adopt
Take a folder that already holds research — papers, notes, code, half-written drafts — into MAGI, without breaking what is…
Turn whatever the human has — a PDF, a link, a DOI, a screenshot of a citation — into raw/ sources this project can compile.
Claude Code
$ npx -y skills add misaka16384/magi --skill ingest --agent claude-code
Codex
$ npx skills add misaka16384/magi --skill ingest -a codex -y
Cursor
$ npx skills add misaka16384/magi --skill ingest -a cursor -y
Gemini CLI
$ npx skills add misaka16384/magi --skill ingest -a gemini-cli -y
GitHub Copilot
$ npx skills add misaka16384/magi --skill ingest -a github-copilot -y
OpenCode
$ npx skills add misaka16384/magi --skill ingest -a opencode -y
Windsurf
$ npx skills add misaka16384/magi --skill ingest -a windsurf -y
name: ingest description: "Turn whatever the human has — a PDF, a link, a DOI, a screenshot of a citation — into raw/ sources this project can compile." commands: ingest: "Ingest papers, articles and notes into raw/." origin: magi
Anything new is arriving: files in `inbox/`, a link or DOI in the chat, a citation the human read out.
1. Files already in `inbox/`, or a path you were given: `magi ingest auto` — it routes by type and finalizes. Stop here unless it refuses. (A run that skipped cleanup: `magi ingest finalize <the file> --project-dir .`.) 2. A link, DOI, arXiv id or citation: identify it first — that step needs you, not the pipeline. `magi ingest url "<id>" --expect "<title fragment>" --go` fetches it and lands what came through unflagged; anything flagged → step 3. 3. `magi ingest review` lists what is waiting. Show its findings to the human before committing — surface `identity-mismatch`, `figure-count-mismatch` and `image-path-not-portable` every time: the file is not what its name says. `math-damage` counts formulas that will fail `magi math check`: say the number. 4. Once the human has seen the list: `magi ingest review --approve-all` (or `--item <id> --decision approve` per row), then `magi ingest review --commit`. 5. After a batch: `magi math check --json` — a `$$` that lost its pair swallows the paragraph after it and is still valid LaTeX. Work the list with `tidy`; papers from an older arXiv-HTML ingest first get `magi math repair --dry-run`.
one sub-agent call per page. Not until a person has said yes; 10 at once.
file: name it and carry on. Report **partial** work as partial — "9 of 12".
ambiguous, stop and ask. Collect the questions the work you delegated could not ask (Invariant 4) and put them all at once.
MAGI 是一个 agent-native 的科研工作环境:人是驾驶员,LLM agent 是机体,确定性的 magi CLI 是拘束具——三者同步率越高,科研越快。它把学术论文(PDF/LaTeX)摄入、编译为 Obsidian 兼容的概念卡片项目,并用三核架构管理完整的科研状态: 进入任意项目先跑 magi sync——它输出同步率、三核状态和逐条可执行的修复提示。magi radar 是文献雷达:定时发现相关新论文,并侦察"知识上应引用我方论文却未引用"的候选。同一份 skills 通吃 Claude
Take a folder that already holds research — papers, notes, code, half-written drafts — into MAGI, without breaking what is…
Answer a question from this project's own knowledge — retrieved, read and cited, never from memory.
Write the one thing a person reads after an unattended run — its report — and the lecture notes that prepare them for it, in…
Compile raw sources into dense, interlinked reference and concept cards, and mine the concepts a compiled card left implicit.
Talk a run over with the person until it can be written down — motivation, directions, what is not worth it, when to stop, their…
Entry point for a MAGI research project: run magi next, do what it says, and call the skill it names.
Use this skill to extract structured Markdown/JSON from PDFs and document images—tables with cell-level precision, formulas as…
Estimates intracellular metabolic fluxes from steady-state carbon-13 isotope-tracing measurements using validated atom maps,…