Ejecuta cualquier Skill en Manus
con un clic

Ejecuta cualquier Skill en Manus con un clic

project-init

Estrellas1

Forks0

Actualizado30 de abril de 2026 a las 10:02

Use this skill to create a new MLClaw project. Triggers when the user wants to start tracking an ML model, set up a project workspace, or initialize project structure. Use for: '新建项目', 'create project', 'initialize', 'set up a new model'. Not for running inference/evaluation (use the stage-specific skills).

Instalación

Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.

Ejecutar en Manus

Fuente

Ryanshuai

Ryanshuai/MLClaw

Abrir repositorio de GitHub Ver repositorios del creador

Descarga

Ejecutar en Manus

Ocupaciones relacionadasSOC

Basado en la clasificación ocupacional SOC

Científicos de datosOcupaciones informáticas y matemáticas·SOC 15-2051

SKILL.md

readonly

name	project-init
description	Use this skill to create a new MLClaw project. Triggers when the user wants to start tracking an ML model, set up a project workspace, or initialize project structure. Use for: '新建项目', 'create project', 'initialize', 'set up a new model'. Not for running inference/evaluation (use the stage-specific skills).

/project-init — Initialize MLClaw Project

One-time project setup. Creates directory structure, project.json, and stage config templates.

Ask one question at a time — multiple questions at once overwhelms users.

Follow the Workflow State Protocol from CLAUDE.md: push on entry, update step as you progress, pop on completion.

Conversation flow

Ask: What is the project name? (lowercase, underscores, e.g., rtdetr_detection)
Resolve workspace. Default = workspace_root: line at the top of CLAUDE.md (read it once, expand ~/). Confirm with user; if they pick a different path, replace the workspace_root: line in CLAUDE.md so subsequent invocations get the new default. If --workspace <path> was passed in the user's prompt, skip the confirm and use it directly. mkdir -p whichever path won — workspaces are cheap, accidentally creating one is harmless.
Ask: Which stages to enable? Options: data, exploration, training, evaluation, inference, deployment
For each enabled stage, ask one at a time: Code source for {stage}? (git URL / local path / skip for now)

Show a full summary with all fields and values (including defaults):

Project: detection
Workspace: ~/agent_space/mlclaw/projects
Root: ~/agent_space/mlclaw/projects/detection
Created: 2026-03-16

Stages:
  data:        disabled
  training:    enabled  repo: https://github.com/...
  inference:   enabled  repo: skipped
  ...

Create? Anything to change?

User confirms -> create everything. User wants changes -> update and show summary again.

Write project.json

Fixed keys, agent only modifies values. Template: lifecycle/project.json.

Each stage has: enabled, code_path (stages/{stage}/code), code_source (source, path, branch, commit, credentials). Paths section: stages, runs_pattern, artifacts_pattern, data_pattern.

Create project

MLCLAW_ROOT=$(python <repo>/lifecycle/scripts/shared/workspaces.py tool)
python "$MLCLAW_ROOT/lifecycle/scripts/project-init/init_project.py" '<project_json_string>' "$MLCLAW_ROOT"

Creates directories, copies templates, writes .gitignore, runs git init + initial commit. $HOME-relative paths (project root, workspace, each stage's code_source.path) are rewritten to ~/-prefixed form in project.json so the file survives rsync across machines.

Fallback: if script fails, manually create directories (stages/{stage}/code, runs, artifacts, data for each enabled stage), copy JSON templates from lifecycle/, write .gitignore, git init + commit.

Clone / Link code

The unified contract (see CLAUDE.md "Code Source Resolution") is code_dir = stages/{stage}/code/_source if exists else stages/{stage}/code. Per source mode:

Git URL (code_source.source == "github"): git clone <code_source.path> into stages/{stage}/code/, record branch + commit (HEAD SHA) in project.json, remove .git so the code becomes plain files tracked by project git. No _source symlink for this mode.
Local path (code_source.source == "local"): init_project.py already creates the symlink stages/{stage}/code/_source -> expanduser(code_source.path) during creation. Do not copy — the user iterates in their own repo, copy creates bidirectional sync friction. Reproducibility comes from code_snapshot.py at run-time, not from a project-wide copy. After rsync to a new machine the symlink will dangle (it stores an expanded absolute path); recreate with ln -sfn $(python -c "import json,os;print(os.path.expanduser(json.load(open('project.json'))['stages']['<stage>']['code_source']['path']))") stages/<stage>/code/_source.
Server (code_source.source == "server"): scp into stages/{stage}/code/, no _source symlink.

Code modifications during runs stay in project git (for github/server) or are captured per-run via code_snapshot.py SHA + dirty patch (for local) — never pushed back to the original repo.

After creation

The chosen workspace was already upserted in step 2d — last_used is fresh and n_projects is now incremented to include this new project. If it wasn't yet upserted (unusual — only when the script was unavailable and you used the bootstrap fallback path), run python lifecycle/scripts/shared/workspaces.py upsert <workspace> now.
Tell user: Project created at {root}.
Downstream: offer /resources and available /{stage}-init skills for enabled stages. For stages without an init skill yet, note it's not available.

Más de este repositorio

mismo repositorio

eval-report

Ryanshuai/MLClaw

Use this skill to generate a self-contained HTML evaluation report from a completed eval run. Includes metrics, baseline comparison, per-class breakdown, training context from upstream DAG, and embedded charts. Triggers for: '生成评伋报告', 'make eval report', 'create HTML report from eval results', 'share eval results with team'. Not for running evaluation (use eval-run).

2026-04-301

eval-run

Ryanshuai/MLClaw

Use this skill whenever the user wants to execute a model evaluation run — testing a trained model on a dataset to measure metrics like mAP, accuracy, IoU, precision, recall, or per-class AP. Trigger for: launching eval runs (debug or production mode), running evaluation on a remote server, checking status of a running/crashed eval job, collecting results when eval finishes, forking a previous eval run with changed parameters (threshold, NMS, confidence, dataset split), and comparing metrics against a baseline. Also trigger for Chinese requests like "跑评估", "测一下", "跑一下eval", "对比baseline". This is the execution skill — not for initial config setup (use eval-init) or HTML report generation (use eval-report).

2026-04-301

infer-run

Ryanshuai/MLClaw

Use this skill whenever the user wants to run inference — feeding data through a trained model to get predictions, detections, embeddings, or any output. Trigger for: running inference on images/video/data, executing a model on new inputs, testing a model quickly, checking inference speed/FPS, running in debug or production mode, checking status of a running inference job, forking a previous run with different inputs or model. Also trigger for Chinese requests like "跑推理", "推一下", "跑一下模型", "测试一下", "看看效果". This is the execution skill — not for initial config setup (use infer-init).

2026-04-301

refactor-run

Ryanshuai/MLClaw

Execute one refactoring round — make changes, run verification, commit or revert. Each invocation is one round; multiple rounds across conversations until the codebase is clean. Trigger for: running a refactoring round, verifying benchmark after changes, reverting a failed round, resuming an interrupted round, checking refactoring progress, promoting refactored code to other stages. Also trigger for: "跑重构", "重构一轮", "跑一下refactor", "验证benchmark", "回退这轮". This is the execution skill — not for initial analysis (use refactor-init) or audit reports (use refactor-report).

2026-04-301

train-run

Ryanshuai/MLClaw

Use this skill whenever the user wants to execute a training run — launching, monitoring, or finalizing a model training job. Trigger for: starting a training run (debug or production mode), checking status of an in-progress training job, diagnosing a crashed run, finalizing a completed run (picking best checkpoint, applying retention), forking a previous training with changed hyperparameters, or continuing/resuming training from a prior checkpoint. Also trigger for Chinese requests like "跑训练", "开训", "继续训", "训练崩了看一下", "训练完了". This is the execution skill — not for initial schema setup (use train-init) or comparing runs (use train-compare, when available).

2026-04-301

train-tune-report

Ryanshuai/MLClaw

Use this skill to render a train-tune session as a human-readable markdown chain (chain.md). Reads all runs belonging to the session, extracts hypothesis / outcome / config diff per run, and assembles a structured narrative: headline, best-so-far curve, coverage map, decision timeline, confirmed/refuted distillation, open questions, and final recipe. Triggers when: /train-tune session closes (auto-invoked), user asks to re-render an existing session report, user asks for mid-session preview. Use for: "render the lr search report", "show me the chain.md for last train-tune session", "渲染 tune 报告", "总结一下这次搜索". This is a pure rendering skill — does not modify runs, does not run training.

2026-04-301