Michael Wood

I suppose one could ban it for coding and use the hell out of it for things like reviewing PRs?

I've had a look and I see they do the Component.tsx + Component.web.tsx thing in some places. The alternative is using the Platform module

OK. At worst you could have Editor.tsx and Editor.web.tsx

Stress Testing Deliberative Alignment for Anti-Scheming Training Highly capable AI systems could secretly pursue misaligned goals -- what we call "scheming". Because a scheming AI would deliberately try to hide its misaligned goals and actions, measuring and mitiga...

Just saw this today:
arxiv.org/abs/2509.15541
I agree it's not being "honest" or "lying" etc., but it kind of seems like that sometimes

I haven't tried it yet, but this looks pretty good:
github.com/software-man...

My grandmother was a weaver and spinner and did card wool before spinning it :)

That page doesn't seem to mention Android. I don't generally access my email on my phone, but the usual Open Source mail clients for Android are K-9 Mail and Thunderbird. The former was always an Android app (maybe iOS too), whereas the latter was a desktop app. I think they are merging now

It does for some things. Images, videos (but not youtube). Maybe that's all

And now OpenClaw

Reminds me of the time I got a massive cramp in my hamstring while I was at a café with my boss and had to jump up and then go outside and stretch it

Switching to a local tool like gemini-cli, codex-cli or claude code is much better than the web interfaces for editing code

bsky.app/profile/nora...

Discovered that a week or two ago as well

Apparently Grok hallucinated that "paid accounts only" restriction and the media believed it

GitHub - mattpocock/evalite: Evaluate your LLM-powered apps with TypeScript Evaluate your LLM-powered apps with TypeScript. Contribute to mattpocock/evalite development by creating an account on GitHub.

Not sure if evalite might be usable for this:
github.com/mattpocock/e...

Get a polarised filter (if you can) :)

Seems sort of related. Haven't tried it: Ralph Wiggum
twitter.com/mattpocockuk...

I was a bit older. Had to reach around the back of the TV cabinet to unplug my ZX Spectrum, but the socket was quite tight, so I readjusted my fingers to get a better grip and touched the live and neutral pins. Hello 230V

Yeah! Galaxy Tab is much better!

/s

Same. I turned off autocompact a while ago. Not sure how much it helped, but it *seemed* to help.

Do you have autocompact enabled? Do you have the context window % in your status line? I'm interested to know if there's any correlation between those sorts of issues and the fullness of the context window.

I find that getting Claude, Gemini and Codex to review the changes afterwards often picks up issues, although Gemini (2.5. Maybe 3 is better) is way too eager to say everything's brilliant.

Review Complete File: reviews/commits_origin-main_to_HEAD_20251226_205556.md Key Findings | Severity | Issue | Fix Effort | |----------|------------------------------------------------|------------| | Critical | Idempotency guard broken - checks wrong string | 5 min | | Critical | Silent regex failure - no validation | 10 min | | High | Zero plugin tests | 60 min | | High | Magic string duplicated | 5 min | Agent Agreement - Quality + Codex: Both caught the idempotency bug - Gemini R1: Incorrectly claimed plugin was "idempotent" and "well-crafted" - Gemini R2: Corrected itself after reading other reviews Verdict Block merge until idempotency guard is fixed. Second npx expo prebuild will break the build. * Insight The idempotency bug is a classic config plugin pitfall - the guard string 'DEV_SETTINGS_PREFERENCES' was likely intended but never made it into the generated code. The fix is trivial: check for the actual comment or SharedPreferences key that IS inserted.

µmging is more about the context win + what phase of the plan to do next than the code changes. What I don't like about plan mode is that it wants to immediately start making changes after finishing the plan. I tell it to stop. I manually switch out of plan mode and go from there
Review screenshot:

I add directories of relevant stuff, but often have to remind it that it has access to it and it doesn't have to search the web

I just finished watching the 2023 Death in Paradise Christmas Special! 😊

I do feel I need to do a lot of micromanaging, though.
I'm trying more unit tests, e2e tests and getting Claude to run them after making changes and propose new tests for new features. So far looks promising, but I've only just added the thing about proposing test, so don't know how well it'll work

like above 30% (sometimes I let it go higher, but don't really like it higher than 30-something), then I compact before doing the next phase.
Then I get Claude to review the changes and also get Gemini and Codex to review them. Codex seems better at this, but they all find issues to fix

I have been getting Claude to write a plan, then it often asks questions about aspects of thd work, which I answer. Then I might ask it about parts of it or tell it what I want changed, etc. Then I get it to do one phase or sub-phase. I have the context % in my status line. If it starts getting high

In all interactions and commit messages, be extremely concise and sacrifice grammar for the sake of concision. # GitHub * Your primary method for interacting with GitHub should be with the GitHub CLI (gh). # Plans * At the end of every plan, give me a list of unresolved questions to answer, if any. Make the questions extremely concise. Sacrifice grammar for the sake of concision. * When saving a plan to file or GitHub issue or comment, save it as a multi-phase plan with checkboxes for easily marking items as done.

Weird. I use it a lot. I have been using these instructions in ~/.claude/CLAUDE.md, based on @mattpocock.com's since a bit before they started autosaving the plans under ~/.claude, and I have been happy with it

Have you tried pressing Shift-Tab to get into plan mode? It should automatically write the plan to ~/.claude/plans/...