← ClaudeAtlas

tz-verifylisted

Verify a TZ has been fully implemented. Cross-checks branch/PR/diff against acceptance criteria using 3 independent LLM critics (any mix of Claude/Codex/Gemini CLIs, OpenRouter, or NVIDIA NIM), per-AC focused evidence bundles with prompt-injection delimiters, citation+grep validation, and a 5-tier verdict (BLOCK / FIX-FIRST / SAFE-TO-COMMIT / SAFE-TO-DEPLOY-AFTER-CHECK / INSUFFICIENT-EVIDENCE). Read-only against source — only writes inside the verifications directory. Invoke after implementing a major TZ, before merge or deploy. Complement to /tz-review (which reviews specs *before* implementation). Works in whatever language the user writes in (Ukrainian, English, any other).
ymaxymovych/tz-skills · ★ 0 · Code & Development · score 75
Install: claude install-skill ymaxymovych/tz-skills
# TZ Implementation Verification Protocol ## Language Work in the language the user writes or dictates in. Detect it from the dictation, the draft, or the first message; if it is mixed or unclear, do not ask — use English. Everything user-facing follows that language: the TZ document, the questions and their recommended answers, the review journal, progress lines, the final report, and any commit messages you write on the user's behalf. Keep verbatim: command names (`/tz-go`), file names, code, config keys, and the code words — «ФІНІШ», "FINISH" and "DONE" are the SAME code word in any language. The Ukrainian text blocks inside this skill are templates of MEANING, not strings to paste: render them faithfully in the user's language. Prompts sent to critic models may stay in English (models handle it best), but every finding you quote back to the user is translated. Never switch language mid-run because a source file or a kit template happens to be in another language. **Skill version:** 4.1.0 (adds opt-in Queen→Workers fan-out — Step 3.5) **Report schema version:** 1.2 (adds audit log + ac_schema_version) Read-only post-implementation verification. Pairs with `/tz-review`. Design grounded in 2026 multi-model critic research: - **Spec-first + parallel independent verification** — Abhishek Ray's "self-congratulation machine" critique. - **Per-criterion focused evidence bundles** with **Microsoft Spotlighting + sandwich defense** for prompt-injection resistance (~17.8% baseli