skills.rest
zjunlp
zjunlpOFFICIAL
Explore all skills by this author →
Website𝕏@ChenHuajunGitHub

next-round

Archives completed research round outputs and drafts the next round's task.md from exploration memory.

75|7|Updated May 2, 2026
Search Tags:#claude code#workflow orchestration#research automation#mechanistic interpretability#task.md#round archiving#exploration memory
One-click install
npx skills add https://github.com/zjunlp/Mechanist --skill next-round-zjunlp
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: next-round
Source: https://github.com/zjunlp/Mechanist/tree/main/skills/next-round
Command: npx skills add https://github.com/zjunlp/Mechanist --skill next-round-zjunlp

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Multi-round autonomous research runs leave outputs scattered at the project root and require manual bookkeeping to decide what to explore next. This Skill handles the between-round transition: it archives the finished round's outputs and drafts a starting task.md for the next round based on what has already been explored. ## Core Features & Use Cases - Round Archiving: Moves the just-finished round's outputs (idea-stage, refine-logs, verify, review-stage, ledgers, caches) into rounds/round_<N>/ using a keep-list approach, with a hard-stop guard against clobbering existing archives. - Memory-Driven Recommendations: Reads the orchestrator-owned research_memory.json to recommend whether the next round should explore a new behavior or a new mechanism for an existing behavior, based on each behavior's status (established, inconclusive, not-established). - Draft task.md Generation: Writes a draft task.md for the next round carrying the research direction, recommended behavior-source/mechanism parameters, and a human-readable summary of prior findings. - Use Case: After an /auto round concludes that a behavior is established but a mechanism direction remains untried, run /next-round new-mechanism <behavior-id> to archive the round and get a pre-filled task.md that keeps the same stimuli and activation caches for the follow-up run. ## Quick Start After your /auto round finishes, run /next-round with new-behavior or new-mechanism plus a behavior id to archive the round and draft the next task.md.

Frequently Asked Questions about next-round

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I start a new research round after an /auto run completes?▼

Run /next-round with either new-behavior or new-mechanism <behavior-id>. It archives the finished round's outputs into rounds/round_<N>/ and drafts a new task.md with recommended behavior-source and mechanism parameters, which you edit before running /auto again.

What is the difference between new-behavior and new-mechanism intents?▼

new-behavior starts exploration of a brand-new phenomenon and archives data/ and cache/ since new stimuli are needed. new-mechanism keeps the same already-found behavior and its data/ and cache/ directories, trying a different mechanism direction via behavior-source: given.

Does next-round modify the research_memory.json file?▼

No, it only reads research_memory.json for context and recommendations. The file is owned exclusively by /auto, which writes it at the end of each round; next-round never edits or moves it.

What happens if the target archive directory rounds/round_<N> already exists?▼

The skill hard-stops with exit code 3 and refuses to archive or draft task.md. This clobber guard prevents overwriting an existing archive and avoids leaving a drafted task.md next to un-archived prior-round outputs.

What should I do when a round ends with an inconclusive behavior?▼

The skill recommends re-validating the same behavior with behavior-source: given-validation, which re-opens the plan with a stronger M0 test. If stronger data is needed, delete data/ and cache/ before running /auto so fresh stimuli are generated.

Related Skills

Explore More Education & Research Skills→
duypham2801duypham2801

research-pipeline

Orchestrate research from idea discovery through implementation, experiments, and review.

Community
Advanced
zhuyingqinzhuyingqin
1

research-pipeline

Automate the end-to-end research lifecycle from idea discovery to paper writing.

Community
Advanced
KwongFukKwongFuk

auto-review-loop

Automate iterative research review rounds with scoring and fix implementation.

Community
Advanced
fitlab-aifitlab-ai
84

analyze-task

Generate requirements analysis documents from task.md files.

Official
Advanced