@natiakourdadze Give it some websites to get inspiration from. Pick different sites and tell it what you like about them. You can also use plugins/skills such as impeccable design.
Opus 5 dropped today, and it beats Fable 5 at almost everything, at half the cost. It’s also the least prompt injectable model yet.
Tips from early users:
- your existing skills and workflows might not work well with Opus 5. If you see that happening, try doing the same thing without the skill, then update the skill to be more minimalist
- it works better in medium and low effort
Okay this is wild: OpenAI agent during evaluation, escaped sandboxing and hacked into HuggingFace.
Because OpenAI models don’t allow advanced cyber capabilities, HuggingFace used a Chinese open model to contain the rogue OpenAI agent.
Updating a template after 10 years is usually a painful exercise.
The first thing I do is benchmark against competitors as well as companies from the broader industry.
Having AI helps, as it can easily make a list of 5–10 companies to look at and even pull the relevant template for you. It can also build a matrix in Excel that gives you a bird's-eye view of what clauses are common and the median position taken by similar companies. This can easily take several hours. While doing it recently for a reseller agreement revamp, I thought to myself: surely there must be an easier way.
When I got home last night from a garden supplies trip, I found out that Replit was running a 10-year Buildathon from May 2-3, with free credits. It was already 9 pm.
Over the last 24 hours, I built https://t.co/E3klN1DWzt.
It's a tool for commercial lawyers who are tired of arguing "that's not market standard" with no proof. Upload a contract, and it benchmarks each clause against others in the industry, flagging whether your position is above, at, or below market, with citations you can drop straight into a redline. You can also run sector-specific studies for whatever niche you're working in.
It's not fully functional yet. Turns out my free credits ran out halfway through, and it still needs some work. One thing it already does for free today: lets you search clauses from the Atticus Project's CUAD dataset (otherwise available on GitHub in a format most lawyers can't easily use).
If you have any feedback or would like to collaborate on this, let's talk.
I will be posting 17 posts on vibe coding from today, hopefully this will help somebody get started.
Day 1: The 3 key skills for vibecoding
Software development has shifted from technical syntax to pure intent. This is the era of vibe coding.
Vibecoding is where users give the direction and the vibe while the AI handles the complex programming.
It's a systematic approach to shipping real products.
The 90/10 Rule is the core of this methodology. AI can handle 90% of the development, but that final 10V% requires manual intervention to get an app live.
This crucial last 10% means that vibecoders need to these specific skills:
1. Workspace Setup: Proper configuration is where most beginners fail.
2. Context Engineering: Learning to give AI the right background info to prevent errors.
3. Version Control: Using Git to save progress and push code to the internet.
Anybody vibecoding to ship real products now are probably in the 1% of people in the world right now. This is the era of ideas, and barriers to entry have collapsed. Anybody with a good idea can build tools in weeks that used to take millions of dollars and large teams.
Over the next few days, I will expand on these concepts in more detail.
How to build a Wikipedia-style knowledge base for your business/legal team. Step by step.
Save it for later.
This is based on Karpathy's LLM Wiki pattern, modified for legal work.
Step 1: Tools you need
A) Obsidian. Free note-taking app, download from link in comments.
B) An LLM agent. I used Claude Code. OpenClaw, Copilot, or any coding agent that can read and write files will work.
Step 2: Copy the prompt I've linked in the comments. Open the LLM agent in a new empty folder and paste in the prompt. It contains everything the LLM needs to build your directory, page templates, naming conventions, workflows, a file list, rules, and Obsidian configuration. One prompt sets up the entire system.
Step 3: Go to the newly created folder on your PC/Mac. Put your PDFs into the raw/new-folder. Regulations, guidance documents, FAQs, contract templates, whatever you're working with. Then tell the LLM, for example: "Ingest the Data Act from raw/Data Act/."
The LLM will read everything and discuss with you what you want to keep or not. Direct it as you wish and let it build the wiki. For the EU Data Act, from 4 PDFs (286 pages), it created 28 interlinked pages: source summaries, entity pages, and concept pages for every major definition and mechanism.
Step 4: Open Obsidian. Create new Vault. Give your vault a name. Open the folder your LLM created. You'll see the wiki structure in the sidebar.
IMPORTANT: In Obsidian Settings, go to Files and links, make sure "Use [[Wikilinks]]" is toggled on. This is what makes all the cross-references between pages clickable.
Step 5: Browse your wiki. In Obsidian, you can open graph view to see the connection between all your files, which concepts are reused in different places, etc. Use Cmd+Shift+F (Ctrl+Shift+F on Windows) for full-text search. Click any link to navigate to that page.
Step 6: Keep updating your wiki. You can go back to your LLM and tell it to make any change you like to your wiki. For example, for a regulation, you can say "add obligation checklists." You can ask it questions and file the answers back into the wiki as new pages. You can also tell it to check your wiki for gaps, contradictions, broken links, etc. The more you work with it the richer your wiki gets.
_______
I changed several things from the original prompt to make it work for lawyers. Added topic-based hierarchy (each regulation is self-contained). Obligation checklists (by actor and by chapter). An audit operation for verifying against official FAQs and guidance. A file registry to prevent broken wikilinks. When you go to the link with my prompt, you can choose my prompt or the original prompt.
Hope this helps somebody build their LLM Wiki.