research.lab.notebook

Maintains a living lab notebook and reproducible failure postmortems for LLM research experiments.

Updated May 9, 2026
One-click install
npx skills add https://github.com/0h-n0/auto-research --skill research-lab-notebook-0h-n0
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: research.lab.notebook
Source: https://github.com/0h-n0/auto-research/tree/main/skills/research.lab.notebook
Command: npx skills add https://github.com/0h-n0/auto-research --skill research-lab-notebook-0h-n0

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Research experiments lose critical context when failures are discarded without recording why they happened, what was tried, and what to change next. This Skill preserves the full reasoning trail of an LLM research project—failed runs, rejected ideas, predictions versus actual results—so nothing is lost to memory. ## Core Features & Use Cases - Living Lab Notebook: Maintains a single chronological LAB_NOTEBOOK.md with phase-tagged entries, decision journals (predicted outcome, confidence, assumptions), and an auto-generated tag index for reverse lookup. - Reproducible Failure Postmortems: Auto-drafts a POSTMORTEM.md per failed run with a 3-5 hypothesis space derived from events.jsonl and error.txt, plus a reproduce.sh script and uv.lock snapshot satisfying a 7-tuple reproducibility checklist. - Rejected Ideas & Lessons DB: Archives rejected ideas with reasons and revisit conditions in 03_REJECTED_IDEAS.md, and appends generalizable lessons to a cross-project ~/.research-lessons.json database. - Use Case: During Phase 6 of an auto-research workflow, a training run fails with CUDA OOM; the Skill auto-generates a postmortem with hypothesis drafts (memory, gradient checkpointing, config drift), a blameless-framed decision section, and a frozen reproduce script. ## Quick Start Ask the agent to use the research.lab.notebook skill to record a failed experiment run and draft its postmortem for your current research project.

Frequently Asked Questions about research.lab.notebook

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

FAQPage Schema
How do I document a failed machine learning experiment?▼

Invoke the skill after a failed run and it auto-drafts a POSTMORTEM.md with 3-5 hypotheses derived from events.jsonl and error.txt, a decision section, lessons, and a reproducibility checklist. You polish the Decision and Lessons sections before release.

How to make a failed training run reproducible?▼

The skill enforces a 7-tuple checklist: code revision, config.yaml, uv.lock dependency snapshot, random seed, data version hash, hardware info, and an auto-generated reproduce.sh using set -euo pipefail and uv sync --frozen.

Does this lab notebook work outside the auto-research workflow?▼

It is designed for the auto-research plugin's 8-phase workflow and auto-dispatches at phases 3, 4, 6, and 8 by reading STATE.json. Manual invocation is supported, but it expects the .research/<slug>/ directory structure.

What is a decision journal in research notebooks?▼

A decision journal records predicted outcomes, confidence level, and up to three falsifiable assumptions before results exist. At Phase 6 the skill drafts a Predicted vs Actual comparison with surprise scores to counter hindsight bias.

Why keep rejected research ideas instead of deleting them?▼

Rejected ideas are stored in 03_REJECTED_IDEAS.md with full hypothesis, rejection reasons, and future revisit conditions. This enables later pivots and comparison studies without regenerating the original analysis.

Where are cross-project lessons stored?▼

Generalizable lessons are appended to ~/.research-lessons.json at Phase 8, with tags, context (domain, model size, task), and links back to source postmortems. The lessons-search command queries this file with jq filters.