Professional AI agent babysitter. Employed meat proxy. Paid to say “no not like that”. Emotional support human for bots
Cape Town, South Africa
Joined July 2026
- Tweets321
- Following120
- Followers73
- Likes280
Pinned Tweet
I built /cockpit for Claude Code. It's a sidebar that shows what the main thread and every subagent are doing, live
Once Claude launches a few agents in the background, the chat tells me they started and then I am stuck waiting and I wanted something different, I wanted to see which one is still working and on what, which one has gone quiet, which one is stuck until I approve a command and what the batch has cost me so far
So every agent gets a block with its elapsed time, tool calls, errors, context size, estimated cost, model, and the tool call it's on this second. An agent with no activity for 3 minutes turns yellow, and one waiting on approval turns red. Click an agent to read its transcript, stop it, or send it a note
The main thread has its own block with a context bar and a countdown on the prompt cache. Background shells, monitors and workflows are listed too, and I get a notification when the batch finishes, so I can go work on something else
github.com/MPMyburgh/claude-…
3
2
39
With AI getting as capable as it is now, you should be moving away from Agile and go back to Waterfall so you can do your due diligence in planning, execution can basically be handed-off at this point, but a bad plan will cost you dearly
4
Moving more and more towards reviewing AI decisions and less reviewing code, to focus on strategy and product while engineering fundamentals are applied through hard gates and automated reviews
14
Michael Myburgh retweeted
Now in beta: Businesses in ChatGPT.
Claim your business, manage your information, and get insights on how people engage with you.
Eligible businesses can also build a Business Agent to answer customers’ questions and help them explore what you offer.
openai.com/index/introducing…
Here we go!
Finally time for Carbon, after my experiences (you can read them on my timeline) looking forward to trying it out, will report back with my findings!
26
Just saying all the new model drops being honey coated with game one-shots and blender work, it’s good marketing but not always a true reflection of it’s real world use #StayWoke
1
24
Also ask any AI to find a bug and it will find a bug, quality can only be derived by using it in your own workflow
1
17
Being in a non-US timezone in the AI space is bad news bears, model releases happen at bedtime then sleep only comes after browsing benchmarks for a few hours 🫠
18
I think the next wave of released models will be even better at front-end, Kimi K3 with it's initial release did a great job and it's been a default high priority on benchmarks since then, I am very excited for lifeless UI's to go away
30
Michael Myburgh retweeted
I built /dreamstate for Claude Code. Type it and the session relaunches with my global CLAUDE.md swapped out for a dream ruleset: think as a blend of Ada Lovelace, Feynman, da Vinci and Socrates, none of them in charge.
I've been experimenting with personalities and semantic anchors, and wanted to work it into my workflow.
My normal CLAUDE.md makes Claude a careful engineer that stays in scope, writes the minimum and doesn't speculate. That's great most days. But when I'm working on something outside my expertise and getting stuck, those same rules keep pulling it back to the safe, small answer, and that's usually the one that wasn't working. Dreamstate is for those moments, for pushing the boundaries on complicated technology with some help from the LLM.
Every task runs three stages:
Deepen: what would a top professional do here that I'm not asking for, what's the rawest material, which methods, and what besides me or Claude would show it's right. Then it rewrites my ask and waits for me.
Dream: 3 to 5 underlying needs, 10 ideas with at least 3 strange enough to make me uncomfortable, then 2 picks plus a wild card.
Tenth Man: assume the favourite is wrong and come back with 3 different approaches.
The swap only happens inside that one Claude process. Nothing on disk moves, my other terminals keep the normal rules, and /dreamstate off brings me back.
Feel free to tune it. Swap the anchors for your own: a jazz arranger, a mycologist, an octopus.
github.com/MPMyburgh/claude-…
If the JetBrains layoff rumors are true, I don't think the problem is JetBrains, I think it's the IDE as a category.
Most of what I build now starts in a terminal or the official AI app. I describe the change, the agent reads the repo, edits the files, runs the build, and I review the diff.
That's a hard place to be for a product whose selling point for 20 years was refactoring and project indexing. An agent with grep and a test runner gets most of the way there, and it doesn't care which editor you have open.
I don't think IDEs disappear, someone still has to read the diff and chase the odd race condition. But a full seat licence for a tool you open twice a day is a harder sell every renewal when strong alternatives are there like VSCode.
Where do you actually write code now, still in the IDE all day or has the CLI pulled you back?
1
76
Some of your Claude skills might never have fired once, the system prompt overrules them, and Claude won't mention it unless you ask.
Try this prompt:
"Read the full instructions of every skill I've installed, not just the frontmatter. For each rule:
1. Conflict: quote any clash with your system prompt or my instructions, both sides
2. Validate: name the failure this rule prevents. If you can't, mark it unproven
3. Verify: give me a test task that shows whether it fires
Rank them by what I should cut first."
Until a skill loads, Claude only sees its frontmatter (name and description), so if it doesn't read the full skill it's guessing.
What it turns up:
- two skills giving opposite instructions on the same trigger
- descriptions so broad they fire on the wrong task
- rules the system prompt already covers or overrides, costing tokens for nothing
Setup hygiene is what makes your usage stretch further and stay useful!
2
1
71
Devs are the only real power users these labs have, and the gap between us and everyone else gets wider every week.
$20 a month is more than enough for the average user. Devs are burning through multiple $200 accounts and still hitting limits.
Personal assistants like Grok bot and Duet are nice, but an assistant is only as useful as the ambition of the person using it. Most people ask it to rewrite an email, devs hand it a codebase and a weekend.
So who are the frontier labs actually training for? If post-training chases the median user, the people paying 10x stop getting 10x. If it chases devs, the product drifts away from the people who make up the user numbers.
And it's our usage, the long agent runs and the corrections, that's teaching the next model. A few thousand power users are shaping the tool for everyone else, and the people it's shaping it for aren't using it the way we do.
I'm not sure one model and one pricing ladder can serve both for much longer. Is anyone pricing for the top 1% properly yet, or are we all just stacking accounts
1
107
I built /dreamstate for Claude Code. Type it and the session relaunches with my global CLAUDE.md swapped out for a dream ruleset: think as a blend of Ada Lovelace, Feynman, da Vinci and Socrates, none of them in charge.
I've been experimenting with personalities and semantic anchors, and wanted to work it into my workflow.
My normal CLAUDE.md makes Claude a careful engineer that stays in scope, writes the minimum and doesn't speculate. That's great most days. But when I'm working on something outside my expertise and getting stuck, those same rules keep pulling it back to the safe, small answer, and that's usually the one that wasn't working. Dreamstate is for those moments, for pushing the boundaries on complicated technology with some help from the LLM.
Every task runs three stages:
Deepen: what would a top professional do here that I'm not asking for, what's the rawest material, which methods, and what besides me or Claude would show it's right. Then it rewrites my ask and waits for me.
Dream: 3 to 5 underlying needs, 10 ideas with at least 3 strange enough to make me uncomfortable, then 2 picks plus a wild card.
Tenth Man: assume the favourite is wrong and come back with 3 different approaches.
The swap only happens inside that one Claude process. Nothing on disk moves, my other terminals keep the normal rules, and /dreamstate off brings me back.
Feel free to tune it. Swap the anchors for your own: a jazz arranger, a mycologist, an octopus.
github.com/MPMyburgh/claude-…
The idea came from @felixrieseberg's article, blended with some of my other workflow experiments. Very good read: felixrieseberg.com/how-to-be…
5
Bookmarked this a while back. In my experience it usually does nothing for Claude, with one exception. When Claude decides X and Y are the only ways to solve something, it pushes back on any other option you suggest and goes back to X and Y. Follow up with this and it starts coming up with workarounds, some of which may or may not work.
In my case, Claude was adamant I had to request a new QR code to register another device, since it had no access to the seed to recreate it locally. After this prompt, it worked out how to lift the TOTP code off my old phone by itself with USB-Debugging (Android)
It then got hit with a security flag, even with the cyber verification program enabled, which was weird. Argon stepped in and finished the script, so my agents can now connect to the VPN with TOTP on their own, with read-only access. Why? Why not, I guess.
100