← ClaudeAtlas

research-bootstraplisted

Use only when a repository or applied-AI problem has no valid Research Engineering canonical state, when that state is unrecoverable, or when the lead architect explicitly asks for a clean research-state re-initialization. Inspects the problem, rules, assets, and environment; creates the minimum durable state needed to begin evidence-driven research. Not a planning skill.
uukuguy/research-engineering · ★ 0 · AI & Automation · score 67
Install: claude install-skill uukuguy/research-engineering
# Research Bootstrap Turn a zero-state repository or problem into the minimum valid Research Engineering state, then hand off to `research-engineering`. The success condition is not a plan. It is that **the first technical uncertainty now has a concrete, cheap, executable next action.** `AGENTS.md` carries the precedence rules, the authority split, the invariants, and the language policy. This skill does not restate them. ## When to run - `research/ACTIVE.json` and the canonical state do not exist; - the repository has been newly adopted into Research Engineering; - existing state is declared invalid or unrecoverable by the architect; - the architect explicitly asks for a clean re-initialization. **Do not re-run bootstrap because a new session started.** If canonical state exists, the job is to resume it, not to rebuild it. Re-initializing over a live state destroys the record of what was already learned, and it is the failure this skill's narrow trigger list exists to prevent. If state exists but looks wrong, that is a reconciliation problem — say so, and do not overwrite it with a fresh skeleton. ## Procedure 1. **Inspect the problem.** Rules, competition statement or requirements, `README`, docs, repository layout, SDK, data, baselines, and whatever execution surfaces already exist. 2. **Separate external facts from assumptions.** "The SDK exposes a controller interception point" and "the rubric scores recovery separately" are facts, checkable in the SDK a