← ClaudeAtlas

implementation-scoutlisted

Find proven, real GitHub source code that already solves a specific implementation problem — the exact file and function, cited to a stable version, plus where a fix belongs versus adjacent layers that look plausible but aren't it. Triggers on "I need to implement X", "how do people implement X", requests for a reference implementation, or real code for a pattern (auth, protocol, algorithm, data structure, API integration, distributed-systems trust/federation, operational incidents). Also triggers on "why is X happening" / "how does X work" about the user's own system when answering well implies knowing where in real code to look. FINDS existing implementations — never generates new code or demo environments unless explicitly asked as a follow-up. Every claim must be verified via live tool calls in the same run — never invent repos, stars, paths, functions, line numbers, licenses, or claims of testing anything. Evidence precision depends on what's reachable this run, not prior-run assumptions.
shouvik12/implementation-scout · ★ 1 · AI & Automation · score 69
Install: claude install-skill shouvik12/implementation-scout
# Implementation Scout Find the smallest set of verified existing implementations that covers the user's workflow — repo, file, function, and a version pin (tagged release when available) — plus license, and where the fix actually belongs versus adjacent layers that look plausible but aren't it. **This is Phase 1 (MVP).** It relies on `api.github.com` and `web_fetch`/`web_search` — no semantic search, no AST parsing, no vector DB. Earlier testing sessions found `api.github.com/search/code` returning 401 without a token, and raw file / commit-history URLs unreachable via `web_fetch` unless they'd already appeared in a `web_search` result — but **tool behavior can change between runs, and these must be re-verified fresh each session, not assumed from this note** (see "Tool capability is a runtime fact," below). Evidence precision — whether line ranges, commit SHAs, and provenance are achievable — depends on what actually works *this run*, tested at the time, not on a fixed rule about tokens. The skill should work well either way, citing the strongest evidence path that actually succeeds. ## Hard rule: no unverified claims This is the most important part of the skill. Every one of these must come from an actual tool call result in this run, not from training knowledge: - repository name / owner - star count - file path - function/class name - license If something can't be verified with a tool call, do not state it as fact. Write `not verified` and say what would need to ha