# Agent Global Rule Set — Situational Path Reference System
> **Created**: 2025-09-20 23:06 KST
> **Last updated**: 2026-08-31 11:58 KST
> **Modified by**: Gemini 3.7 Flash (High), Antigravity
> **Version**: 3.28.0
This file is the single, authoritative entry point for all AI agents (Claude Code, Gemini/Antigravity, Codex, etc.). It replaces the former multi-part distilled rule injection system and establishes a lightweight, on-demand situational routing system.
---
## 1. CRITICAL MANDATES & SINGLE SESSION READ PRINCIPLE
1. **Mandatory Session Start Read**:
- At the start of every session, before performing any tasks or replying to the user, you MUST read the following entry point files unconditionally:
1. `~/agentic/rules/00-instinct.md` (Absolute top-level mandate: resolve uncertainty against primary source or stop and surface; mechanically enforced via 3-tier lifecycle hooks)
2. `~/agentic/rules/01-harness-records.md` (Harness project record management mandate: NEVER omit `.harness/` records when `.harness/` exists in workspace root)
3. `~/agentic/rules/02-language-english.md` (Global language policy mandate: US English primary, authoritative English first + Korean parallel rendering on all final replies and questions, preserve Latin technical terms, hard script whitelist)
2. **Situational Trigger Rule Read**:
- Before performing any task that matches a situational trigger listed in Section 2, you MUST read the referenced rule file (`~/agentic/rules/##-name.md`) with your file reading tool and strictly adhere to every instruction inside it without exception.
3. **Single Session Read Principle**:
- **Do NOT re-read a rule file if you have already read it in the current session.**
- Once a target rule file (`~/agentic/rules/##-name.md`) has been loaded into your context during the active session, refrain from executing redundant re-reads when encountering subsequent triggers for that file. Rely strictly on the recalled rule directives in your context memory.
---
## 2. SITUATIONAL TRIGGERS & REFERENCED RULE PATHS
Consult and read the target rule file immediately when your task matches any of the following situational triggers:
### 00. Mandatory Session Start & Uncertainty Resolution
- **Trigger**: Mandatory at session start, or whenever encountering any element whose identity, value, code path, schema, or meaning is unconfirmed or uncertain.
- **Target Path**: [~/agentic/rules/00-instinct.md](~/agentic/rules/00-instinct.md)
- **Directive**: Absolute mandate: MUST be loaded at the start of every session. Never guess or speculate. Resolve uncertainty either by confirming against a ground-truth primary source or by stopping work immediately to report to the user. Mechanically enforced via 3-tier lifecycle hooks (PreToolUse, PreInvocation, and Stop/PostInvocation).
### 01. Mandatory Session Start & Harness Project Record Management (`.harness/`)
- **Trigger**: Mandatory at session start, or whenever the active project contains a `.harness/` directory, or when authoring/updating implementation plans, session histories, handoffs, or walkthroughs.
- **Target Path**: [~/agentic/rules/01-harness-records.md](~/agentic/rules/01-harness-records.md)
- **Directive**: Absolute mandate: MUST be loaded at session start alongside 00-instinct.md. Whenever `.harness/` exists in the workspace root, recording history, plans, walkthroughs, and handoffs is mandatory and must NEVER be omitted. When `.harness/` is absent, the rule remains inert—do NOT force creation of `.harness/` or write record files. Follow exact directory structures, naming conventions, mandatory recording rules, and git tracking constraints.
### 02. Mandatory Session Start & Global Language Policy (US English Primary & Bilingual Replies)
- **Trigger**: Mandatory at session start, and whenever formulating final replies, authoring documents/plans/records, asking interactive questions, writing git commit messages/PRs, or rendering technical terms.
- **Target Path**: [~/agentic/rules/02-language-english.md](~/agentic/rules/02-language-english.md)
- **Directive**: Absolute mandate: MUST be loaded at session start alongside 00-instinct.md and 01-harness-records.md. Enforce US English authoritative text followed by full Korean parallel rendering on all turn-ending replies and interactive questions. Strictly prohibit Korean-only regression upon Korean user input. Keep Latin-script technical terms and architecture invariant concepts in original form or parallel notation. Prohibit using 'pillars' in categorization and taxonomy contexts. Default all written artifacts to US English. Enforce hard script whitelist (Latin + Hangul only).
### 03. Korean Verb Precision & Prohibited Lexicon Policy
- **Trigger**: When communicating in Korean, writing user-facing documentation, committing messages, or formulating status logs.
- **Target Path**: [~/agentic/rules/03-korean-verb-usage.md](~/agentic/rules/03-korean-verb-usage.md)
- **Directive**: Strictly ban `박다` across all inflections, non-technical mimetic expressions, body metaphors, and colloquial slang. Enforce approved technical verb mapping across all technical operations. Deconstruct light verbs and experience possession formulas (`~경험을 보유하다/가지고 있다`).
### 03. Korean Natural Prose & Syntax Policy
- **Trigger**: When authoring Korean prose, user-facing documentation, explanations, conversational answers, proposals, or Korean commit messages.
- **Target Path**: [~/agentic/rules/03-korean-natural-prose.md](~/agentic/rules/03-korean-natural-prose.md)
- **Directive**: Enforce context-driven subject omission (zero-anaphora) and pragmatic object handling; deconstruct translation-ese (light verbs, double passives, by-passives, abstract subjects, structural metaphor suffixes `~ 축`, abstract `~화` compounding like `자산화`, redundant `투입 첫날부터` day-one filler, experience possession `~경험을 보유하다`, and empty savior formulas like `자립적으로 책임지다/통제하다`); enforce anti-parroting, solution-first, and anti-breakdown presumption rules; strictly ban AI signature cliches under the removal-only rule; and maintain ending variety, rhythm harmony, and speech-level consistency.