Retrospective

Capture experiment learnings into a reusable skill registry. The failed attempts table is the most valuable artifact.

Overview

The Retrospective skill captures what was learned during a session — what worked, what failed, exact hyperparameters, and key insights — into a structured registry entry that can be searched and surfaced by future investigations.

This is not for mid-session notes (use investigation scratch/ for that). Retrospectives are end-of-session artifacts that preserve knowledge across /clear boundaries.

When It Triggers

The auto-suggestion is a recommendation, not a gate — the user can decline:

"We've been working for a while without capturing learnings.
Consider running a retrospective before /clear — the failed
attempts table is the most valuable artifact, and it won't
survive a context reset."

The Experiment Registry

Registry entries are stored in scratch/registry/{YYYY-MM-DD}-{experiment-short-name}/SKILL.md and follow a strict format:

Setup Section

Each entry records the exact experimental configuration:

• Model: architecture, size, variant
• Dataset: name, size, preprocessing
• Framework: JAX/PyTorch/etc, key libraries and versions
• Hardware: GPUs, memory, distributed setup
• Config: path to config file or inline key parameters

The Failed Attempts Table

This is the most valuable part of any registry entry. It prevents repeating mistakes across /clear boundaries and across team members.

Other Sections

• What Worked: Specific things that worked with exact hyperparameters, copy-paste ready
• Key Hyperparameters: Exact values in a table with the reasoning for each choice
• Findings: Key insights from the session
• Next Steps: What to try next based on results

Registry Structure

The registry lives in scratch/registry/ (gitignored):

scratch/registry/
  2026-03-15-rvq-depth-sweep/SKILL.md
  2026-03-18-fsq-codebook-size/SKILL.md
  2026-03-22-commitment-loss-scaling/SKILL.md

How It Feeds Back

Registry entries are searchable by the investigation skill. When a new investigation starts, matching entries are automatically surfaced based on the description and tags in each entry. This creates a feedback loop:

1. Session A runs experiments and captures a retrospective
2. Session A ends (or /clear happens)
3. Session B starts a new investigation in a related area
4. The investigation skill finds matching registry entries and surfaces them
5. Session B avoids repeating Session A's failed attempts

Key Principles