菜单
Modal remote GPU compute. Use for CUDA workloads, fine-tuning, large training, or anything the local machine can't fit. Requires explicit user approval per plan (cost > 0).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Create or revise high-quality academic and research presentation slides in Markdown using Marp, following a story-first multi-phase workflow. Use when the user needs conference talk slides, lab or group meeting decks, thesis defense presentations, research seminars, invited talks, mixed-audience or industry research talks, OR wants to create, revise, or edit a talk deck or a specific slide of it (e.g. 'revise slide 7 of my conference talk'). Starts from an audience-facing story spine before slide titles, frameworks, screenshots, or feature lists; then applies Assertion-Evidence and Mayer multimedia principles. For teaching/lecture slides use `teaching-marp-slides`; for written documents use `paper-writing` or `scientific-writing`; for individual figures use `scientific-visualization`.
Create or revise lecture/teaching slides in Markdown using Marp for upper-undergraduate and graduate courses across any academic discipline. Use when the user needs classroom lecture slides, course materials, tutorial slides, or workshop sessions — including a request framed as creating or editing a slides Markdown file (e.g. a '*_slides.md' in a course or week folder) for teaching a class — or wants to revise existing teaching slides. Enforces Cognitive Load Theory, worked-example effect, scaffolding, and retrieval practice principles. Produces both a lecture version and a handout version. For research/conference talks use `academic-marp-slides`; for written documents use `paper-writing` or `scientific-writing`.
Strategically revise an existing CS/AI/Systems conference paper draft for resubmission or submission readiness. Focuses on framing diagnosis, claim crystallization, narrative unification, evidence strengthening, reviewer defense, and venue-aware polish.
Write publication-ready ML/AI/Systems conference papers (NeurIPS, ICML, ICLR, ACL, AAAI, COLM, OSDI, NSDI, ASPLOS, SOSP). Use when the user requests a conference manuscript, needs conference-specific templates/checklists, or format conversion between venues. For journal articles or technical reports, use scientific-writing instead.
Core skill for high-quality research strategy, topic exploration, idea generation, scope/framing critique, contribution assessment, and mentor-style research direction refinement. Use when the user asks whether an idea is worth pursuing, wants brainstormed research directions, needs to pivot, sharpen a paper/proposal scope, evaluate novelty, identify the strongest objection, or turn a vague area into a concrete, testable research plan. Produces a small number of high-bar candidates with premise audits, nearest alternatives, decisive experiments, and kill criteria.
Prompt guidance for the generate_diagram tool — how to describe different scientific diagram types (flowcharts, pathways, architecture, circuits, networks, conceptual) so the tool produces publication-grade output on the first iteration.
| name | compute-modal |
| description | Modal remote GPU compute. Use for CUDA workloads, fine-tuning, large training, or anything the local machine can't fit. Requires explicit user approval per plan (cost > 0). |
| category | Compute |
| tags | ["compute","modal","gpu","training","remote"] |
| triggers | ["modal","train large model","fine-tune","gpu cloud","cuda"] |
| license | MIT |
| metadata | {"skill-author":"Research Copilot"} |
Modal backend (backend: 'modal', toolPrefix: 'modal') shells out to the modal CLI to run scripts on Modal's remote GPU infrastructure. Bills the user's Modal account. Every plan requires explicit user approval in the Compute tab before execution — no auto-approval.
The user must have:
pip install modal (or uv pip install modal) somewhere on PATHmodal token new once, OR MODAL_TOKEN_ID + MODAL_TOKEN_SECRET in Settings → API KeysIf either is missing, list_compute_backends will show available: false with missingRequirements. Surface the message to the user — they need to install Modal CLI / configure credentials before Modal plans can execute.
list_compute_backends().compute_plan({ backend: 'modal', command, script_path, task_description }) — runs a planning sub-agent that reads the script and extracts the declared Modal image + GPU configuration, then estimates cost.requires_approval: true and message — tell the user a Modal plan is awaiting approval in the Compute tab. Do NOT call modal_execute until they approve.modal_execute({ plan_id }) — submits the run. If approval is still pending, returns waiting_for_approval: true; if user rejected, returns rejected: true with the rejection comments.modal_status / modal_wait. Stop with modal_stop.import modal
image = (
modal.Image.debian_slim(python_version="3.11")
.uv_pip_install("torch==2.8.0", "numpy", "pandas")
)
app = modal.App("research-copilot-job")
@app.function(gpu="A10G", image=image, timeout=60 * 60)
def run_job():
import torch
# Heavy imports + work inside the function body
print('##PROGRESS## {"step": 1, "total": 3, "phase": "setup"}')
# train / evaluate / process here
@app.local_entrypoint()
def main():
run_job.remote()
The planner extracts the GPU declared by @app.function(gpu=…) and feeds it to the cost estimator.
| GPU | Use case | ~Rate ($/hr) |
|---|---|---|
| T4 | Small inference / light eval | 0.59 |
| A10G | Default for ordinary PyTorch/TensorFlow training | 1.10 |
| A100 | Larger models / memory-heavy training | 3.72 |
| A100-80GB | Larger context / longer sequences | 4.28 |
| H100 | Very large or perf-critical jobs only | 8.10 |
| L4 | Cheap inference, video / image workloads | 0.80 |
When multiple GPUs are declared, the cost estimator picks the highest and adds a warning.
compute_plan returns a cost_estimate with estimatedTotalUsd, hourlyRateUsd, and coverage: 'lower_bound' (only GPU time is modeled; CPU/RAM/idle-container are not). The user's configured Modal cost threshold (Settings → Compute) auto-kills runs whose elapsed cost exceeds the threshold.
Mention the estimated cost when surfacing the plan to the user so they know what they're approving.
Same convention as local: print ##PROGRESS## {...} on stdout; the runner extracts them via modal_status/modal_wait.
uv_pip_install over pip_install when available — it's faster.torch==2.8.0, not just torch.@app.function(gpu=…), not just on the image build, so cost estimation finds it..add_local_dir of the whole workspace — only the directories the script reads at runtime.