Route each unit of work to the CHEAPEST resource that fits — the experiment engine (0 Claude tokens) for compute, weaker Claude models (haiku/sonnet) for mechanical/moderate agent work, Opus only for genuine judgment. Apply BY DEFAULT before dispatching any…
The single project realignment skill for the neural-simulator repo — run it whenever working here and Claude has drifted, stalled, declared a wall, is about to wrap up or hand off, or just needs to re-anchor priorities and workflow. Combines the mission +…
The neural-simulator project's PLAIN-LANGUAGE task board (Vikunja at vikunja.dant123.com) — the owner's monitor + Claude's durable next-action source-of-truth that survives context compaction (supplements RAG/findings, does not replace them). AUTONOMOUS: read…
Use when code OR findings change may have made docs stale — and run it SAME-CYCLE when a committed finding changes a wall/gap STATUS, the CURRENT FRONTIER, or a next-action. Two layers. (1) MECHANICAL drift — line counts, class line numbers,…
Before COMMITTING or REPORTING a positive result — a GO, a "surpass", a milestone, an "it works / it's inert / byte-identical / tests pass" — run independent ADVERSARIAL skeptics that each try to REFUTE it from a distinct angle, then commit only what survives…
Periodically review what's WORKING and what's RECURRINGLY FAILING in this project's workflows, then make INCREMENTAL updates to the applicable skills (or a one-line CLAUDE.md pointer) so the workflows compound instead of re-learning the same lessons. Run when…