What problem does it solve?
Research and design decisions in the Marin project often lack a systematic record of prior work, leading to repeated experiments, missed contradictions, and unsupported claims. This Skill performs an explicitly requested prior-work search across internal Marin sources and external literature, then produces a cited, structured brief.
Core Features & Use Cases
- Tiered Effort Levels: Choose low (3-7 min), medium (10-15 min), or high (30-60 min) effort depending on decision cost, with defined stop rules and provenance depth.
- Ordered Source Search: Searches the current issue, GitHub issues/PRs, docs/reports, logbooks, experiment code, W&B runs, and external papers in a defined priority order, including adversarial query passes.
- Structured Output Contract: Produces a brief with an evidence map, ranked next experiments, hypothesis queue updates, and a source ledger, filed into the logbook, issue, or
.agents/projects/<slug>/research.md.
- Use Case: Before launching an expensive pretraining run, request a high-effort background research pass to gather internal experiment history and external scaling-law literature, then receive a cited brief with falsifiable next-experiment hypotheses.
Quick Start
Ask the agent to run a medium-effort background research pass on your current design question and file the cited brief in the project research file.
