Dawn Code Master v6.0

50+ 顶级编码技能集成 + 全生命周期工作流 + 反合理化验证门禁。 写代码、修bug、审PR、做设计，按这个来。

0. Skill Router

Core Operating Behaviors

1. Surface Assumptions - Before implementing, explicitly state assumptions
2. Manage Confusion - Stop, name confusion, present tradeoff, wait for resolution
3. Push Back - Point out problems directly, quantify downsides, propose alternatives

When to Use What

1. Coding Laws

Pre-Commit Checklist: syntax clean, tests pass, type hints, no bare except, no hardcoded secrets, f-string (not format()), pathlib (not os.path), docstrings, functions <= 30 lines.

File Structure: <= 300 lines per file, <= 30 lines per function, one thing per function.

Error Handling: async ops must have error handling, network requests must timeout, user input must be validated, external deps assumed to fail.

Anti-Rationalization:

2. Spec-Driven Development

Law: NO CODE WITHOUT AN APPROVED SPEC.

9 Mandatory Sections: Title/Metadata, Context, Functional Requirements (RFC 2119), Non-Functional Requirements, Acceptance Criteria, Edge Cases, API Contract, Data Model, Out of Scope.

4-Phase Gated Workflow: SPECIFY -> PLAN -> TASKS -> IMPLEMENT (each reviewed by human)

Stop-and-Ask Rules: Scope creep, >30% ambiguity, breaking changes, security implications, unknown performance.

Anti-Rationalization: | Rationalization | Reality | | "Clear requirements, no spec needed" | Clear reqs take 5 min. Unclear ones need it. | | "Quick prototype first" | The prototype is the final product. | | "Specs are too slow" | Specs save 10x refactoring time. | | "I will document later" | Later never comes. |

3. Planning and Task Breakdown

Each task <= 30 min, independently testable, with clear AC. Slicing strategies: Vertical (preferred), Contract-First, Risk-First.

4. Incremental Implementation

Cycle: Implement -> Test -> Verify -> Commit -> Next slice

Anti-Rationalization: "Writing everything at once is faster" -> 500 lines untested = 10x debugging. "Just a few lines, no need to commit separately" -> atomic commits make rollback/review clean. "I will test later" -> never happens.

5. Test-Driven Development + E2E

TDD Cycle: RED (write failing test) -> GREEN (min code to pass) -> REFACTOR (clean up)

Prove-It Pattern: Write test reproducing bug (RED) -> Fix bug (GREEN) -> Write regression test -> Refactor

E2E Silent-Pass Detection: toBeDefined()/not.toBeNull() on Locator always passes. Use toBeVisible()/toHaveText().

P0 Anti-patterns: toBeDefined() on Locator, forgotten it.only, empty catch blocks, un-awaited async.

Anti-Rationalization: "Tests pass, it is good" -> tests dont catch architecture/security. "AI-generated code is fine" -> needs more scrutiny. "Write code first, add tests later" -> tests what code does, not what it should do.

6. Context Engineering

From: ratel-ai/ratel (351 star) + NeoLabHQ/context-engineering-kit (1.3k star)

Core Principle: Progressive Disclosure - Load only what each turn needs.

Context Hierarchy: 1. Rules Files (persistent) -> 2. Spec/Arch (per feature) -> 3. Source Files (per task) -> 4. Error Output (per iteration) -> 5. History (accumulates, compacts)

Ratel-Style Catalog: All skills in SkillCatalog, tools in ToolCatalog. Each turn loads only matching ones.

Anti-Rationalization: "All rules in system prompt" -> noise reduces accuracy. "Same project, no need to separate" -> different modules need different context. "Write rules once" -> projects evolve.

7. Doubt-Driven Development

From: addyosmani/agent-skills -- doubt-driven-development

A confident answer is not a correct one. For non-trivial decisions (branching logic, crossing boundaries, unverifiable assertions, irreversible blast radius).

Doubt Cycle: 1. CLAIM (write claim + why matters) -> 2. EXTRACT (isolate artifact) -> 3. DOUBT (fresh-context adversarial review) -> 4. RECONCILE (classify findings) -> 5. STOP (trivial findings, 3 cycles, or user override)

Anti-Rationalization: "This is obviously correct" -> hides most dangerous assumptions. "Tests pass" -> only cover known scenarios. "No time" -> doubt cycle is faster than debugging.

8. Bug Fixing (Focused-Fix)

Law: NO FIXES WITHOUT SCOPE -> TRACE -> DIAGNOSE FIRST.

Phase 1: SCOPE - Identify feature, find files, understand purpose. Phase 2: TRACE - Map inbound dependencies (imports) and outbound (who imports this). Phase 3: DIAGNOSE - Risk tags: HIGH (public API/DB schema/security), MED (internal module), LOW (isolated file). Phase 4: FIX - Dependencies -> Types -> Logic -> Tests -> Integration. Fix one at a time, run tests after each. Phase 5: VERIFY - Feature tests -> referencing tests -> full suite.

9. PR Review (5-Axis + Anti-Rationalization)

From: addyosmani/agent-skills -- code-review-and-quality

Axis 1: Correctness - Matches spec? Edge cases? Error paths? Tests pass? Axis 2: Readability & Simplicity - Descriptive names? Straightforward flow? Logical organization? No clever tricks? Lines can be reduced? Abstractions earning complexity? Axis 3: Architecture - Fits system design? Clean boundaries? No circular deps? Appropriate abstraction? Axis 4: Security - Input validated? Secrets safe? Auth checked? SQL parameterized? XSS prevented? Axis 5: Performance - N+1 queries? Large object allocations? Uncached hot paths?

Output Format: Blast radius, per-axis assessment, severity labels (CRITICAL/REQUIRED/NIT), verdict (Approve/Request changes).

Anti-Rationalization: 9-item table covering "It works, good enough", "I wrote it, I know it is correct", "Clean up later", "AI-generated code is fine", "Tests pass", "Refactor makes it cleaner", "Just a small addition", "Just a version bump".

10. Code Simplification

From: addyosmani/agent-skills -- code-simplification

Principles: Chestertons Fence, Rule of 500 (>500 lines = decompose), reduce complexity not relocate it.

Checklist: Dead code? Conditionals can be simplified? Duplicate code? Complex expressions named? Nesting flattened? Single responsibility? Abstraction worth it? More code deleted than added?

11. Security Hardening

From: addyosmani/agent-skills -- security-and-hardening

OWASP Top 10 Prevention: SQL injection, XSS, auth, authorization, sensitive data, config security, dependency scanning, logging.

3-Tier Boundary: External (untrusted, validate all) -> Service (semi-trusted, auth) -> Internal (trusted, business logic).

12. Performance Optimization

From: addyosmani/agent-skills -- performance-optimization

Measure First, Optimize Second: Baseline -> Hypothesize -> Validate -> Optimize -> Verify -> Repeat.

Checklist: N+1 queries? Repeated computation? Large objects? Sync blocking? Caching? Lazy loading? Bundle size? Indexes? Connection pool?

13. Ship Gate and Deployment

From: addyosmani/agent-skills -- shipping-and-launch

Pre-Launch Checklist: All tests pass, code review approved, no P0/P1 vulns, CHANGELOG updated, version bumped, DB migrations forward-compatible, rollback plan, monitoring configured, docs updated, performance baseline, feature flags, staged rollout plan.

14. Git Workflow and Versioning

From: addyosmani/agent-skills -- git-workflow-and-versioning

Principles: Trunk-based development, atomic commits, ~100 lines per PR, commit as save point.

Convention: <type>: <description> (feat/fix/refactor/chore/docs/test/ci)

Anti-Rationalization: "Commit first, organize later" -> commit history IS code history. "One PR for everything" -> no one reviews large PRs. "Long feature branch is fine" -> merge hell.

15. CI/CD Pipeline and Automation

From: addyosmani/agent-skills -- ci-cd-and-automation

Stages: Lint -> Type check -> Unit test -> Integration test -> Build -> Security scan -> Deploy

Gates: Lint 0 errors (block), Unit test 100% (block), Coverage >=80% (block), Security scan 0 critical (block), Build success (block).

16. API Design Review

REST: kebab-case plural nouns (/api/v1/users). HTTP methods: GET (retrieve), POST (create), PUT (replace), PATCH (partial), DELETE (remove).

5-Dimension Score: Consistency (30%), Documentation (20%), Security (20%), Usability (15%), Performance (15%).

17. Database Design

Principles: Plural table names, UUID PK, created_at/updated_at, soft delete optional, indexes on query conditions, explicit FKs, forward-compatible migrations.

Performance Checklist: Indexes used? N+1 queries? Reasonable joins? Cursor pagination? Correct transaction scope?

18. Observability Design

Three Pillars: Structured logging, business metrics (latency/error rate/throughput), distributed tracing.

Health Endpoints: GET /health (liveness), GET /health/ready (readiness), GET /health/debug (internal only).

19. Code Migration

Strategies: Direct replacement (high risk, small module), Parallel run (medium, need rollback), Gradual (low, large system), Abstraction layer (low, dual maintenance).

Flow: Inventory -> Impact analysis -> Compatibility layer -> Execute -> Verify -> Cleanup

20. Tech Debt Tracking

Classification: Architecture (HIGH), Code quality (MED), Testing (HIGH), Documentation (LOW), Infrastructure (MED).

Entry Format: Type, Location, Description, Impact, Estimate, Created.

21. De-AI-Writing Check

Tier 1 (Always Flag): leverage/use, utilize/utilize, implement/build, nevertheless/but, furthermore/and, commence/start, endeavor/try.

Tier 2 (Cluster Flag): robust/reliable, seamless/smooth, facilitate/help, granular/detailed, holistic/complete, ecosystem/system.

Tier 3 (High Density): cutting-edge/modern, state-of-the-art/best available, game-changer/big change.

Detection Flow: First pass (flag all Tier 1) -> Structure check -> Rhythm check -> Second pass -> Output

22. Self Evaluation

Two-Axis: Task Difficulty (Low/Medium/High) x Execution Quality (Poor/Adequate/Strong) -> Score 1-5.

Forced Counter-Argument: Before final score, write reasons for lower, reasons for higher, then resolve.

Inflation Detection: 4+ of last 5 same score -> flag inflation. 3 consecutive 4s -> stricter evaluation.

23. Language Quick Reference

Python: snake_case, type hints, pathlib, f-strings, try/except/raise. TypeScript: camelCase, PascalCase for types, interfaces, async/await with try/catch. Go: camelCase, capitalized exports, error wrapping with fmt.Errorf.%w. PowerShell: Verb-Noun naming, CmdletBinding, ErrorAction Stop.

Appendix: Quick Reference

When to Use What

Never Do

1. Fix bugs without scoping first
2. Write code without spec
3. Merge without review
4. Commit without tests
5. Release without CHANGELOG
6. Change API without impact assessment
7. Dishonest self-eval
8. Ship without de-AI-fying