Agent skill · research science · aradotso
auto-deep-researcher-24x7
auto-deep-researcher-24x7 — an agent skill for coding agents.
Why this skill is useful
Adds a fully autonomous deep learning experiment agent with executable scripts for managing experiments and monitoring GPU usage.
What it needs
Requires git, python installed locally. Requires anthropic, openai account access. About 8k tokens when loaded. Last updated 2026-07-09. 70 stars on the source repository.
What this skill does
bash git clone https://github.com/Xiangyue-Zhang/auto-deep-researcher-24x7.git cd auto-deep-researcher-24x7 pip install -r requirements.txt bash For Anthropic (Claude) export ANTHROPICAPIKEY=yourkeyhere For OpenAI (Codex) export OPENAIAPIKEY=yourkeyhere bash mkdir my-experiment cd my-experiment markdown Goal Train a ResNet-50 on CIFAR-100 to reach 80%+ top-1 accuracy. Codebase Create the training code from scratch in PyTorch. What to Try Start with a basic ResNet-50 baseline (lr=0.1, SGD, cosine schedule). If accuracy < 75%, improve optimizer and learning rate schedule. If accuracy is 75-80%, try data augmentation (CutMix, AutoAugment). If accuracy > 80%, stop and write final report. Constraints Use GPU 0 only. Max 100 epochs per run. Do not change the backbone architecture. bash Inside Claude Code or Codex CLI: /auto-experiment --project /path/to/my-experiment --gpu 0 my-experiment/ ├── PROJECTBRIEF.md # ← You write this (required) ├── HUMANDIRECTIVE.md # ← Optional: override next cycle direction ├── config.yaml # ← Optional: override defaults └── workspace/ # ← Auto-created by agent ├── MEMORYLOG.md # Rolling memory of results and decisions ├── experiments/ # Per-run code, configs, logs ├── progresstracking/ # Local text notes (if no Obsidian) │ ├── Dashboard.txt │ └── Daily/YYYY-MM-DD.txt └── results/ # Parsed results per experiment bash Start autonomous experiment loop /auto-experiment --project /path/to/project --gpu 0 Check current status (goal, best result, cycle count, running jobs) /experiment-status Generate structured progress report /progress-report Manually sync Obsidian notes (if configured) /obsidian-sync Run a single THINK-EXECUTE-REFLECT cycle (manual trigger) /run-cycle --project /path/to/project yaml config.yaml — place in your project root agent: model: "claude-opus-4-5" # or "gpt-4o", "claude-sonnet-4-5" maxcycles: 50 # stop after N cycles (0 = unlimited) cycletimeouthours: 6 # max hours per experiment before abort reflectonfailure: true # still …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills aradotso/auto-deep-researcher-24x7