slice-workflow

name	slice-workflow
description	Skill-first workflow for planning, implementing, reviewing, and handing off a meaningful repo slice. Use when starting a slice, rendering planner/implementer/ reviewer prompts with mise, deciding what evidence to keep, or evaluating prompt quality with holdout sample tasks.
allowed-tools	Read, Edit, Glob, Grep, Bash

Slice Workflow

Use this skill to keep a slice legible from first plan through handoff.

The public interface is mise. This skill explains the workflow and judgment standard behind those tasks.

The deterministic implementation lives in cli/, a tiny uv-backed Python project that uses only the standard library at runtime. The .mise/tasks/* wrappers call its slice-workflow command, so the capability moves with the skill when scaffold generates a new repo.

Command Flow

mise run plan -- <slug>
mise run slice-plan -- --task path/to/task.md
mise run slice-implement
mise run slice-review
mise run sync-check

slice-* tasks render prompts into the active plan's prompts/ directory. They do not launch Codex, Claude, or any other harness. Paste the rendered prompt into the agent session you are already using.

Use mise -q run slice-status -- --json when another agent or script needs the active slice state.

Responsibilities

Planning

• Make META.yaml truthful before significant code changes.
• Turn vague work into concrete TODOs.
• Declare the evidence and review required for handoff.
• Record likely durable decisions in DECISIONS.md.

Implementation

• Follow the active TODO order unless the work teaches you it is wrong.
• Update LEARNING_LOG.md when scope, assumptions, or approach change.
• Record validation commands and outcomes as they happen.
• Keep scratch evidence out of durable context unless it is summarized or intentionally declared.

Review

• Treat review as adversarial verification, not a confidence ritual.
• Use the rubrics named in META.yaml.
• Write findings and final disposition into REVIEW.md.
• Update META.yaml review_backend to match the review artifact.

References

• references/artifact-policy.md - what to keep, summarize, or ignore
• references/handoff-rubric.md - what makes a slice handoff trustworthy
• references/holdout-sample-tasks.md - sample tasks for prompt-quality checks

Rules

• Skills own workflow policy and judgment.
• mise tasks own deterministic file creation, prompt rendering, status output, and validation exit codes.
• V1 is provider-neutral prompt rendering only.
• CI should block only deterministic contract failures. Fuzzy prompt quality belongs in holdout sample review until it becomes mechanically checkable.