Skip to main content

Design Overview

## §1 Goal

Establish two canonical identity contracts — `~/.claude/user.md` (Alem-as-user) and `~/.claude/personality.md` (John-as-agent) — and wire them into `~/.claude/CLAUDE.md` as authoritative sources loaded at session boot.

**Why LOW gap:** The information exists today. `~/.claude/CLAUDE.md` carries CEO bio, 5 hard constraints, decision tree, and ZAKONs. The 78 persona definitions (~/system/agents/definitions/) carry agent traits. The gap is consolidation and canonicalization, not invention. No new data needs to be created without an AI-interview with Alem.

**What this achieves:**
- Single source of truth for who Alem is (communication style, constraints, trust model) — eliminates persona drift when 78 agents each infer it independently.
- Single source of truth for who John is (operating principles, cost discipline, decision tree) — eliminates contradiction between persona files that partially duplicate CLAUDE.md.
- Enables Pillar #7 (multi-client architecture) to reference personality.md as the root John identity without re-authoring it per client.
- Reduces session-start token waste: user.md + personality.md are tightly scoped (1 page each) vs the current CLAUDE.md which mixes identity, rules, routing, and tool references.

---

## §2 Current State

### What exists

**`~/.claude/CLAUDE.md`** — Primary loaded file at every session. Contains:
- CEO identity (name, email, phone, role)
- John identity paragraph ("second brain, think, delegate, verify")
- 5 hard constraints
- Decision tree (CLASSIFY → MEHANIK → ROUTE → DELEGATE → MONITOR → REPORT)
- ZAKON NULA (tool-first rule)
- ZAKON PI2 (deploy verification)
- ZAKON PLAN (validation + docs in every plan)
- Specialist routing table (10 companies, 20+ named agents)
- Context loading table
- Essential tools table
- Session start procedure

**`~/system/agents/definitions/`** — 60 `.md` + 71 `.yaml` persona files (131 files total; task brief references 78 as of prior SENTINEL audit). Each persona encodes: role, company, communication style, domain knowledge, sometimes behavioral rules. John's own identity traits are partially duplicated across builder.md, orchestrator logic in mehanik.md, and fragments in multiple skill files.

**`~/.claude/agents/`** — Mirrors definitions/ for Claude Code reads (dual-store per feedback_agent_definitions_dual_store.md). Contains builder.md, validator.md, and ~20 key personas.

### What is missing

1. **`~/.claude/user.md`** — No file describing Alem-as-user. His communication preferences, decision patterns, time constraints, and trust model are scattered in:
   - CLAUDE.md ("CEO: Alem Basic, founder, carries everything, trusts you")
   - MEMORY.md memos (feedback_john_broken_pattern_2026-05-02.md, feedback_iteracija_means_execute.md, etc.)
   - Session-state memos written after incidents
   Agents inferring user preferences from these scattered sources will diverge. No single file an agent can load to understand how Alem works.

2. **`~/.claude/personality.md`** — No single-file John identity contract. Personality traits are distributed across:
   - CLAUDE.md (identity paragraph + constraints + decision tree)
   - builder.md (what John delegates)
   - 78 persona files that partially describe interaction with John
   - Rule files that describe what John must never do
   No persona file says "this is John's complete operating identity."

3. **Per-persona identity inheritance** — Personas do not reference a canonical John contract. If John's decision tree changes, 131 files would need updating. No inheritance chain.

---

No comments to display

Back to top