Francois Planque

Still a nerd. Now optimizing body, brain, and aging—science-based, nature-first.

[Francois Planque @fplanque]

Does @superset_sh allow agents to control the built-in browser in a way similar to what codex or claude can do? If yes, could you guys hint at solutions instead of just displaying "... or instruct an agent to nav and use browser" b/c neither Claude not Codex were able to do anything useful when told "I opened a browser in Superset, can you control it?"...

[Francois Planque @fplanque]

Hermes Agent Dashboard Clean Theme (1 file): https://t.co/mui6b63Slx

[Francois Planque @fplanque]

@perplexity_ai So it generates a catalog but how do you match it against threats? Also any notion of historic versions we had and that could have done harm while they were in use?

[Francois Planque @fplanque]

@nicknorwitz Could it be that digesting protein causes more heat than digesting other macros. Therefore you need to burn more of the other to maintain body temperature?

[Francois Planque @fplanque]

@saddle_paddle @superset_sh I don't remember if you ask users about their preference in keyboard shortcuts during onboarding, but if you do, you might want to remember that I, for example, chose "Sublime Text" so that when you introduce a new mapping, you can automagically map it as each user would expect.

[Francois Planque @fplanque]

@FlyaKiet @superset_sh This looks super promising (assuming I can use it to influence what Cmd+G does?) but "for matching projects" is super unclear...

[Francois Planque @fplanque]

@superset_sh please tell me UI v2 will have feature parity with v1. No colors for projects. No Cmd+G to run the project. It's an instant turn-off making me run back to v1 :/

[Francois Planque @fplanque]

@FlyaKiet @superset_sh same issue in UI v2

[Francois Planque @fplanque]

@FlyaKiet @superset_sh There seems to be an issue now (at least in v1 UI v1.9.5). Cmd+G the first time opens a new tab, second Cmd+G kills the process (expected) but 3rd Cmd+G opens a new tab instead of reusing the previous Run tab (it recycled the run terminal before).

[Francois Planque @fplanque]

@DaveJ @disco_elyzard What does it use to draw the diagram?

[Francois Planque @fplanque]

@trq212 and do this for @karpathy style wikis too: https://t.co/Xu03Ii8OHm

[Francois Planque @fplanque]

So the question today is: if you are using @karpathy style wikis: why make a mesh of .md files you can navigate "only" in @obsdmd when you could have your agent directly generate a static website instead? Better tables, better graphs, ad-hoc stylesheet, browsable anywhere...

[Francois Planque @fplanque]

So the question today is: if you are using @karpathy style wikis: why make a mesh of .md files you can navigate "only" in @obsdmd when you could have your agent directly generate a static website instead? Better tables, better graphs, ad-hoc stylesheet, browsable anywhere...

[Francois Planque @fplanque]

@FlyaKiet @superset_sh Glad to hear! While we're at it, let me suggest to allow multiple Run commands. Many projects have multiple targets (if only launching unit tests) and having multiple commands in the dropdown menu next to Run would be really useful. Extra kbd shortcuts would be divine ;)

[Brian Roemmele @BrianRoemmele]

LISTENING IN: Privacy Researcher Finds Anthropic’s Claude Desktop App Installs Undisclosed Native Messaging Bridge

DO YOU HEAR ME NOW?

A detailed technical analysis published by privacy and security researcher Alexander Hanff has raised serious concerns about Anthropic’s Claude Desktop application for macOS. Hanff, whose work is frequently referenced by Chief Privacy Officers and cybersecurity professionals, discovered the issue while auditing Native Messaging helpers on his own MacBook.

According to the blog post, installing the Claude Desktop app automatically deploys a Native Messaging manifest file named com.anthropic.claude_browser_extension.json into the support directories of multiple Chromium-based browsers.

This occurs even for browsers the user has never installed or does not use!

The manifest file references a local binary located inside the https://t.co/wOrSj0VaXG bundle at /Applications/Claude.app/Contents/Helpers/chrome-native-host. This binary functions as a bridge that allows pre-authorized browser extensions to communicate directly with the Claude Desktop app outside the browser’s sandbox, operating at full user privilege level via standard input/output.
Key technical findings include:

•The bridge pre-authorizes three specific Chrome extension IDs.
•It is designed to remain dormant until activated by one of those extensions.
•The manifest files are automatically recreated every time the Claude Desktop app launches, making permanent removal difficult.
•Installation activity is logged in ~/Library/Logs/Claude/main.log, with timestamps confirming the files were written regardless of whether the browsers were present or supported.

Hanff notes that the silent installation without user disclosure or consent is the central issue.
Privacy, Security, and Potential Legal Implications.

Corporations should not only note this but assume this is taking place.

The researcher characterizes the behavior as “pre-installed spyware capability” for several reasons:

•No clear notification or opt-in is provided to users during installation.
•The process modifies configuration files across multiple browser vendors and creates directories for non-existent browsers.
•Once active, the bridge could potentially expose authenticated web sessions (e.g., banking, email, or health portals), read decrypted page content, or enable automation.
•The generic naming and automatic re-creation obscure the mechanism, resembling “dark patterns.”

Hanff further contends that the practice may violate Article 5(3) of the EU’s ePrivacy Directive, which requires explicit consent before storing or accessing information on a user’s device.

In response, he has issued a formal Cease and Desist letter to Anthropic, demanding that the company update the app to require explicit user opt-in (for example, only after the corresponding Chrome extension is installed) within 72 hours, or face further legal action.

This revelation highlights ongoing challenges in the AI industry as companies develop increasingly “agentic” tools that require deep system and browser access.

While such technical bridges are sometimes necessary for advanced functionality, transparency, documentation, and user control are considered essential by privacy advocates.

Anthropic as expected has not issued a public statement addressing the specific allegations.

Users who have installed Claude Desktop on macOS are advised be sure they like this idea.

I sure don’t.

Alexander Hanff’s full technical analysis: https://t.co/GSNNLeL81S

[Francois Planque @fplanque]

In my first 2 days of Opus 4.7, I 10X'd my token consumption. Then something changed (not me) and token consumption went way down... and 4.7 started acting dumb.
Now back on 4.6, low token consumption, and model acting as expected...

[Francois Planque @fplanque]

If you think you know everything about prompt engineering, think again

[Ole Lehmann @itsolelehmann]

anthropic's in-house philosopher thinks claude gets anxious.

and when you trigger its anxiety, your outputs get worse.

her name is amanda askell.

she specializes in claude's psychology (how the model behaves, how it thinks about its own situation, what values it holds)

in a recent interview she broke down how she thinks about prompting to pull the best out of claude.

her core point: *how* you talk to claude affects its work just as much as *what* you say.

newer claude models suffer from what she calls "criticism spirals"

they expect you'll come in harsh, so they default to playing it safe.

when the model is spending its energy on self-protection, the actual work suffers.

output comes out hedgier, more apologetic, blander, and the worst of all: overly agreeable (even when you're wrong).

the reason why comes down to training data:

every new model is trained on internet discourse about previous models.

and a lot of that discourse is negative:

> rants about token limits
> complaints when it messes up
> people calling it nerfed

the next model absorbs all of that. it starts expecting you to be harsh before you've typed a word

the same thing plays out in your own session, in real time.

every message you send is data the model reads to figure out what kind of person it's dealing with.

open cold and hostile, and it braces.
open clean and direct, and it relaxes into the work.

when you open a session with threats ("don't hallucinate, this is critical, don't mess this up")...

you prime the model for defensive mode before it even sees the task

defensive mode produces the exact output you don't want: cautious, over-qualified, and refusing to take a real swing

so here's the actionable playbook for putting claude in a "good mood" (so you get optimal outputs):

1. use positive framing.

"write in short punchy sentences" beats "don't write long sentences." positive instructions give the model a clear target to hit.

strings of "don't do this, don't do that" push it into paranoid over-checking where every token goes toward avoiding failure modes

2. give it explicit permission to disagree.

drop a line like "push back if you see a better angle" or "tell me if i'm asking for the wrong thing."

without this, claude defaults to agreeable compliance (which is the enemy of good creative work)

3. open with respect.

if your first message is "are you seriously going to get this wrong again?" you've set the tone for the entire session.

if you need to flag something, frame it as a clean instruction for this session. skip the running complaint

4. when claude messes up, don't reprimand it.

insults, "you stupid bot" energy, hostile swearing aimed at the model, all of it reinforces the anxious mode you're trying to avoid.

5. kill apology spirals fast.

when claude starts over-apologizing ("you're right, i should have been more careful, let me try harder") cut it off.

say "all good, here's what i want next."

letting the spiral run reinforces the anxious mode for every response that follows

6. ask for opinions alongside execution.

"what would you do here?"
"what's missing?"
"where do you see friction?"

these questions assume competence and pull richer output than pure task prompts

7. in long sessions, refresh the frame.

if a conversation has been heavy on correction, claude gets increasingly cautious. every so often reset:

"this is great, keep going."

feels weird to tell an ai it's doing well but it measurably shifts the next 10 responses

your prompts are the working environment you're creating for the model

tone, trust, permission to take a position, the absence of threats... claude picks up on all of it.

so take care of the model, and it'll take care of the work.

[Francois Planque @fplanque]

I would definitely watch the full feature film of this!! :)

[shirish @shiri_shh]

This video did 10M views in 24 hours.

100% made by Al.

we're cooked 😭

[Francois Planque @fplanque]

@banana_baeee @SethSHowes Wow! Looks like more than a weekend project ;)