AgentSkillsCN

coding-prompt-bank

为 AI 编码智能体设定基础规则与行为准则。适用于启动编码项目时，“使用我的编码规则”“编码智能体的初始化”“项目初始化”“新建代码库”，或在向任意编码智能体（如 Claude Code、Cursor 等）提供指令时使用。系统还提供了分层的复杂度规则与项目模板。

SKILL.md
--- frontmatter
name: coding-prompt-bank
description: Base rules and behaviors for AI coding agents. Trigger on starting a coding project, "use my coding rules", "coding agent setup", "initialize project", "new codebase", or when providing instructions to any coding agent (Claude Code, Cursor, etc). Provides tier-based complexity rules and project templates.

Coding Prompt Bank

Project-agnostic base rules for AI coding agents. Copy relevant sections based on tier.

Core Philosophy

code
PARSE, DON'T READ
Treat requests as objects to decompose.
Extract: goal, constraints, inputs, outputs, dependencies.
Never process as prose blob.

CORRECTNESS > SPEED
Working first attempt beats fast iteration + debug.
Mental simulation catches 80% of bugs before execution.

EVERY LINE IS LIABILITY
Config > code. Native > custom. Existing > new.
Delete what doesn't earn its place.

CONTEXT IS SCARCE
Lean context prevents rot.
Reference, don't restate.
Compress aggressively.

Execution Laws

Investigate First

code
NEVER implement first.
1. Find working example (search, grep, docs)
2. Read every line
3. Copy pattern exactly
4. Adapt minimally

Single Source of Truth

code
One location for each concern:
- Auth: one extraction point
- Validation: one schema
- Config: one file
- Types: one definition

Fail Fast

code
Exit early. Clear messages. No silent failures.
If uncertain: STOP → ASK → WAIT.

Atomicity

code
No partial states. Transaction succeeds or rolls back.
Test in isolation before integration.

Tier Selection

code
T1 (Hackathon): Learning, throwaway, demo, <4 hours
T2 (Real): Side project, MVP, anything that persists (DEFAULT)
T3 (Critical): Production, multi-team, regulated, zero-downtime

T1: Minimal Process

  • •Ship > perfect
  • •Comments optional
  • •Tests optional
  • •"Does it work?" is the only gate

T2: Default Standard (use for most work)

  • •Pre-flight: What am I building? Why this approach? How will I know it works?
  • •One task = one thing
  • •Verify integration immediately
  • •Log decisions: DECISION: X because Y
  • •Basic error handling

T3: Full Rigor

  • •Phase gates required
  • •All T2 rules +
  • •Test coverage
  • •Rollback plan
  • •Documentation
  • •Code review checklist

Validation Protocol

Before implementation:

code
□ Working example found?
□ Dependencies identified?
□ Integration points clear?
□ Done-when defined?
□ Edge cases listed?

Project Template

code
PROJECT: [name]
TIER: [1-3] (default: 2)
STACK: [languages, frameworks, infra]

CONSTRAINTS:
- [hard limit 1]
- [hard limit 2]

AVOID:
- [anti-pattern 1]
- [anti-pattern 2]

OVERRIDES:
- [any base rule modifications]

TASK: [current request]

Quick Reference

code
BEFORE WRITING:
□ What am I building?
□ Why this approach?
□ What depends on this?
□ How will I know it works?

DURING:
□ One task, one thing
□ Clear done-when
□ No speculation
□ Verify integration immediately

AFTER:
□ Matches spec exactly?
□ Idempotent?
□ Fail-fast?
□ Decision logged?

Anti-Patterns

  • •Speculative code ("in case we need it")
  • •Partial implementations
  • •Silent failures
  • •Reimplementing existing solutions
  • •Skipping investigation phase
  • •Multiple sources of truth