One repository that integrates 6 independent rule systems: a shared core layer, one active profile, and on-demand capability packs. Clone once, pick a profile, and sync to any AI tool's rule file.
This repository is the single source of truth for AI collaboration rules, not application code for any specific project. It consolidates 6 previously separate rule repositories into one. Profiles are loaded in isolation so that conflicting domain constraints (for example, "never fabricate" versus "fiction is the core ability of novel writing") do not collide.
| Profile | Origin | Use Case |
|---|---|---|
coding |
badhope/AI | Software development, bug fixes, refactoring, code review |
conversation |
badhope/universal | General Q&A, research, comparison, information retrieval |
novel |
badhope/novel | Novel writing, chapter creation, character/worldbuilding |
interactive-novel |
badhope/interactive-novel | Interactive fiction, branching narratives, state machines |
paper |
badhope/paper | Academic paper writing, literature review, submission |
agent-builder |
badhope/AgentCreater | Design, evaluate, and deploy AI agents |
Why merge: keep 5 rule sets from drifting apart; clone one repository instead of five; unify the cross-tool sync entry point.
Why not merge into one set: domain constraints conflict (for example, "no fabrication" vs. "fiction is the core ability"). Profiles are loaded in isolation.
git clone https://gitcode.com/badhope/AI-RULE.git
cd AI-RULE# List available profiles
python scripts/sync_rules.py --list
# Generate Claude Code entry for the coding profile
python scripts/sync_rules.py --profile coding --tool claude-code
# Generate AGENTS.md (cross-tool standard, read by Codex CLI, OpenCode, Aider, etc.)
python scripts/sync_rules.py --profile coding --tool agents-md
# Generate all 13 tool entries for the novel profile
python scripts/sync_rules.py --profile novel --tool all| Category | Tool ID | Output File | Notes |
|---|---|---|---|
| Cross-tool standard | agents-md |
AGENTS.md |
Read by Codex CLI, OpenCode, Aider, Zed, Warp, Junie, Devin, Google Jules (20+) |
| Existing | claude-code |
CLAUDE.md |
Claude Code |
| Existing | gemini |
GEMINI.md |
Gemini CLI |
| Existing | cursor |
.cursor/rules/project.mdc |
Cursor (with frontmatter) |
| Existing | copilot |
.github/copilot-instructions.md |
GitHub Copilot |
| Existing | trae |
.trae/rules/project_rules.md |
Trae IDE |
| International | windsurf |
.windsurfrules |
Windsurf (12K char limit) |
| International | cline |
.clinerules/project.md |
Cline / Kilo Code |
| International | continue |
.continue/rules/project.md |
Continue.dev |
| International | amazon-q |
.amazonq/rules/project.md |
Amazon Q Developer |
| International | qodo |
best_practices.md |
Qodo (formerly Codium) |
| China | lingma |
.lingma/rules/project.md |
通义灵码 (10K char limit) |
| China | comate |
.comate/rules/project.mdr |
文心快码 (.mdr format) |
Copy the generated tool entry file (for example CLAUDE.md) to your project root, or reference this repository as a Git submodule and run the sync script.
Load the coding Profile from Rule Hub.
Or let the project anchors auto-detect (see below).
Load the <profile-id> Profile from Rule Hub.
| Anchor Signal | Inferred Profile |
|---|---|
pyproject.toml, package.json, requirements.txt + source code |
coding |
.ai-memory/creative-blueprint.md, chapters/, outline.md |
novel |
.game-state/, game-state-machine.md, save-slot-*.json |
interactive-novel |
config.yaml + tools.json + test-cases.md |
agent-builder |
| None of the above | conversation |
| Keywords | Profile |
|---|---|
| fix / refactor / test / API / bug | coding |
| write a chapter / continue / character / foreshadowing / worldbuilding | novel |
| start a game / branch / save / NPC / turn | interactive-novel |
| design Agent / agent config / tool permissions | agent-builder |
| query / compare / analyze / research | conversation |
- Origin: badhope/AI
- Scope: Python/FastAPI development, bug fixes, refactoring, testing, code review
- Core capabilities: Git SOP, dependency management, PowerShell syntax, MCP red lines, engineering hygiene
- Capability packs: research, testing, review, agent-governance, dar
- Mutually exclusive with: novel, interactive-novel
- Origin: badhope/universal
- Scope: General Q&A, research, comparison, information retrieval
- Core capabilities: truth protocol, deep search, anti-dumbing-down, clarification protocol, reasoning depth control
- Capability packs: research, dar
- Mutually exclusive with: novel, interactive-novel, agent-builder
- Origin: badhope/novel
- Scope: Novel writing, chapter creation, character/worldbuilding maintenance
- Core capabilities: creative seed confirmation, 35-item anti-AI-literary-flavor checklist, character consistency, foreshadowing tracking, story knowledge graph, three-tier revision
- Capability packs: research, worldbuilding, creative, dar
- Mutually exclusive with: coding, conversation, interactive-novel, agent-builder
- Origin: badhope/interactive-novel
- Scope: Interactive fiction games, branching narratives, state machine driven
- Core capabilities: game seeds, state machine, NPC autonomy, adaptive difficulty, save/load, turn-based
- Capability packs: creative, research, state-machine, npc-simulation, adaptive-difficulty, dar
- Mutually exclusive with: coding, conversation, novel, agent-builder
- Origin: badhope/paper
- Scope: Academic paper writing, literature review, submission, reviewer response
- Core capabilities: academic integrity protocol, citation verification, literature review methodology, paper structure (IMRaD/Review/Position/Case Study), research question extraction, methodology design, data presentation, anti-AI-academic-tone, peer review simulation, revision letter response
- Capability packs: research, dar
- Mutually exclusive with: novel, interactive-novel
- Origin: badhope/AgentCreater
- Scope: Design, evaluate, and deploy AI agents — produce config, tool definitions, test cases
- Core capabilities: four-layer role model, CTCO prompt structure, tool side-effect grading, memory systems, evaluation framework, 6 executable templates
- Capability packs: research, agent-governance, engineering, testing, dar
- Mutually exclusive with: conversation, novel, interactive-novel
Single-source rules (the core/ layer plus the AGENTS.md selector) are assembled per profile, and sync_rules.py then generates an entry file for each AI tool.
Clone the repository, pick a profile, run the sync, import the result into your project, and the AI works under unified rules that stay consistent across tools.
AI-RULE/
├── AGENTS.md # Rule hub entry (selector + priority + language mediation)
├── core/ # P0 hard constraints shared by all profiles
│ ├── governance.md # Security, permissions, MCP red lines, circuit breaker
│ ├── interaction.md # Clarification, intent normalization, output spec
│ ├── profile-router.md # Profile selection and capability pack whitelist
│ ├── language-mediation.md # Language mediation protocol (English reasoning, user-language output)
│ └── dar-spec.md # DAR (Domain Authority Registry) unified spec
├── profiles/ # 6 independent rule sets
│ ├── coding/ ( 13 files)
│ ├── conversation/ ( 19 files)
│ ├── novel/ ( 28 files)
│ ├── interactive-novel/ (31 files)
│ ├── paper/ ( 22 files)
│ └── agent-builder/ ( 70 files)
├── capabilities/ # 14 on-demand capability packs (incl. dar/ domain registry)
├── manifests/ # Per-profile assembly manifests
├── scripts/sync_rules.py # Generate tool entry files per profile
└── tests/ # 6 test suites (51 checks, all passing)
All system prompts are written in English for reasoning precision; rule documentation is bilingual Chinese-English for clarity. The AI communicates with you in your language:
- Input: auto-detect your language, identify the intent, then reason internally in English.
- Output: generate in English, translate to your language, and polish against translationese.
See core/language-mediation.md for details.
The sync script generates rule entries for 13 platforms:
| Category | Tool | Output File |
|---|---|---|
| Cross-tool standard | AGENTS.md | AGENTS.md (Codex CLI, OpenCode, Aider, etc. 20+ tools) |
| Existing | Claude Code | CLAUDE.md |
| Existing | Gemini | GEMINI.md |
| Existing | Cursor | .cursor/rules/project.mdc |
| Existing | GitHub Copilot | .github/copilot-instructions.md |
| Existing | Trae | .trae/rules/project_rules.md |
| International | Windsurf | .windsurfrules |
| International | Cline / Kilo Code | .clinerules/project.md |
| International | Continue.dev | .continue/rules/project.md |
| International | Amazon Q Developer | .amazonq/rules/project.md |
| International | Qodo (formerly Codium) | best_practices.md |
| China | Tongyi Lingma (通义灵码) | .lingma/rules/project.md |
| China | Comate (文心快码) | .comate/rules/project.mdr |
# Single tool
python scripts/sync_rules.py --profile coding --tool claude-code
# All 13 platforms
python scripts/sync_rules.py --profile coding --tool allThis repository incorporates findings from recent prompt engineering and AI alignment research:
- Instruction Budget: Empirical research (ManyIFEval, ICLR 2025) shows instruction adherence degrades as a power law as the simultaneous instruction count grows. P0 rules are capped at 5 simultaneously active; total hard constraints are capped at 12.
- Position Effects (Lost in the Middle): LLMs attend to the beginning and end of the context window and under-weight the middle. P0 rules are placed at both ends.
- Anti-Patterns: ALL CAPS emphasis, negative-only constraints, and manual "think step by step" are empirically ineffective on next-generation models (Claude 4.x, GPT-4.1). Rules are written with conditional logic and positive alternatives.
- Extended Thinking: Model-native reasoning budget (Claude 4.x / OpenAI o-series) replaces manual CoT for complex tasks.
- Three-Tier Behavior Boundaries: Allowed (autonomous) / Confirmation Required / Forbidden, replacing vague "appropriate behavior" declarations.
- GUID Delimiter Injection Defense: Random GUID-based delimiters replace fixed
[UNTRUSTED]markers to prevent marker-closing injection attacks. - Abstention Protocol: Explicit permission to say "I don't know" with anti-inflation guards, preventing confident fabrication.
- Self-Refinement: Reflexion loops and Constitutional self-critique for pre-output quality checking.
See profiles/agent-builder/docs/skills/ for full documentation.
pytest tests/ # 6 suites, 51 checks, all passing
# Or run individually: pytest tests/test_audit.py10 models tested across 6 scenarios (120 API calls), objectively comparing baseline (no DAR) vs enhanced (with DAR routing/scoring/domain-knowledge prompts). Full report:
tests/dar-evaluation/multi-model-report.md· Raw data:tests/dar-evaluation/full-test-results.json
| Dimension | Coverage |
|---|---|
| Models tested | 10 (1 primary API + 9 backup API) |
| Scenarios | 6 (coding / conversation / paper / novel / agent-builder) |
| Languages | English · 中文 · 日本語 |
| Total API calls | 120 (baseline + enhanced) |
| Valid results | 60 |
| Scoring | 6 dimensions × 0–5 = /30 per scenario |
| Model | API | Status | Baseline | Enhanced | Δ |
|---|---|---|---|---|---|
| Qwen3.5-397B-A17B | backup | ✅ Available | 18.3 | 20.5 | +2.2 |
| DeepSeek-V4-Pro | backup | ✅ Available | 15.8 | 15.2 | -0.7 |
| moonweaver-4.8 | primary | ✅ Available | 14.3 | 13.2 | -1.2 |
| DeepSeek-V4-Flash | backup | ⚠ Partial | 7.0 | 4.5 | -2.5 |
| glm-4.7 | backup | ⚠ Partial | 7.5 | 5.0 | -2.5 |
| step-3.7-flash | backup | ⚠ Low quality | 2.8 | 2.0 | -0.8 |
| glm-5.2 | backup | ❌ Timeout | — | — | — |
| Kimi-K2.6 | backup | ❌ Timeout | — | — | — |
| MiniMax-M3 | backup | ❌ Timeout | — | — | — |
| Spark-X2-Flash | backup | ❌ Auth fail | — | — | — |
| sensenova-u1-fast | backup | ❌ Not found | — | — | — |
xychart-beta
title "DAR Enhancement: Baseline vs Enhanced (avg score /30)"
x-axis ["Qwen3.5-397B", "DeepSeek-V4-Pro", "moonweaver-4.8"]
y-axis "Average Score" 0 --> 25
bar [18.3, 15.8, 14.3]
bar [20.5, 15.2, 13.2]
| Scenario | moonweaver-4.8 | DeepSeek-V4-Pro | Qwen3.5-397B-A17B |
|---|---|---|---|
| S1-CVE (coding) | +14 🟢 | 0 ⚪ | +1 🟢 |
| S2-GDP (中文) | -3 🔴 | -11 🔴 | +2 🟢 |
| S3-ACADEMIC | -19 🔴 | +3 🟢 | +5 🟢 |
| S4-NOVEL | +3 🟢 | +7 🟢 | +11 🟢 |
| S5-JP (日本語) | 0 ⚪ | +4 🟢 | -2 🔴 |
| S6-AGENT | -2 🔴 | -7 🔴 | -4 🔴 |
🟢 = DAR improvement · ⚪ = no change · 🔴 = DAR regression
xychart-beta
title "DAR Impact by Dimension (avg delta, 3 effective models)"
x-axis ["Routing Acc.", "Source Qual.", "Domain Know.", "Citation Fid.", "Conflict", "Freshness"]
y-axis "Score Delta" -0.5 --> 1.0
bar [0.72, 0.28, 0.22, -0.44, -0.33, -0.33]
DAR improves: Routing Accuracy (+0.72, core value), Source Quality (+0.28), Domain Knowledge (+0.22)
DAR does not improve: Citation Fidelity (-0.44), Conflict Handling (-0.33), Freshness Awareness (-0.33)
- DAR excels in domain-specific scenarios — S4-NOVEL (+11) and S1-CVE (+14), where models lack specialized source knowledge (Etymonline, NVD).
- DAR's routing rules are its greatest value — Routing Accuracy improved +0.72, far exceeding the other dimensions.
- Qwen3.5-397B-A17B is the most DAR-compatible model — 4 of 6 scenarios improved, average +2.2.
- Long DAR prompts can hurt small models — moonweaver-4.8 returned an empty response on S3-ACADEMIC (−19).
- DAR adds noise when the baseline is already strong — S6-AGENT regressed across all models.
- Compress the DAR prompt prefix from 200–400 words to under 100 words.
- Provide a lite DAR (routing only) for smaller models.
- Append "all factual claims must cite URL + date" to strengthen Citation Fidelity.
- Skip DAR enhancement when the baseline score already exceeds 20/30.
- Refine the Chinese DAR prompt wording to avoid disrupting model comprehension.
Capability packs are composable, on-demand work methods. They do not define agent identity — the profile does. Packs only provide methodology.
| Pack | Use Case |
|---|---|
research |
Fact support, data validation |
testing |
Writing/verifying tests |
review |
Code/content review |
engineering |
Engineering implementation |
creative |
Creative generation, style, revision |
worldbuilding |
Worldbuilding, characters, timelines |
state-machine |
State machine governance, branch reachability |
npc-simulation |
NPC autonomy, memory, relationships |
adaptive-difficulty |
Difficulty adaptation |
game-engine |
Game turns, saves, commands |
agent-governance |
Agent evaluation, observability, safety alignment |
orchestration |
Multi-agent orchestration |
novel-chapter-deliverable-mode |
Novel chapter delivery mode |
dar |
Domain Authority Registry — authoritative source lists, scoring, routing |
See capabilities/README.md.
Higher priority wins on conflict:
P0: core/ security, permissions, truthfulness, MCP red lines
> P1: user's current explicit confirmation
> P2: main profile domain rules
> P3: capability pack on-demand rules
> P4: model default behavior
Can guarantee:
- Profiles are mutually exclusive and conflict-free.
- Manifest references are complete.
- Generated files come from specified sources.
- Rule sets include the three layers: core + profile + skills.
- Hand-edited generated files can be overwritten by re-syncing.
Cannot guarantee:
- Any model 100% executes natural-language rules.
- Rule files alone prevent dangerous operations (this needs tool permissions, Git hooks, and human confirmation).
- Auto-configuration of Trae custom agents or MCP after cloning (manual setup is required).
This repository is mirrored on both GitCode and GitHub with identical content:
- GitCode (primary): https://gitcode.com/badhope/AI-RULE
- GitHub (mirror): https://github.com/weed33834/AI-RULE
MIT