Seonglae Cho절대로 타협하지마
claude code 지멋대로 시킨거 안따르고 타협해서 쉬운 길로 돌아가거나 가짜 데이터 사용
- exact correct way to grep previous chat hsitory schema
Research hyperparameters should be polynomial so that every combination could be covered, but this is not a computationally feasible way to explore. Use heuristics and polynomial design instead.
실험 설계를 기본적으로 infinity run 으로 하되 언제 끝내도 의미있게 결과 일정기간마다 저장하거나 report 완성형 되도록. 다만 inifnite 라도 항상 똑같은거 반복이 아니라 일정 스텝 이후에는 새 run 으로 다른거 exploration 하거나
Research experiments design with control groups
비교군으로 증명가능한거만 말해 . 다른 config 나 code 다 같은데 하나만 다른 run 을 해야함. to prove and verify. Claims require control groups.** Without comparison runs, we cannot attribute effects.
| Claim | Requirement |
| "X improves accuracy" | Run with X vs run without X, all else equal |
| "X is better than Y" | Runs with X vs runs with Y, all else equal |
| "X is key factor" | Multiple runs varying only X |
Single Variable Rule: Change ONE hyperparameter at a time.** Multiple changes make it impossible to attribute effects.
Hook
- git hook 사용
- push 전 build/test
- commit 전 format/lint
plugin/mcp 사용
- notion
- figma
- github
- vercel
- playright
- claude hook 사용
Personalization
Put this information in claude.md, and store each role in its respective md file using - - - metadata
- state.md - stateful information which stores information in the git history for experiment trajectory tracking
- need to be commit when edited
- claude.md - ROUTING to other md memory|understading|implementation mapping documents, permanent and general rules
- Commit frequently
- In PRs, focus only on the relevant feature and keep PRs independent by pushing only to their respective branches to maintain PR intention
- use only english by myself
- project specific rules
- project type specific rules - research, web, application etc.
- metainfo such as subagent, hook etc
- always run the experiments and use compute to explore and make variants, update state.md with results
- results.md
- at least triplet tuple map
- for each experiment result, script path, result json path, and description
- roadmap.md
- tasks
- experiments
- todo list
- hosts.md ignored
- local path
- shell, os, gpu
- example command
- package manager such as uv
- JAX, PyTorch, etc. OS installation methods
- Special notes for each node
- memories/
- concept
- method
- jargon
- baseline.md
- literature.md
- findings.md
- results.md
- experiments.md
Typescript
- vercel 처럼
- typing
- hook is important
- 보통 reserach일경우 결과를 serving하는 page 로 url argument 와 router 잘사용해서 접근 바로 되게 클릭하면 reactive 바뀌고 같은 페이지내에서
Research Python
- JAX or for torchjaxtypingpatrick-kidger • Updated 2026 Mar 9 17:59
- typing
- code snapshot
- like DeepMind, Anthropic, OpenAI
- portable options
Remote hosts
requirement: github, huggingface login using token, clone repo, uv install, .env
- pull before experiment
- rsync results after experiment ends
- if code is not compatible with host OS or environment, do not commit from remote
- in other words, preferably code: local → nodes, experiment results: nodes → local
memories/ folder
- generate for each {concept}.md markdown file
- concept can be a paper method name or notion of the domain or implementation
- specific terminology and jargon for this repo or project should be here
- each memory contains not only a conceptual explanation. it should be a guideline to actionable requests that map a user task query to exact implementation location to be fixed
History
- use Claude Code Hook to generate a chat history before context limit triggered summarization
- just leave user messages somewhere and put chat history folder to track and retrieve