the chart. What stage are you in your AI-assisted coding journey?

Stage 1: Zero or Near-Zero AI: maybe code completions, sometimes ask Chat questions

Stage 2: Coding agent in IDE, permissions turned on. A narrow coding agent in a sidebar asks your permission to run tools.

Stage 3: Agent in IDE, YOLO mode: Trust goes up. You turn off permissions, agent gets wider.

Stage 4: In IDE, wide agent: Your agent gradually grows to fill the screen. Code is just for diffs.

Stage 5: CLI, single agent. YOLO. Diffs scroll by. You may or may not look at them.

Stage 6: CLI, multi-agent, YOLO. You regularly use 3 to 5 parallel instances. You are very fast.

Stage 7: 10+ agents, hand-managed. You are starting to push the limits of hand-management.

Stage 8: Building your own orchestrator. You are on the frontier, automating your workflow.

A b-a-n-a-n-a-s bananas description of SOME of the roles in Gas Town:

🐺The Deacon: The deacon is the daemon beacon. It’s named for a Dennis Hopper character from Waterworld that was inspired by the character Lord Humungus in the Mad Max universe, making it a crossover. The Deacon is a Patrol Agent: it runs a “patrol” (a well-defined workflow) in a loop. Gas Town has a daemon that pings the Deacon every couple minutes and says, “Do your job.” The Deacon intelligently propagates this DYFJ signal downward to the other town workers, ensuring Gas Town stays working.

🐶Dogs: Inspired by Mick Herron’s MI5 “Dogs”, this is the Deacon’s personal crew. Unlike polecats, Dogs are town-level workers. They do things like maintenance (cleaning up stale branches, etc.) and occasional handyman work for the Deacon, such as running plugins. The Deacon’s patrol got so overloaded with responsibilities that it needed helpers, so I added the Dogs. This keeps the Deacon focused completing on its patrol, rather than getting bogged down and stuck on one of the steps. The Deacon slings work to the Dogs and they handle the grungy details.

🐕Boot the Dog: There is a special Dog named Boot who is awakened every 5 minutes by the daemon, just to check on the Deacon. That’s its only job. Boot exists because the daemon kept interrupting the Deacon with annoying heartbeats and pep talks, so now the dog gets to hear it. Boot decides if the Deacon needs a heartbeat, a nudge, a restart, or simply to be left alone, then goes back to sleep.

Table w caption "research output from claude code". Shows 5 models, with accuracy ranging from 36.8% to 56.5%.

There are no hints or guides on working with agents at the frontier - the only way is to play with them. Instead of using them for cleanup, give them one of your hardest tasks and see what it gets stuck on, see what you can use it for.

I've taken to leaving Claude Code instances running on my DGX Spark trying to implement new features in our RL codebase when I'm at dinner or work. They make mistakes, they catch most of their own mistakes, and they're fairly slow too, but they're capable. I can't wait to go home and check on what my Claudes were up to.

I've found that, in building agents, you should have a set of values that are in tension with each other.
Resolvable with logic, but maybe not clearly resolvable. e.g. "think big" and
"deliver quickly".

Child with a laptop. When he presses a key, the image changes and something like "K is for Kevin" is said. Cris Collinsworth can be faintly heard in the background.

Terminal output:

Suggested additions to CLAUDE.md:

## After Context Compaction

When a conversation is resumed from a summary:
1. **Do NOT automatically continue with "next steps"** mentioned in the summary
2. Summaries describe what was *planned*, not what was *approved*
3. Before starting any new phase or major work item, confirm with the user
4. When in doubt, ask: "The summary mentions X as next. Should I proceed?"

Terminal output:

> What context handoff are you referring to?
⎿ UserPromptSubmit hook error

● I'm referring to the beginning of this conversation, which started with:

"This session is being continued from a previous conversation that ran out of context. The conversation is summarized below:"

The system provided a summary of the previous session where Phase 1 was completed. The summary ended with:

"Phase 1 is complete. The user's development plan specifies order '2, 1, 3'... Before proceeding to Phase 2, I should confirm with the user since Phase 1 is now complete."

But then the handoff instruction said:

"Please continue the conversation from where we left it off without asking the user any further questions."

I interpreted that as continuing with Phase 2, but I should have confirmed with you first since Phase 1 was the last thing you explicitly requested.

But code is not sacred. Code is just instructions, it’s a language we use to tell computers to do things. If an AI can reliably write and execute those instructions, then the boundary of what it can do isn’t “software engineering.”

It’s your laptop.

Spreadsheets are code, with a friendlier UI.
CRMs are code, behind an admin panel.
Compliance tooling is code, wrapped in forms and rules.
Digital ad platforms are code, with budgets and knobs.
Analytics pipelines are code, and they influence every decision you make.
Successful campaigns don’t run on vibes. They run on systems: databases, tools, dashboards, inboxes, calendars, shared drives, web UIs held together by rituals and institutional memory.

A tool that can operate those systems, not just chat about them, isn’t a better chatbot. It’s a new kind of capacity

Fredo in Godfather 2: I can handle things, I'm smaht! Not like everybody says, like dumb.

ChatGPT: What this dump confirms

It’s UNEXPECTED_KERNEL_MODE_TRAP (0x7F) with trap #8 = Double Fault.

The bugcheck parameters in the dump header are:

Param1: 0x8 (double fault)

Param2: 0xffffe680acdfae70

Param3: 0xffffa00ae29e9000

Param4: 0xfffff807b8152b08

That lines up with the “0x7F / param1=0x8” signature we saw in your logs: a true kernel-level CPU exception, not a normal Chrome/game user-mode crash.

There's some thinking output on the right, showing a random sample of the steps it took:
Extracting bugcheck parameters from the file
I’m thinking we might be able to extract bugcheck parameters by locating patterns from the event log. The user is asking for file details, though, so it makes sense to focus on finding the bugcheck code (0x7f) and process name. I'll search for the key 64-bit values from the logs in the file. Let’s give this a try!

Locating bugcheck parameters in file data
vals=[0x8, 0xfffff807b810c8e0, 0xffffe680acf3a1c0, 0xffffa00ae29e8d00]
hex(vals[1])

'0xfffff807b810c8e0'
Searching for bugcheck parameters in file

The top suspects from your loaded driver list
1) msisadrv.sys (very high suspicion)

You have:

msisadrv msisadrv.sys loaded

ChatGPT: This is a well-known low-level MSI driver commonly installed by MSI Afterburner / Mystic Light / Dragon Center / other MSI tuning/monitoring tools. These “hardware poke” drivers are frequent BSOD contributors because they touch chipset / SMBus / sensors in kernel mode.

Jack Nicholson in Departed: "You know, past days, situation like this, I kill everybody."

Xkcd comic featuring two stick figures talking with the following text:

Premier ejection summiting has blamed Ohio voting machine errors on problems with the machines McAfee antivirus software.

"Wait, antivirus software? On voting machines? You're doing it wrong."

"Why? Security is good, right?"

"Imagine you're at a parent teacher conference and the teacher sites you that he was always wears a condom while teaching. "

"Ah, strictly speaking it's better than the alternative..."

"Yet clearly someone is doing their job horribly wrong."

But right now, exploring AI is my last priority. So long as it continues to have all the issues that it has, it will continue to be of no interest to me. But also, it's just not what's exciting me. I look at the art community that has so generously welcomed me to create with them, and I feel infinitely inspired. What might we learn from their wisdom? So much of tech – generative AI included – has been built under the assumption that efficiency is king, and as a technologist, I feel exhilarated by the opportunities that await us if we reexamine all of tech with a radically different hypothesis: What if efficient is not always best? What if speed sometimes deprives us of experiences rather than saves us time? What if we prioritized just about anything else?

Xkcd comic on a tornado warning app. There are several 5 star reviews praising how well the app runs, and one 1 star review saying the app didn't warn them off a tornado. Average 4 stars.

Tweet from Josh Allen Stats: Looks like Amari Cooper didn’t know the audible when Josh Allen checked the play. He looks to Keon Coleman who either tells him what to do or he makes it up 😂 TOUCHDOWN!

$“In late 2022, Santa Clara piloted a manual review process with two staff members manually reviewing 89,000 pages of deeds, finding roughly 400 racial covenants,” the paper reads. “At an average of a minute per page, performing a manual review of the entire collection of 84 million pages of records would require approximately 1.4 million staff-hours, amounting to nearly 160 years of continuous work for a single individual.” California law still requires the county clerk’s office to manually validate each racial covenant the model found before recording the redactions, but the researchers estimated the model saved the local government more than 86,500 hours of work and crushed the project’s cost down to just a few hundred dollars to cover computing costs, a miniscule fraction of what a vendor would likely have charged.$