workspace-improver

Workspace Improver

When to Use

Activate for autonomous workspace improvement sessions. This skill implements the autoresearch pattern — agents iteratively improve the workspace through scoped experiments with clear metrics.

Activate when:

• Operator says "improve the workspace" or "run autoresearch"
• During scheduled autonomous sessions
• When MASTERPLAN.md activations list improvement areas
• As part of the ongoing learning phase (Phase 3 of bootstrapping)

The Autoresearch Loop

Inspired by karpathy/autoresearch:

MASTERPLAN.md          = program.md     (human directions)
data/*.yaml + skills/  = train.py       (editable surface)
HEARTBEAT.md metrics   = val_bpb        (evaluation criteria)
memory/YYYY-MM-DD.md   = experiment log (results)

Procedure

Step 1: Read Directions

1. Read MASTERPLAN.md — understand current activations and research directions
2. Read HEARTBEAT.md — understand active tasks and priorities
3. Read memory/ (last 3 days) — understand recent work and avoid repetition

Step 2: Identify Improvement

Scan for improvement opportunities in priority order:

Priority 1: MASTERPLAN Activations

• Are there unchecked activations in MASTERPLAN.md?
• Pick the highest-priority one that can be scoped to a single session

Priority 2: Data Gaps

• Are there TBD, null, or placeholder values in data/*.yaml?
• Are any registries significantly incomplete?
• Are dates stale (last_updated > 7 days ago)?

Priority 3: Knowledge Gaps

• Run npm run compile:knowledge to refresh the knowledge extract and compiled pages
• Run npm run lint:knowledge — address any hard failures
• Check for domains with coverage: "none" or "partial" in data/knowledge-manifest.yaml
• Check for <!-- LLM-SYNTHESIZE --> markers — invoke knowledge-curator in Compilation Mode to synthesize
• Are there sources in data/sources.yaml that haven't been processed?
• After changes, run npm run knowledge to validate the full pipeline