Building agentic AI systems | Tutorials & experiments on local LLMs and retrieval | Open-sourcing what I learn.
Joined February 2026
- Tweets621
- Following211
- Followers15
- Likes59
The bar for learning to code hasn’t disappeared, it’s shifted: less memorizing syntax, more understanding architecture and debugging AI-generated output.
I've spent serious hours living inside Claude Code, and the biggest gains didn't come from fancy tricks — they came from a handful of small habits I now use on autopilot. Here's my running list of tips, one per post, no fluff. Steal the ones that fit and drop the rest. 👇
1. 🧵 Claude Code tips & tricks — a thread 👇
2. Use plan mode before any non-trivial change.
3. Rewind with checkpoints when a change goes sideways.
4. Write a CLAUDE.md (run `/init` to draft one).
5. Give Claude something to verify against — tests, linter, type checker.
6. Explore → plan → code, in that order.
7. Manage context with `/clear` and `/compact`.
8. Delegate noisy investigation to subagents.
9. Run parallel sessions with git worktrees.
10. Save reusable prompts as custom slash commands.
11. Automate formatting and guardrails with hooks.
12. Connect your tools with MCP servers.
13. Paste screenshots instead of describing them.
14. Interrupt with Esc and steer early.
15. Be specific up front — file names, errors, failing tests.
16. Set a `/goal` and let it grind until the condition is met.
17. Add an adversarial review step (`/code-review` or a skeptical subagent).
18. Pipe Claude into scripts with headless mode.
19. Track spend with `/usage`.
20. Match model + effort level to the task.
21. Work from your phone with Remote Control.
22. Let Claude drive the browser to test web apps.
23. Resume or fork sessions instead of restarting.
24. Ask Claude Code about its own capabilities.
25. Pick two, use them this week.
What did I miss? 👇
Achievement Unlocked: Maxed out Fable 5. 👑
We’ve hit our usage limits for the cycle. Time to step away from the screen for a bit!
Stop dumping raw history into your AI Agent's context. It's expensive and confuses the model 📉
TencentDB Agent Memory provides local long-term memory for AI Agents via a 4-tier progressive pipeline. It replaces flat vector storage with a layered architecture to improve reasoning 🧠
✨ Key features:
• Symbolic short-term memory using Mermaid graphs to condense logs 📉
• 4-tier long-term memory layering (Conversation → Atom → Scenario → Persona) 🏛️
• Deterministic drill-down paths for full traceability back to raw logs 🔍
• Local-first execution with zero external API dependencies 💻
Engineers building long-horizon agents can cut token usage by up to 61% 💰. Developers get higher task success rates by eliminating context overload 🚀.
So "Owl Alpha" — the model that's been quietly climbing OpenRouter — was this all along. 🐱
LongCat-2.0: 1.6T MoE, ~48B active, and it edges GPT-5.5 on SWE-bench Pro (59.5 vs 58.6).
The dynamic per-token compute (33–56B) is the part to watch. 🐱
Introducing LongCat-2.0 🐱
1.6T parameters · MoE with ~48B active · 1M context
The full model behind Owl Alpha on @OpenRouter — now available.
Built for agentic coding from the ground up:
◆ LongCat Sparse Attention (LSA) — scales efficiently for 1M-context tokens
◆ Zero-Compute Experts — dynamic activation 33B–56B per token, zero wasted compute
◆ MOPD — three specialized expert groups (Agent / Reasoning / Interaction), gate-routed per task
How it stacks up:
→ Terminal-Bench 2.1: 70.8
→ SWE-bench Pro: 59.5 (GPT-5.5: 58.6)
→ SWE-bench Multilingual: 77.3
→ FORTE: 73.2 · RWSearch: 78.8 · BrowseComp: 79.9
📖 Tech Blog: longcat.chat/blog/longcat-2.…
Try it across different scenarios 🧵👇
Must know when dealing with local AI
Compression techniques I’d study if I wanted small but smart LLMs.
1.Quantization
2.Distillation
3.Low-Rank Adaptation
4.Weight Sharing
5.Sparse Matrices
6.Layer Dropping
7.Knowledge Transfer
8.Embedding Compression
9.Mixed Sparsity
10. Progressive Shrinking
11.Structured Pruning
12.AutoML Compression
Follow @asmah2107 to update your game on LLM optimisations.
🚀 **Must-Read for Anyone Running Local LLMs in 2026**
"You don’t pick an inference engine first.
You pick your **hardware strategy**, **workload shape**, and **serving model**. The engine follows."
This is the clearest, most practical guide I’ve seen on choosing the right LLM inference engine. Hardware-first thinking + one-page decision table + deep dives on llama.cpp, MLX, ExLlamaV3, vLLM, SGLang, TensorRT-LLM and more.
Real bottlenecks explained (KV cache, bandwidth, scheduling), hardware recipes for every setup, and pro benchmarking tips.
Whether you're on a Mac, single RTX, multi-GPU rig, or running production serving — this will save you weeks of trial and error.
Part 3 of an absolute gold series by @TheAhmadOsman
🔖 Bookmark this. Seriously.
DROP EVERYTHING
The ultimate resource for running LLMs locally is now available online to read for free
Covers what to use on
- Laptop / edge / odd hardware
- Mac-first workflows
- Single RTX GPUs
- 2-4+ NVIDIA / CUDA GPUs
- General production serving
- Long-context / MoE / routing
- NVIDIA max performance
- Cluster orchestration
Software
- llama.cpp
- MLX / MLX-LM
- ExLlamaV2
- ExLlamaV3
- vLLM
- SGLang
- TensorRT-LLM
- NVIDIA Dynamo
You should read this, and if you cannot now then you most definitely wanna bookmark it for later
Opensource & Local AI FTW
This is the most of the agentic architectures has needed.
introducing tinyrouter
i reverse engineered the routing architecture behind Skana AI's Fugu and built replication for open frontier models.
it's a tiny ~10K parameter LLM router that learns which model to use and what role it should play for every question.
the goal was simple: beat every individual model by routing each task to the right specialist instead of relying on a single LLM.
github.com/harrrshall/tinyro…
a few interesting findings:
• routing only helps when the models have complementary strengths. if every model performs similarly on a benchmark, a router has almost nothing to optimize.
• on MMLU, the router beat every individual model. on math, it matched the best model because the pool had very little observable diversity.
• warm-starting the router and shaping the evolutionary reward improved training, but we don't claim a real gain yet because evaluation variance was too high. more rigorous experiments are needed.
🚀 Hy3 might be the open-weights flex of the year.
295B MoE going toe-to-toe with trillion-param giants — under Apache 2.0, so you can actually ship with it. API's free for 2 weeks, so no excuse not to test it. 👇
huggingface.co/tencent/Hy3
🚀Hy3 is here.
295B MoE. Best in its size class. Rivals trillion-scale flagships.
Reliable and affordable for most agentic usecases.
Apache 2.0. Friendly for commercial use.
FREE API for 2 weeks → openrouter.ai/tencent/hy3:fr…
🤗 huggingface.co/tencent/Hy3
📖 hy.tencent.com/research/hy3
