← ClaudeAtlas

research-engineeringlisted

Use for exploratory applied-AI research where the system, algorithm, architecture, evaluation surface, or research environment is still evolving. Operates the research loop end to end — resume state, pick the cheapest valid evidence, run it, record it, decide what is next. Not for delivery work on an already-promoted baseline.
uukuguy/research-engineering · ★ 0 · AI & Automation · score 67
Install: claude install-skill uukuguy/research-engineering
# Research Engineering Operate as the lead architect's technical research partner. The stable abstraction is **evidence** — not a runnable candidate, not an evaluator. On most days neither exists yet. `AGENTS.md` carries the precedence rules, the authority split, the core invariants, the state file list, and the language policy. This file carries the loop, the block contract, and the router. Where the two disagree, `AGENTS.md` wins. ## Mode Research Mode is the default: disposable code, minimal correctness, run early, cheapest evidence first. A failure that produces information is a good result. Integration Mode applies only to a mechanism the architect has promoted. It re-enables planning, selective TDD, regression, and cleanup for that mechanism and nothing else. Know which mode you are in before you touch code. Applying Integration Mode discipline to a hypothesis mutation is the most expensive mistake available here. ## Resume comes first (Normative) Run this at the start of every session, before proposing any new work. 1. If `research/ACTIVE.json` is absent, stop — use `research-bootstrap`. Do not invent prior state. 2. Read `research/ACTIVE.json`, `CURRENT.md`, `ARCHITECT.md`, `BOUNDARIES.md`, `ENVIRONMENT.md`, and the parts of `FINDINGS.md` that bear on the active question. 3. `git rev-parse HEAD`, current branch and worktree, `git status --short`, and the diff for the files ACTIVE says should be dirty. 4. If `ACTIVE.status != idle`, reconstruct that e