AgentSkillsCN

build-app-workflow

当您需要通过 Skills + Shell + Compaction 实现长期运行的代理，并在软件构建、修改、验证或交付过程中，追求可重复的执行效果时使用。涵盖阶段性门控工作流、脚本驱动的验证、回滚规范，以及跨技术栈（Node/Python/Rust/Tauri 等）的制品化交付。

SKILL.md
--- frontmatter
name: build-app-workflow
description: Use when building, modifying, validating, or shipping software with long-running agents and you need repeatable execution via Skills + Shell + Compaction. Covers stage-gated workflow, script-driven verification, rollback discipline, and artifact-based delivery across tech stacks (Node/Python/Rust/Tauri and others).
compatibility: hosted-shell, local-shell

build-app-workflow

Convert the Skills + Shell + Compaction methodology into a reusable, stage-gated software delivery workflow.

When to use

Use this skill when the user is:

  • •Implementing features or fixes that require real command execution, not advice only.
  • •Running multi-step tasks likely to exceed short context windows.
  • •Standardizing engineering workflow with verification and rollback gates.
  • •Building reusable team SOPs across repositories or stacks.

Do not use this skill for:

  • •Pure conceptual discussion with no repository changes.
  • •One-off copywriting tasks without build/test requirements.
  • •Highly specialized flows that already have a dedicated skill with tighter contracts.

Inputs to collect first

Collect these inputs before implementation:

  1. •Goal: one-sentence expected outcome.
  2. •Scope: files/modules that are in or out of scope.
  3. •Stack/runtime: Node/Python/Rust/Tauri/other and package managers.
  4. •Acceptance gate: what must pass (lint/test/build/smoke/etc.).
  5. •Rollback unit: the exact reversible boundary (single commit or file group).

If any input is missing, ask focused questions before execution.

Workflow (stage-gated)

Stage 1 - Plan

  1. •Define change scope and impact surface.
  2. •Define rollback point before edits.
  3. •Run bash scripts/preflight.sh.
  4. •Choose execution mode:
  • •hosted shell for isolation/repeatability.
  • •local shell for internal tools and fast iteration.

Reference: references/workflow.md sections Plan and Safety.

Stage 2 - Do

  1. •Apply minimal, reversible changes.
  2. •Keep API/type/caller/docs synchronized when interfaces change.
  3. •Produce concrete artifacts instead of narrative-only output.

Reference: references/workflow.md sections Do and Artifact Contract.

Stage 3 - Check

  1. •Run bash scripts/verify.sh --stack auto (or force node|python|rust when needed).
  2. •Validate against templates/acceptance-checklist.md.
  3. •If failed, run bash scripts/postmortem.sh "<task-id>" and fix by failure category.

Reference: references/workflow.md sections Check and Failure Routing.

Stage 4 - Act

  1. •Summarize change, risk, and rollback path.
  2. •Record run outcome in run-log.md.
  3. •Apply compaction policy for long runs.

Reference: references/workflow.md section Act and Compaction.

Compaction policy

Use compaction when any condition is met:

  • •8-12 consecutive implementation rounds.
  • •Before major refactor or broad file edits.
  • •Before release/handoff summary.

Do not compact repeatedly in short intervals without a milestone.

Stability gate and split policy

Treat this skill as "stable" only when all are true in the latest 5 real tasks:

  • •verify pass rate >= 4/5.
  • •zero process failures (missed required checks/docs/rollback notes).
  • •ad-hoc prompt additions <= 1 per task.
  • •acceptance records complete.

Split into sub-skills when two or more are true:

  • •workflow spans >= 3 distinct phases with frequent branching.
  • •SKILL.md becomes long/ambiguous for routing.
  • •repeated process errors >= 2 in 5 tasks.
  • •frequent skipped or misordered steps.

Governance details: ../docs/skill-lifecycle.md.

Output format required from assistant

Return outputs in this order:

  1. •Conclusion first.
  2. •Plan and impact surface.
  3. •Implemented changes.
  4. •Verification commands and outcomes.
  5. •Risks and rollback.
  6. •Optional next optimizations.
  7. •Sources used.

Use template: templates/agent-report.md.

Non-negotiables

  • •Do not place secrets/tokens in code, logs, docs, or examples.
  • •Do not invent unverified APIs or packages.
  • •Every change must be reversible.
  • •API changes require caller/type/doc updates together.
  • •Prefer script execution evidence over natural-language claims.

Quick runbook

  1. •bash scripts/preflight.sh
  2. •Implement minimal scoped changes.
  3. •bash scripts/verify.sh --stack auto
  4. •bash scripts/postmortem.sh "<task-id>" if verification fails.
  5. •Fill templates/acceptance-checklist.md and append run-log.md.

Sources