이 가이드는 AI 에이전트를 위한 문서를 코드와 함께 레포 안에서 버전 관리하고, 유효 기간과 적용 범위에 따라 계층으로 나누라고 권한다. 에이전트는 세션마다 기억이 초기화되므로 레포에 명시된 내용만 사실로 다루도록 설계해야 한다. 상시 규칙은 CLAUDE.md에 짧게 적고 상세 내용은 docs 폴더에 두어 필요할 때만 읽게 한다. 요구사항이 바뀌면 코드보다 스펙 문서를 먼저 수정하고, 기술 결정은 ADR로 기록해 추가만 하는 방식을 따른다. 반복되는 실수는 문서 한 줄보다 테스트, 린트 규칙, Hook 같은 기계적 장치로 막는 것이 권장된다.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
An article argues that the term 'AI slop' is more slippery than it appears, since people often use it to mean anything they dislike rather than a clearly defined category. It notes that AI-generated text once had recognizable traits like stiff, formulaic phrasing, but modern language models can now mimic casual, professional, and personal writing styles convincingly. The author points out that generic, predictable writing predates generative AI, so surface features such as polished grammar or repetitive structure are weak evidence of machine authorship. The piece suggests that judgments about quality and authenticity are shaped by personal opinion and familiarity as much as by objective signals. It concludes that identifying AI content reliably requires looking beyond style to how we define quality and understanding.
With OLLAMA_MAX_LOADED_MODELS set to 1, a request for an unloaded model prompts Ollama to schedule the currently loaded model for eviction once it goes idle. If a request for the loaded model arrives during this wait and includes keep_alive, that value overrides the pending unload, so the loaded model stays in memory for the full keep_alive period. Tests on Ollama 0.40.8 with CPU-only inference showed the second model responding in about 9 seconds without keep_alive but not within 150 seconds with keep_alive set to 30 minutes. The server reports no error and /api/ps looks healthy, and the relevant scheduler messages appear only at debug level. The issue is tracked upstream as ollama/ollama#19182, with a proposed fix in #19190.
Running several Claude Code or Codex sessions on the same repository often leads one agent to overwrite another's fixes, even though neither acted incorrectly. The article proposes borrowing three restaurant-kitchen practices: giving each session its own git worktree and branch, so conflicting edits surface as readable merges. It also suggests a shared claims.md file where each agent registers its task under a file lock to prevent duplicate work. Claims carry timestamps so that stale ones from crashed or closed sessions can be reclaimed after a set period. The author notes these rules live in instruction files that agents can skip, and that the approach works best for a solo developer, becoming harder to manage across several people and machines.
A developer has built LazyDesk (SlothAgent AI), a minimalist, local-first workspace for intentional productivity, submitted to the Hacktoberfest Open-Source AI Challenge. The app combines a glassmorphism dashboard with focus timers, task and note managers, and an optional AI chat assistant powered by open-weight models run through Ollama. All tasks, notes, and preferences are stored in the browser's localStorage, so personal data stays on the user's device. The project runs with a simple Python server and requires no npm installation or external cloud subscriptions. Its creator says the single workspace reduces context switching across separate timer, planning, note, and AI apps.
Discussion (0)
Log in to join the discussion and vote.
Log in