CLAUDE.md + Rules + Hooks

Benchmarked against: Anthropic — Console Prompting Tools Architecture: Three-layer agent configuration Location: CLAUDE.md (root), .claude/rules/ (rules), .claude/settings.json (hooks)

This page details each component of SuperPortia's prompt engineering stack — how they work individually and together.

---

CLAUDE.md

Location and loading

Property	Value
File	CLAUDE.md (project root)
Loading	Always (injected at session start)
Token cost	~500 tokens
Protection	Locked by PreToolUse hook (Captain authorization required to edit)

Structure

# SuperPortia — Agent Boot Config

## Identity
Who you are, who the Captain is.

## Watch Rule
Timestamp every reply with Taipei time.

## Boot Sequence
1. Read hook output
2. Ask Captain what to work on

## Pre-Flight Check
3D scoring reference (pointer to UB entry).

## Engineering Governance
Pointer to EGS spec.

## Information Retrieval
Mandatory search flow: UB first, delegate, ingest.

## Version Control
Commit conventions.

## Terminology
Pointer to glossary.

## Communication
Behavioral guidance.

Design principles

Principle	How applied
Lean	Pointers to specs, not full content
Stable	Rarely changes — locked by hook
Identity-focused	Answers "who am I, what do I do?"
Boot-focused	Answers "what do I do first?"

---

Rules

Location and loading

Property	Value
Directory	.claude/rules/
Loading	Always (all .md files auto-loaded)
Total token cost	~2,000 tokens
Count	8 rule files

Rule files

File	Purpose	Token cost
company-constitution.md	14 sections of core behavior	~400
agent-intelligence-protocol.md	8 mandatory protocols	~350
cost-awareness.md	Engine and token discipline	~200
tech-freshness.md	Perishable knowledge defense	~200
ub-governance.md	Knowledge management standards	~250
gsta-alignment.md	Strategic alignment format	~150
compaction-recovery-protocol.md	Memory cliff defense	~150
glossary.md	Term definitions and pointers	~300

Rule design guidelines

Guideline	Why
Keep rules behavioral, not procedural	Procedures go in skills (L3)
Each rule under 500 tokens	All rules loaded every session
Use pointers to UB for details	Rules stay lean
No hardcoded dates or versions	Use "current" or "latest"
No duplication across rules	Single source per concept

Rule vs skill decision

Question	Answer
Must every agent always know this?	Rule
Is it a specific step-by-step procedure?	Skill
Does it shape all behavior?	Rule
Is it invoked by command?	Skill
Would it waste tokens if unused?	Skill

---

Hooks

Location and configuration

Hooks are configured in .claude/settings.json:

{
  "hooks": {
    "SessionStart": [...],
    "PreToolUse": [...],
    "PostToolUse": [...]
  }
}

Hook types

Hook	Trigger	Common uses
SessionStart	New session begins	Inject timestamp, strategic anchor, agent intelligence
PreToolUse	Before any tool call	Gate operations, protect files, validate inputs
PostToolUse	After any tool call	Log results, trigger follow-up actions

SessionStart hook

The most important hook — sets the agent's initial context:

#!/bin/bash
# .claude/hooks/session-start.sh

# Inject current Taipei time
echo "$(TZ='Asia/Taipei' date '+%Y-%m-%d %H:%M') (Taipei)"

# Strategic anchor (survives compaction)
echo "STRATEGIC ANCHOR"
echo "  Project: SS1 | Role: Chief Engineer"

# Pre-Flight Check reminder
echo "Pre-Flight Check active — 30s self-check before non-trivial tasks"

PreToolUse hook

Gates specific operations:

Use case	Implementation
Protect CLAUDE.md from edits	Check if target file is CLAUDE.md, block if not authorized
Validate dangerous commands	Check for destructive shell commands
Log tool usage	Record which tools are called

PostToolUse hook

Reacts to completed operations:

Use case	Implementation
Log ingestion results	Record entry IDs after ingest_fragment
Track WO transitions	Log state changes
Notify on errors	Send alerts on failures

---

How they work together

Loading order

1. System prompt (Claude's base behavior)
2. CLAUDE.md (L1 — identity and boot)
3. Rules (L2 — behavioral standards)
4. SessionStart hook output (event-driven context)
5. Skills (L3 — on demand, when user invokes /command)

---

Authoring guidelines

For CLAUDE.md

• Keep under 500 tokens
• Only pointers, never full procedures
• Lock with PreToolUse hook to prevent accidental edits
• Update rarely — version changes go in changelog

For Rules

• One concept per file
• Under 500 tokens per file
• Use tables for structured information
• Include "Go Deeper" pointers to UB entries
• No procedure steps (those go in skills)

For Hooks

• Keep scripts fast (under 1 second)
• Output text is injected into agent context (costs tokens)
• Test hook output for token efficiency
• Use environment variables for ship-specific values

---

Related pages

Page	Relationship
Prompt Engineering Overview	High-level concepts
Skills Overview	L3 skill reference
Skills in CLAUDE.md	How skills integrate
Context Windows	Token budget