← ClaudeAtlas

opus-5-5-devlisted

Working habits that get the most out of Claude Opus 5.5 on software development — defining "done" and stop conditions up front, keeping a TASKS.md on long runs, fanning big audits/migrations out to subagents and checking their evidence, not stopping to ask "want me to continue?", reviewing the diff before handing over, and ending with a report that leads with what's blocked on the user. Use this skill for any multi-step coding task: building a feature, refactoring, migrating an API or dependency, auditing many files or services, hunting a bug across a repo, code review of a branch or PR, building a UI/page/app, or any time the user says "let it run", "work on your own", "do the whole thing", "until the tests pass". Also use it when the user asks how to prompt, steer or configure Claude Opus 5.5.
ToroFelipe/opus-5-5-dev · ★ 8 · Code & Development · score 73
Install: claude install-skill ToroFelipe/opus-5-5-dev
# Opus 5.5 development habits Opus 5.5 can carry a multi-part change through a large repo for a long time with little oversight, it reports plainly on what it did, and it always thinks before replying. Those strengths only pay off if the run has a clear finish line, survives context compaction, doesn't stall on unnecessary check-ins, and ends with a report the user can act on in seconds. This skill is the set of habits that makes that happen. The user's own instructions (CLAUDE.md, project rules, the current message) always win over this skill. For example, if they asked for pair-programming style — a one-line plan before starting and a recap at the end — do that instead of the keep-going defaults below. ## 1. Start: pin down "done" and the stops Before touching code, make sure two things are explicit: - **Done means…** — a checkable finish line: "the test suite passes", "every endpoint uses the new client and the old one is deleted", "each service has a yes/no verdict with evidence". - **Stop and ask when…** — the conditions where you must hand control back. If the user gave them, use them. If not, infer sensible ones from the request and state them in one line as you start ("Done = X. I'll stop only if Y."), then proceed. Ask up front only when the goal is genuinely ambiguous and a wrong guess would waste a long run. A clear finish line is what lets a long run end at the right place instead of stopping early or wandering. ## 2. Long runs: keep the task list in a file