writing-driven-autoresearch

Sync an Overleaf project via git — pull the latest state, apply generated artifacts (LaTeX tables, figures, prose, bib entries) into the .tex sources, commit, and push back to Overleaf. When writing prose, matches the tone of the existing manuscript and a cached personal "voice profile" learned from the user's past Overleaf projects (personal voice only — venue formatting always follows the current paper). Always pulls first (rebase) because the user may have edited on the Overleaf web UI in parallel; never force-pushes. Triggered by requests like "push to overleaf", "apply results to the paper", "update the draft", "overleaf에 반영해줘", "논문에 push해줘", "내 문체로 써서 반영해줘", "내가 평소 쓰던 톤으로 써줘". Requires the Overleaf git URL and a git auth token on first use; building the personal voice profile additionally requires the user to supply the git URLs of past Overleaf projects (one-time).

Manage a herdr multi-agent terminal topology, local-only or with ssh remotes. Installs herdr if missing, discovers and remembers the current topology (servers, sessions, panes, agents, ssh remotes), builds/revives pane layouts, bridges persistent remote (ssh) agents into the LOCAL agents tab — which herdr does NOT do natively — relays messages between agents/panes, and watches + auto-judges (approve/deny/dismiss) AND nudges agents' permission/idle prompts against a user-defined policy, including agents running locally on one machine. Use when the user says "/herdr", "herdr", "set up herdr", "revive the panes", "bring the agents back", "wire my remote agents", "watch my local agents", "sync the agents tab", "watch the agents", "judge/approve the permission prompts", "nudge the stuck agent", "go live / dry-run", or asks to inspect or restore a herdr agent view.

Use when preparing Ralphthon research specifications or 연구 명세, automated experiments, a Track 1 short paper or 논문, a Track 2 review agent or 리뷰, or an evidence-backed submission under the event deadline.

Welcome and orient Ralphthon @ICML attendees, operators, sponsors, and QR/POP traffic. Use when the user asks for Ralphthon ICML onboarding, attendee welcome copy, event orientation, QR or POP copy (QR코드, POP), guestbook engagement, hackathon (해커톤) context, or skill surface framing for Auto Research (오토리서치) and world model (월드모델) work.

Use when signing up or 가입 for current VESSL Cloud, installing or authenticating vesslctl, checking organizations and teams, or preparing a first billable GPU Workspace, Job, or 기본 세팅.

Use when signing up or 가입 for W&B Cloud, configuring a personal API key, checking authentication, or creating a first experiment-tracking run or 실험 추적 project.

Analyze live W&B Models and Weave project data using the W&B MCP server, including runs, histories, configs, traces, evaluations, artifacts, registries, comparisons, diagnoses, and reports. Use when the user asks what happened in a W&B project, which run or model is best, why a run or trace failed, how evaluations performed, what changed between versions, or to save an evidence-backed analysis as a W&B Report.

Add W&B Models experiment tracking and guarded alerts to iterative experiment workflows, including Claude Code or Ralph-loop autoresearch. Use when the user asks to instrument experiments, create one W&B run per candidate, group related runs, record hypotheses and code provenance, log metrics and summaries, detect failed or stalled experiments, or validate that runs are comparable and reproducible. Use wandb-project-analyst for analysis of data already stored in W&B.

Add W&B Weave tracing to an existing LLM, agent, RAG, or tool-using application with minimal behavioral change. Use when the user asks to integrate Weave, instrument code with @weave.op, trace model or tool calls, choose trace boundaries, account for framework auto-patching, protect sensitive trace data, debug missing traces, or verify a real trace tree. Keep this skill focused on tracing; route evaluations, assets, and general W&B documentation elsewhere.

Turn a world-model, 월드모델, simulation, predictive-state, embodied-agent, or intervention-model idea into ranked falsifiable research questions and a compact experiment handoff. Use when selecting a world-model research direction, deciding whether an idea needs a world model, defining an action-conditioned or counterfactual comparison, or preparing Ralphthon work for Auto Research Track 1 or Track 2.