Picard1701ent
- 13 skills
- 0 followers
- 9 hours ago last updated
- ▌ Paper Ideation · picard1701ent bundleStructured reasoning workflow for research/paper work: (1) propose paper ideas from a topic/problem, (2) summarize an existing method/paper, (3) identify gaps/limitations in prior work and turn them into testable research directions, (4) draft positioning/contributions plus an experiment plan. Use when the user asks (often in Chinese/English) for 论文idea/选题、研究gap、related work梳理、方法总结、创新点与实验设计。
- ▌ Plan Md Ledger · picard1701ent bundlePersist the latest approved plan into a repo-local markdown ledger, auto-save it to `plan.md` when execution begins, archive replaced plans under `old_plan/`, and keep progress updated with strikethrough instead of deletion.
- ▌ Remote Ec2 Gpu2 · picard1701ent bundleUse when the user wants Codex on this machine to connect to, inspect, or operate the peer EC2 server over SSH. Covers connectivity checks, remote command execution, tmux-based long runs, log inspection, and file transfer conventions for the remote host aliased as gpu2.
- ▌ LLM Benchmark Ops · picard1701ent bundleRun and monitor LLM benchmark matrices with a smoke-to-full workflow. Use for multi-model, multi-benchmark execution with tmux/nohup, structured logs, and resumable operations.
- ▌ LLM Run Preflight · picard1701ent bundleValidate expensive or long-running LLM pipelines before launching the full run. Use when a workflow has multiple stages such as data generation, training, evaluation, or checkpoint selection and failures would waste significant time, compute, or API cost.
- ▌ Repo Research Sync · picard1701ent bundleMaintain a project-local research.md while reading a repository, folder, or codebase for analysis or planning. Use when Codex must inspect multiple files or infer project structure, behavior, or intent before proposing a plan, especially when the user wants to verify Codex's understanding and prevent planning drift.
- ▌ Repro Orchestrator · picard1701ent bundlePlan and supervise end-to-end model/paper reproduction projects. Use when a user asks to reproduce a method, define an experiment roadmap, manage resources/risks, or produce a final reproducibility report.
- ▌ Task Memory Ledger · picard1701ent bundleMaintain durable task memory for complex, multi-step, or multi-folder work. Use when Codex needs to track long-running progress across sessions for reproductions, baselines, experiments, implementations, audits, or any complex task with multiple units that each need their own memory.md plus a root-level memory.md and memory_index.md.
- ▌ Experiment Governance · picard1701ent bundleStandardize experiment observability, artifact structure, metric schema, and incident handling. Use when runs must be auditable, comparable, and reproducible across models and benchmarks.
- ▌ Github Selective Push · picard1701ent bundleCommit and push only specific user-selected files from a local Git repository to a GitHub remote, without accidentally staging unrelated work. Use when Codex needs to set up Git identity or SSH access, verify the remote, stage a narrow file list, handle ignored paths with force-add when explicitly requested, resolve first-push divergence with fetch or rebase, and push the result to GitHub.
- ▌ Sandbox Aware Test Ops · picard1701ent bundleRun tests, smoke checks, training jobs, and benchmark pipelines when sandbox, network, GPU, package, or environment issues may block normal execution. Use when a task involves pytest/unittest, experiment scripts, conda environments, API-backed evaluation, model downloads, CUDA, or other commands that may need escalated permissions or a different runtime than the default shell.
- ▌ Long Run Experiment Recovery · picard1701ent bundleRecover, resume, and triage long-running experiment jobs with minimal wasted work. Use when detached runs stall, fail mid-pipeline, lose terminal visibility, or must be resumed safely from existing logs, progress files, summaries, or checkpoints.
- ▌ Reproduction Boundary Governance · picard1701ent bundleEnforce strict task, data, method, and evaluation boundary checks for paper or model reproduction before implementation or long runs. Use when reproducing papers, adapting methods to available data, classifying faithful versus adapted baselines, or requiring explicit user approval for any adaptive implementation.