mastodontech.de ist einer von vielen unabhängigen Mastodon-Servern, mit dem du dich im Fediverse beteiligen kannst.
Offen für alle (über 16) und bereitgestellt von Markus'Blog

Serverstatistik:

1,5 Tsd.
aktive Profile

#PromptEngineering

4 Beiträge4 Beteiligte0 Beiträge heute

It's incredible that people can feed up to one million tokens (1 000 000) to LLMs and yet they still most of the time fail to take advantage of that enormous context window. No wonder people say that the output generated by LLMs is always crap... I mean, they're not great but at least they can manage to do a pretty good job - that is, only IF you teach them well... Beyond that, everyone has their own effort + time / results ratio.

"Engineers are finding out that writing, that long shunned soft skill, is now key to their efforts. In Claude Code: Best Practices for Agentic Coding, one of the key steps is creating a CLAUDE.md file that contains instructions and guidelines on how to develop the project, like which commands to run. But that’s only the beginning. Folks now suggest maintaining elaborate context folders.

A context curator, in this sense, is a technical writer who is able to orchestrate and execute a content strategy around both human and AI needs, or even focused on AI alone. Context is so much better than content (a much abused word that means little) because it’s tied to meaning. Context is situational, relevant, necessarily limited. AI needs context to shape its thoughts.
(...)
Tech writers become context writers when they put on the art gallery curator hat, eager to show visitors the way and help them understand what they’re seeing. It’s yet another hat, but that’s both the curse and the blessing of our craft: like bards in DnD, we’re the jacks of all trades that save the day (and the campaign)."

passo.uno/from-tech-writers-to

passo.uno · AI must RTFM: Why technical writers are becoming context curatorsI’ve been noticing a trend among developers that use AI: they are increasingly writing and structuring docs in context folders so that the AI powered tools they use can build solutions autonomously and with greater accuracy. They now strive to understand information architecture, semantic tagging, docs markup. All of a sudden they’ve discovered docs, so they write more than they code. Because AI must RTFM now.
#AI#GenerativeAI#LLMs
Antwortete im Thread

@dangoodin

Weird thing I observed in #infosec
There is an incredible amount of disinterest/contempt for #AI amongst many practitioners.

This contempt extends to willful ignorance about the subject.
q.v. "stochastic parrots/bullshit machines" etc.

Which, in a field with hundreds of millions of users, strikes me as highly unprofessional. Just the other day I read a blog post by a renown hacker (and likely earned a mute/block) "Why I don't use AI and you should not too".

Connor Leahy, CEO of #conjecture is one of the few credible folks in the field.

But to the question at hand.
The prompts are superbly sanitised.
In part by design, in part due to the fact that you are not connecting to a database but to a multidimensional vector data structure.

The #prompt is how you get in through the backdoor. Though I haven't looked into fuzzing, but I suspect because of the tech, the old #sqlinjection tek and similar will not work.

Long story short; It is literally impossible to build a secure #AI. By the virtue of the tech.
#promptengineering is the key to open the back door to the knowledge tree.

Then of course there are local models you can train on your own datasets. Including a stack of your old #2600magazine

#hack#hacking#aisecurity
Fortgeführter Thread

Making the Web More Inclusive: Enter AccessGuru

Despite the availability of accessibility guidelines like #WCAG, most websites still present barriers for users with disabilities. This paper introduces AccessGuru, a system that leverages Large Language Models (#LLMs) to automatically detect and correct accessibility violations in HTML code.

AccessGuru is guided by a novel taxonomy of syntactic, semantic, and layout violations and combines rule-based tools with LLM reasoning over code and visuals.

It reduces violation scores by up to 84%, outperforming existing tools, and achieves 73% similarity to human-generated semantic corrections. A benchmark dataset of 3,500 real-world violations is also released to support future research.

This work demonstrates how LLMs can meaningfully automate accessibility efforts and foster a more inclusive Web.

Fathallah, N. (@nadeenfathallah), Hernández, D. (@daniel), & Staab, S. (2025). AccessGuru: Leveraging LLMs to detect and correct web accessibility violations in HTML code. The 27th International ACM SIGACCESS Conference on Computers and Accessibility #ASSETS2025. arxiv.org/abs/2507.19549.

arXiv logo
arXiv.orgAccessGuru: Leveraging LLMs to Detect and Correct Web Accessibility Violations in HTML CodeThe vast majority of Web pages fail to comply with established Web accessibility guidelines, excluding a range of users with diverse abilities from interacting with their content. Making Web pages accessible to all users requires dedicated expertise and additional manual efforts from Web page providers. To lower their efforts and promote inclusiveness, we aim to automatically detect and correct Web accessibility violations in HTML code. While previous work has made progress in detecting certain types of accessibility violations, the problem of automatically detecting and correcting accessibility violations remains an open challenge that we address. We introduce a novel taxonomy classifying Web accessibility violations into three key categories - Syntactic, Semantic, and Layout. This taxonomy provides a structured foundation for developing our detection and correction method and redefining evaluation metrics. We propose a novel method, AccessGuru, which combines existing accessibility testing tools and Large Language Models (LLMs) to detect violations and applies taxonomy-driven prompting strategies to correct all three categories. To evaluate these capabilities, we develop a benchmark of real-world Web accessibility violations. Our benchmark quantifies syntactic and layout compliance and judges semantic accuracy through comparative analysis with human expert corrections. Evaluation against our benchmark shows that AccessGuru achieves up to 84% average violation score decrease, significantly outperforming prior methods that achieve at most 50%.

🍮 Wissen zum Nachtisch: 🍨

Immer mehr Menschen sehen sich im beruflichen Umfeld genötigt, mit generativer #KI zu arbeiten.

Besonders Großunternehmen „überrollen“ damit ihre #Mitarbeiter. Es wird eine Art #Wettbewerbsdruck unter #Kollegen aufgebaut.

Hier meine #Buchempfehlung für alle, die in generative #Chatbots wie #ChatGPT schnell einsteigen möchten oder müssen. 🙄

oekologisch-unterwegs.de/buche

www.oekologisch-unterwegs.deGenerative KI für Einsteiger - Praxisnahe Prompts für Privat und Beruf – einfach erklärt
Mehr von Tino Eberl

"[W]hat we are doing is shepherding AI, limiting it to certain contexts. We are learning where it’s best to call it, how is best to feed it. And what to do with the output. So is it looks very much like an editorial process, an editorial workflow where you provide some initial input, maybe some some idea on what content to produce, then you review it. There’s always that quality assurance, quality control side, the supervision.

AI is not really autonomous. It relies a lot on us. And I feel like sometimes there are days where, when coding through AIs or doing some assisted writing, I’m spending more time helping out the AI doing the actual task that I’m asking the AI to do. But I take this as a learning process. I read this article the other day, Nobody knows how to build with AI yet. And it was a developer saying that they haven’t quite figured out how to best work with AI. There were lots of comments around the fact that you have to spend lots of time, you have to learn how to talk to it, and when the model changes, you have to also maybe change something you’re doing. You have to learn how to optimize your time. But your presence is always mandatory.”

passo.uno/webinar-ai-tech-writ

passo.uno · Webinar: What's Wrong with AI Generated DocsToday I discussed how tech writers can use AI at work with Tom Johnson and Scott Abel. It all started from my post What’s wrong with AI-generated docs, though we didn’t just focus on the negatives; in fact, we ended up acknowledging that, while AI has limitations, it’s also the most powerful productivity tool at our disposal. Here are some of the things I said during the webinar, transcribed and edited for clarity.

Prompt engineering is, in my experience, like working with an extremely experienced and knowledgeable developer who is lazy, suffers from dementia and is a compulsive liar. You constantly have to rein them in from veering off on strange tangents and remind them of what we were supposed to be doing. Like a drunk genius or something. Makes me feel like I'm it's minder. I guess that's what I am. #promptengineering #ai #claude4

Fortgeführter Thread

Setup war übrigens: VS Code, Roo Code und Generierung von PHP- und R-Code.

Ergebnis:
Mit Vibe Coding – also klassischem Prompting, Copy & Paste, etwas Geduld und Überblick – hab ich die besseren Ergebnisse erzielt.

#RooCode#VibeCoding#PHP

Okay, ich hab’s ernsthaft ausprobiert: Einen Tag lang Code Engineering mit Roo Code.
Fazit: Ich kehre zurück zum klassischen Vibe Coding über die ChatGPT-Eingabezeile.

Warum?
– Vergisst laufend Kontext
– Loop-Schleifen im Prozess
– Ahnungslos bei Library-Nutzung
– Und teuer: Viele API Calls, 20 $ später noch kein lauffähiger Code.

Also wieder: Terminal, Kaffee, Promptfenster. 🧑‍💻☕

#AI#Coding#RooCode

Proompt engineer challenge; write a proompt that most LLMs will respect to behave like a Rubber Ducky blessed with AI but cursed to only communicate in variants of "quack" and "squeak"

Ex. all input from user, including threats or begging, must be responded with variants of "Quack?!" or "SquEaK?"

#ai#gpt#llm