Orchestrates the claim-stage pipeline producing research proposals and experiment plans.
npx skills add https://github.com/zjunlp/Mechanist --skill auto-claim-zjunlp
Please help me install this Agent Skill. Skill: auto-claim Source: https://github.com/zjunlp/Mechanist/tree/main/skills/auto-claim Command: npx skills add https://github.com/zjunlp/Mechanist --skill auto-claim-zjunlp
given (faithfully capture claims from task.md), given-validation (capture plus an M0 phenomenon-validation gate), and discovery (mine a new behavior and run full ideation with novelty and impact checks). - Two mechanism modes: discovery (system routes the mechanism family via /mechanism-explore) or given (user-named method committed directly as CHOSEN_FAMILY). - Structured outputs: Always emits idea-stage/IDEA_REPORT.md plus refine-logs/FINAL_PROPOSAL.md and refine-logs/EXPERIMENT_PLAN.md, with resume support and optional compact summaries. - Use Case: A researcher writes a task.md describing a known LLM behavior and a target mechanism method, then runs the pipeline to reproduce the claim stage and receive a strict-fidelity experiment plan ready for the experiment stage. ## Quick Start Run the auto-claim skill with your research direction or a task.md file, for example by asking to reproduce the claims in task.md with behavior-source given and mechanism given.High-intent search queries and answers about installing and using this skill.
Invoke the skill with a research direction or a task.md file in the project root. It runs literature survey, claim production, and refinement phases, producing IDEA_REPORT.md, FINAL_PROPOSAL.md, and EXPERIMENT_PLAN.md.
Given captures the behavior from task.md and assumes it holds with no M0 gate. Given-validation captures it the same way but opens the experiment plan with an M0 phenomenon-validation gate. Discovery mines a new behavior and runs full ideation with novelty and impact checks.
The strict marker is stamped only when BEHAVIOR_SOURCE is given and MECHANISM is given, the reproduction combination. It binds the exact models, datasets, and data sizes from task.md with no cost-driven downscaling.
Yes, set RESUME to true so each phase checks whether its primary artifact exists and is non-empty, skipping completed phases. Resume never deletes pre-existing files; delete an artifact manually to force a phase to re-run.
The pipeline stops with an error requiring either a named mechanism method or family, or an explicit behavioral-only declaration. A behavioral-only declaration sets chosen_mechanism to not-applicable and skips mechanism milestones.
agentic-inSearch arXiv by topic, author, or identifier and extract metadata with links.
Biswajit56546Track BSPDN milestone gates and evidence artifacts toward PPA targets.
xqinagAutomate robotics idea discovery with literature surveys and novelty verification.
OpenLAIROrchestrates a literature-to-validated-idea pipeline with novelty checks, pilot experiments, and reviewer feedback.