job
Run multi-step work as a background minimise job — authors a plan YAML, wires the review gates, runs it, and reports what landed.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
Run multi-step work as a background minimise job — authors a plan YAML, wires the review gates, runs it, and reports what landed.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Author a mini job or loop plan under control — triage loop-vs-job, run a retrieval-aware interview pre-filled from existing personas and recent runs, then preview the exact YAML behind an approve/edit/reject gate before handing off to the unchanged `mini job new` / `mini loop new`. Use when starting a new job or loop from a one-liner and you want to review the plan before anything runs.
Run open-ended iteration on one artifact as a minimise refinement loop — plan → implement → evaluate, repeating until the goal is met or max_iterations is hit.
Reviews what a minimise job actually IMPLEMENTED against its plan — a NON-BLOCKING advisory reviewer that reads the plan (YAML) from stdin, inspects the job's git diff, and reports findings without ever aborting the job. Use when invoked as `/minimise:review-implementation`, or when a minimise `post_plan` hook runs `claude -p '/minimise:review-implementation'` to critique the finished work. Always exits reporting; it never fails the job. Checks correctness, completeness vs. the plan's goals, obvious bugs, missing tests, and over-engineering — mirroring ralphex's post-implementation review, trimmed to the essentials.
Reviews a minimise job plan or loop spec as a pragmatic BLOCKING quality gate, reading it (YAML) from stdin and printing a machine-readable REVIEW: PASS / REVIEW: FAIL verdict. Use when invoked as `/minimise:review-plan`, or when a minimise `pre_plan` hook runs `claude -p '/minimise:review-plan'` to gate a job before implementation. Reports ONLY severe issues (correctness bugs, data-loss/destructive risk, missing steps that make a task unimplementable, internal contradictions, factually wrong claims about the codebase, a loop goal with no stopping condition) and ignores style, wording, and nice-to-haves.
Verify and install the prerequisites for minimise — the `mini` CLI, the `claude` CLI, and a git repo.
| name | job |
| description | Run multi-step work as a background minimise job — authors a plan YAML, wires the review gates, runs it, and reports what landed. |
A mini job runs each task in a fresh agent session that sees only the previous task's
diff and completion report, so context does not rot across a long build the way it does here.
The plan is a YAML file the user reviews, edits, and commits — the pipeline lives in the
repo instead of evaporating in a transcript. Failed tasks retry; a crash mid-run resumes
from the first incomplete task. And hooks gate the work: a pre_plan review can block a
bad plan before any code runs, and a post_task review can re-run a task with the findings
fed back in.
If mini --help fails, mini is not installed. Tell the user to run /minimise:setup and
stop. Do not install it from here.
The user asked for a job; they have not yet seen the shape of it. Never author a plan, never
run mini job new, before they agree to the breakdown. Put in front of them:
pre_plan /minimise:review-plan hook.Then wait. If the user says no, do the work inline without further mention of minimise.
reference/plan-schema.md. Read it — estimated_duration_min
is required on every task and hook, and a plan missing it fails at mini job new.
Write it to worklogs/scratch/ if that exists, otherwise a plans dir the project already
uses (docs/plans/), otherwise ask where it should live. Do not drop YAML in the repo root.pre_hooks entry running
/minimise:review-plan (exact shell string in the reference). It is a blocking gate: a
failing review aborts the job before any task runs.mini job new --plan <file> # → Job ID; validates the plan, creates it PENDING
mini job start <id> # runs in the foreground until done
mini job status <id> (add --format json to poll from a script) and
mini job logs <id> -f for live narration.mini job results diff <id> and mini job results logs <id>. Summarize
what landed and what the review hooks said — the user delegated so they would not have to
read the logs.If the job fails, mini job start <id> again resumes from the first incomplete task; already-
committed tasks are not re-run. Read the failing task's logs before re-running blindly.