← ClaudeAtlas

workflow-reviewlisted

Score the sessions in a time window and propose harness and workflow improvements.
praisethebit/skills · ★ 1 · AI & Automation · score 77
Install: claude install-skill praisethebit/skills
## Opus guard Do this check first, and again at the start of each turn until the skill ends. The skill ends when step 8 is done, or when step 2 finds 0 sessions. Your system prompt names your model. If the model is not Opus and not Fable, stop. Do not do the next step. Tell the user: "This skill needs Opus or Fable. Press `s` in `/model`, pick Opus (the default) or Fable, then send your last message again." Why: a `model:` key in a skill lasts for one turn only. Each reply of the user then goes back to the session model. ## Task Grade the workflow and the harness. Score the sessions in one window on 6 rubrics, compute 7 scores and `overall`, and propose changes. The goal is fewer tokens, simpler steps and better efficiency. Use evidence from the logs, not opinion. Everything runs on this machine: do not upload a transcript. Each step ends on its completion criterion. Local extension: read `~/.claude/local/workflow-review.md` when it exists, and follow its steps. Let `SKILL_ROOT` be the directory that holds this `SKILL.md`. Shell variables do not carry from one Bash call to the next: in each Bash call, write the absolute paths of `SKILL_ROOT` and `REPORT_DIR`, or set them again. ## Step 1: Scope Ask with `AskUserQuestion`: "Which sessions should I score?" Options: this repo (recommended), all sessions, chosen projects. For chosen projects, ask for the paths. Each path must be a git repository. Set the scope keys. "This repo" has one key: the absolute path of the main