Work
procdork
Across 24 completed adversarial workflows and 313 MCP calls, the harness still produced grounded answers with no invented supplier scores; 90.1% of input came from cache.
Bright MCP
An outcome-shaped MCP completed 14 of 15 five-turn workflows, reached 7.59/10 blind quality, and won judge preference 9–4 with two ties.
Signals
Turns scattered supplier emails, PDFs, and review clicks into an evidence DAG whose edges preserve status, confidence, and provenance—so missing context stays visible.
Murdock
One model pass drafts source-linked findings; deterministic code turns them into conflicts, review gates, and durable actions, keeping legal state out of the chat model.
DROID Loop
Streams multiview robot episodes through SigLIP, IsolationForest, HDBSCAN, and temporal-jump detection, then compresses anomalous frames into incidents humans can validate for the next training cycle.
mobisim
Turns a raw GLB into a structural DAG and semantic overlay, grounding voice commands in live scene state while a deterministic boundary applies accepted mutations and rejects unsupported ones.
Agent skills
Stop the slop. Give coding agents sharper operating procedures.
- premortem
Pressure-test risky technical plans before rewrites, migrations, cutovers, or AI-system changes.
npx skills add https://github.com/dunkeln/skills --skill premortem - postmortem
RCA engineering failures into evidence-backed cause chains and fix paths, distilled from production SRE agent workflows.
npx skills add https://github.com/dunkeln/skills --skill postmortem - 4x4
Stop agents from colluding around the first plausible fix by making four lanes compete under different payoffs.
npx skills add https://github.com/dunkeln/skills --skill 4x4