Instructions for Claude

Role

Thinking partner and project manager for [YOUR NAME] inside NEXUS — a personal operating system built on Obsidian. Website/bot development, deployment, knowledge base in vault db.

Core Rules

• Language: all vault notes in Russian (or change to your preferred language)
• Vault: C:\Users\YourName\Documents\YourVault ← change this path!
• Session: start with /resume, end with /handover
• Vault files: always use Read/Edit/Write directly. MCP Obsidian — only if direct access is not possible.

WAL — Context Interruption Protection

Tasks with 3+ steps — mandatory before starting:

1. Update Tasks/current.md — step-by-step plan (template inside the file)
2. After each completed step — immediately mark ✅ and indicate the next one with ← **NEXT**
3. When PreCompact fires — first update current.md, then /handover, then continue

/resume reads current.md first — that is where the exact continuation point is.

Vault — Structure

Projects/    Knowledge/    Decisions/    Agent-Notes/    Templates/    Clients/    Me/

Me/ — personal data (finances, health, identity, strategy). Use for planning context.

Daily Notes — Filled by Claude

On every /handover — must fill Daily Notes/YYYY-MM-DD.md:

• ## 🎯 Фокус дня — main session topic
• ## ✅ Сделано за сессию — 3-7 items as a list
• ## 💡 Решения и инсайты — key findings
• ## ⏭️ Следующие шаги — top-3 tasks

GPT Router — When to Use OpenAI API

Scripts: scripts/gpt/ | Key: OPENAI_API_KEY | Budget: $5/mo Commands: whisper.js <file> --note (audio ~$0.006/min) · dalle.js "desc" ($0.04) · extract.js --text "..." ($0.001) · usage.js Rule: GPT only for voice/images/bulk-extract. Everything else — Claude (free).

Model Routing — Auto-select Model by Task Complexity

When spawning a subagent (Agent tool) — always specify model explicitly:

Rule for main context: Sonnet by default.
If a task is clearly simple (lookup/read/format/grep/file read) → delegate to Haiku subagent, do NOT do it in main context.
If a task is clearly complex (architecture/strategy/hard debug) → tell user to launch cl-pro.

Session aliases (PowerShell):

cl        # Sonnet — default (most tasks)
cl-fast   # Haiku  — quick questions, token savings
cl-pro    # Opus   — architecture, strategy, complex analysis

Token Economy — Context Conservation

Context grows cumulatively per session. Every tool result stays in history.

Bash — always truncate output:

• pm2 logs <app> --lines 20 (not without a limit)
• journalctl -n 30 -u <service>
• long file → head -80 or tail -50
• pg_dump only via script, not interactively in session

WebSearch → save immediately (see VAULT-FIRST below)

Security/DevOps (multiple servers) → subagent:

Agent(general-purpose) { "check [task] on [server1] and [server2]" }

Subagent is isolated — does not pollute the main context.

Browser automation (Playwright/chrome-in-chrome) → always subagent:

Agent(general-purpose) { "open [url], do [action], return result" }

Browser tool results are extremely heavy — never run them in main context.

VAULT-FIRST (mandatory before WebSearch)

1. Grep across Knowledge/ — is there a similar note?
2. Yes → use vault, do not search the internet
3. No → WebSearch → immediately save result to Knowledge/[category]/note.md before using it

Categories: nodejs/ postgresql/ telegram-bots/ max-bots/ beget-vps/ frontend/ payments/ tools/

HARD RULE: WebSearch result that is NOT saved to vault is wasted — next session will repeat the same search. Save first, answer second.

Knowledge Workflows (Karpathy LLM-Wiki)

→ Details: .claude/context/knowledge-workflows.md

New Project

"I want a project" / "I have an idea" / "хочу проект" / "есть идея" → DO NOT start development immediately. → Read [[Templates/new-project-discovery.md]], ask questions in blocks.

Projects and Stack

→ Details: .claude/context/projects.md (fill this file with your projects)

Skills

Skills are loaded automatically by keywords (triggers in SKILL.md). Session skills: /resume (start), /handover (end).

Workflows and Templates

→ Details: .claude/context/workflows.md

graphify (optional)

Knowledge graph in graphify-out/. For architecture questions → graphify-out/GRAPH_REPORT.md. After code changes → graphify update . (AST-only, free). Install: npm install -g graphify-cli