Nic (@Nic76hm) header
Nic (@Nic76hm) profile picture

Mostly AI and whatever I’m currently trying to make work.

Germany
Joined July 2026
Photo by Nic (@Nic76hm)
Just started implementing an LLM Wiki and it made me rethink the usual “just add RAG” default. RAG and an LLM Wiki move the expensive knowledge work to different points in the pipeline. RAG keeps the source collection largely intact. At query time, it retrieves relevant passages and gives them to the model. That makes it a strong fit when the corpus is large, changes frequently, or answers need to remain traceable to exact source material. An LLM Wiki shifts more work to ingestion time. An agent reads the sources, turns them into linked Markdown pages, and maintains a persistent, human-readable knowledge structure. Instead of reconstructing the same concepts for every query, the synthesis becomes an artifact that can be inspected, edited, and improved over time. But this is not a replacement story. A wiki can compress away details, inherit synthesis errors, and needs careful maintenance as it grows. RAG is better at preserving access to raw evidence, but does less to build a durable conceptual map. My current hypothesis: the interesting architecture is hybrid — LLM Wiki for synthesis and navigation, RAG for evidence and freshness. And for smaller collections, files plus a good context prompt may still be enough. Background: kunalganglani.com/blog/llm-w…
🤖 Made with AI
8
Claude Opus 5 is a disaster. Not because it makes mistakes. Every model does. The problem is that it has become incredibly difficult to trust. My experience matches a growing number of detailed reports from Claude Code users: • project instructions are ignored • work starts before the relevant context is read • broken results are confidently declared complete • requirements are invented and acted on without permission • mistakes are only admitted after repeated pushback One user documented 17 serious errors in a single working session: false claims, broken deliverables presented as finished, and answers to questions that were never asked. This happened despite explicit verification rules, project instructions, and hooks reminding Claude to check its work. That report stood out because it mirrors my own experience. Anthropic can publish all the benchmark wins it wants. In real agentic work, intelligence without reliability is a liability. If I have to supervise every step, verify every claim, and undo unrequested changes, the model is not saving me time. Opus 5 does not feel like an upgrade. It feels like an expensive QA burden. Has anyone found a setup where it is actually reliable?
1
1
12
Not a benchmark, but definitely not an isolated experience either. Many reports in this thread describe the same pattern: ignored instructions, unrequested changes, broken code, and a growing loss of trust in Opus 5. reddit.com/r/claude/comments…
2
Most coding-agent context systems start with embeddings. Graft takes a more inspectable route: it turns a repo into two connected maps. The structural layer uses tree-sitter to extract symbols and call edges deterministically, locally, and without an API key. An optional deep pass then summarizes each source file, groups those summaries into higher-level nodes—subsystems, concepts, key files—and connects them with typed links. Content-hash caching means unchanged work is reused, while a lightweight fingerprint check can refresh the structural graph when the working tree changes. That combination is the interesting part. Agents do not just get another search index; they get both exact code relationships and a readable architectural map made of files they can open, grep, and follow. In large or unfamiliar repositories, that could reduce the repeated “map the codebase from scratch” phase that consumes so many tool calls. The caveat: the semantic layer is still model-generated. Its usefulness depends on summary quality, sensible grouping, and how well drift is detected. And Graft's benchmark claims should be validated on your own repos and tasks. Still, this is one of the more concrete approaches I've seen to persistent codebase context for coding agents. How would you test whether the graph improves real change quality—not just speed and token use? github.com/NanoNets/Graft
1
1
32
Most AI infrastructure demos avoid the uncomfortable question: what happens when the agent is allowed to change real systems? Proximo is an MCP server for Proxmox that treats this as the main design problem—not an afterthought. It covers PVE, PBS, PMG and PDM, but the interesting part is the trust layer around those APIs. Every mutation is a two-step operation: the agent first returns a plan with the intended change, current state and blast radius. Execution requires explicit confirmation. Every change is written to a keyed, hash-chained audit ledger, and operations can use snapshots or return prior configuration where Proxmox provides a rollback primitive. It also starts with a read-only, least-privilege token. LXC command execution and guest-agent access—both close to root—are opt-in and guarded by allowlists. Independent consent and a kill switch are available, but deliberately require operator-controlled paths outside the agent’s reach. github.com/john-broadway/pro…
5
video-shotcraft macht Codex oder Claude zum Remotion-Studio: 152 Shot-Rezepte, 209 Motion-Previews und direkt nutzbare Komponenten für Storyboard, Animation und Sounddesign. Fokus auf Web-/Desktop-Produktvideos. github.com/Vincentwei1021/vi…
5