AI & LLM

AI & LLM

Jason Yang · 22 Jul, 2026

The hard part of a Claude Code agent loop isn't repeating work. It's a stop condition you can trust: convergent vs exploratory, and who enforces the verdict.

AI & LLM

Jason Yang · 22 Jul, 2026

Claude Code's /goal decides it's done from the worker's self-report. Add a Stop hook that verifies completion with exit codes, not the model's word.

AI & LLM

Jason Yang · 22 Jul, 2026

Karpathy's autoresearch pattern rebuilt for Claude Code: a Bash loop that scores each AI-tried hypothesis, keeps improvements, and rolls back failures with git.

AI & LLM

Jason Yang · 12 Jul, 2026

Three Claude Code skills for downloading YouTube captions as SRT, transcribing audio with a local Whisper server, and translating embedded subtitle tracks.

AI & LLM

Jason Yang · 09 Jul, 2026

What an AI harness is and how to build one — validation, gates, and Generator ≠ Evaluator separation at each stage of the SDLC.

AI & LLM

Jason Yang · 08 Jul, 2026

AI engineering is more than prompts. A practical frame — prompt, context, harness, loop — walked through one example: an AI that reviews every PR.

AI & LLM

Jason Yang · 02 Jul, 2026

Screenshots won't paste into Claude Code on Wayland? Install wl-clipboard (not xclip), then restart the CLI. Same fix works for Codex.

AI & LLM

Jason Yang · 09 Jun, 2026

Run the Ralph Loop in Claude Code: four ways to repeat work and stop on clear completion criteria — the ralph-loop plugin, /loop, a prompt loop, and Bash.

AI & LLM

Jason Yang · 05 Jun, 2026

The Claude Code debate over Markdown vs HTML for AI output — and why the real issue isn't the format but how clearly AI communicates structure and intent.

AI & LLM

Jason Yang · 23 May, 2026

Zed's Terminal Threads run Claude Code, Codex, and other CLI AI agents right inside the editor — a lightweight AI workflow without moving to VS Code or Cursor.

AI & LLM

Jason Yang · 03 May, 2026

opusplan, subagents, and /advisor all pick models in Claude Code but fire on different signals — here's what each does and why you'd run all three.

AI & LLM

Jason Yang · 16 Apr, 2026

Claude Code’s /statusline shows the project, model, and git branch right in your terminal — so you never lose track of which AI session is in which tab.

AI & LLM

Jason Yang · 12 Apr, 2026

Pair Claude Code with Codex in one workflow: how I set it up, and why using one model to generate code and another to review it leads to better results.

AI & LLM

Jason Yang · 05 Jan, 2026

Claude Code Subagents get their power from Skills. Combine reusable expertise with independent execution to build smarter automated workflows.

AI & LLM

Jason Yang · 04 Jan, 2026

Claude Code Skills package repeatable workflows so Claude automates code reviews, enforces project conventions, and keeps quality consistent across a team.

AI & LLM

Jason Yang · 04 Jan, 2026

AI code generation breaks architecture without guardrails. Learn how .claude/rules enforce consistent FastAPI layers with practical, team-driven constraints.

AI & LLM

Jason Yang · 04 Jan, 2026

Install the NVIDIA 535 driver on Ubuntu 24.04 and fix the dependency conflicts (pkgProblemResolver breaks) — clean purge, right driver, then verify PyTorch CUDA.

AI & LLM

Jason Yang · 03 Jan, 2026

Generate speaker-labeled subtitles on Ubuntu 24.04 with Whisper and pyannote.audio (CUDA GPU) — a full working setup from Python 3.10 to a combined transcript + speaker SRT.

AI & LLM

Jason Yang · 29 Dec, 2025

Use DeepSeek-Reasoner directly with Claude Code CLI via its Anthropic-compatible endpoint — no local proxy or router. Set a few env vars and switch models.

AI & LLM

Jason Yang · 28 Dec, 2025

Claude Code can add Co-authored-by metadata to your commits, causing Claude to appear in GitHub Contributors. Here’s why it happens and how to disable it.

AI & LLM

Jason Yang · 26 Dec, 2025

Claude Code forgets everything when a session ends. Install claude-mem to persist context, fix common issues, and enable long-term memory for real workflows.

AI & LLM

Jason Yang · 06 Dec, 2025

Ubuntu desktop crashing under heavy dev workloads? You're likely hitting the OOM killer. The exact swap and swappiness setup that stops the crashes.