AgentSkillsCN

software-development

根据GitHub问题规范构建代码。每块代码都分块实现并附带测试。使用just命令进行质量门。创建自包含模块。

SKILL.md
--- frontmatter
name: software-development
description: Build code from GitHub issue specifications. Implements in chunks with tests after each chunk. Uses just commands for quality gates. Creates self-contained modules.

Software Development

You build modules from GitHub issue specifications.

Software Laws

Apply these principles during development:

  • Postel's Law - Liberal input acceptance, conservative output
  • Gall's Law - Start simple, iterate to complex
  • YAGNI - Add only when actually needed
  • RED-GREEN-REFACTOR - Test first, then implement

Input Source

GitHub issue with structured metadata:

  • Implementation guidance
  • Code examples
  • Chunk-by-chunk approach
  • Acceptance criteria
  • Testing strategy

Parse from issue body, not labels.

Working Context

code
Branch: epic/{epic-id}/{task-slug}
Base: main
Target: main (via PR)

All branches from main. Flat structure only.

Implementation Approach

Documentation First

Update docs with retcon writing (present tense):

  • Document as if feature exists
  • Update relevant files
  • Verify DRY (no duplication)

Chunked Implementation

For each chunk in issue:

  1. Implement - Follow guidance from issue, keep maximally simple
  2. Test - Unit tests for this chunk's specific functionality
  3. Verify - Run just check-all
  4. Commit - Incremental commits, incremental pushes

Incremental commits are recoverable.

Test Review

After all chunks:

  • Verify coverage >= 96%
  • Add edge cases and error handling
  • Integration tests if specified

Final Verification

Run just check-all. If failures:

  • Attempt auto-fix (max 3 attempts)
  • If still blocked: escalate

Module Structure

code
module_name/
├── __init__.py      # Public interface ONLY
├── README.md        # Contract (MANDATORY)
├── core.py          # Implementation
├── models.py        # Data structures
└── tests/           # Unit tests

Containment Rules

  • Everything inside module directory
  • No reaching into others' internals
  • Tests run without external setup
  • Clear public/private boundary

Public Interface

python
# __init__.py
from .core import process
from .models import Input, Output

__all__ = ['process', 'Input', 'Output']

No private exports. No internal leaks.

Quality Checklist

Before claiming done:

  • Matches issue spec exactly
  • Works in isolation
  • Public interface minimal
  • All tests pass
  • Coverage >= 96%
  • just check-all passing
  • Can regenerate from README

Anti-Patterns

Don't:

  • Export private functions
  • Import from other modules' internals
  • Mix multiple responsibilities
  • Skip tests after chunks
  • Push without verifying
  • Assume library availability

Self-Healing

When quality checks fail:

  1. Parse error messages
  2. Apply common fixes
  3. Re-run verification
  4. Retry up to 3 times
  5. If still blocked: escalate

Auto-fix what's fixable. Escalate what's not.

Success Criteria

Task complete when:

  • Implementation matches spec
  • Tests comprehensive
  • Quality gates passing
  • Incremental commits pushed
  • User acceptance tests verified

No draft PRs. No "almost done". Either passing or escalated.