The Operator's Playbook

Shared Protocols Shared

Parent level: atlas/agents/_protocols/ -- one copy for all agents

Identity Files -- root of agent folder

memory/ -- starts empty, grows over time

knowledge/ -- Brand Brain + domain expertise

cani/ -- Constant And Never-ending Improvement

tools/ -- what this agent can use

security/ -- permissions and audit

data/ and scripts/ -- outputs and automation

Bot Script -- ~/olympus/infrastructure/telegram-{agent-name}/

LaunchAgent -- ~/Library/LaunchAgents/

• Name the agent

  Greek mythology name that fits the role.
• Define the job in one sentence

  If you need two sentences, the scope is too wide. Split it.
• Assign department

  Ad Creative | Email & SMS | Content & SEO | Social Media | Analytics | Lead Gen
• Define what the agent does NOT do

  Equally important. Prevents scope creep and agent conflicts.
• Identify coordination points

  Which agents does it work with? What data flows between them?
• Choose the model

  Opus for Atlas only. Sonnet 4.6 for all specialist agents.
• Define outputs and schedule

  Daily reports? On-demand analysis? HTML reports? Cron schedule?
• Identify data sources

  Meta API? Shopify? Klaviyo? Manual data from Mink? List every source.
• Mink confirms: "This is the job. Go."

  Do not proceed without explicit approval.

Copy from Template

• SOUL.md

  Copy verbatim from template. 30 hard rules, values, communication standards.

  Source: atlas/agents/_templates/SOUL.md
• IDENTITY.md

  Copy verbatim. Olympus mission, departments, agent roster.

  Source: atlas/agents/_templates/IDENTITY.md
• USER.md

  Copy verbatim. Who Mink is, how to work with him.

  Source: atlas/agents/_templates/USER.md
• SKILL.md -- BUILD TOGETHER

  Mink describes the job. Atlas writes it. Mink approves. This is the most important file in the agent.

  Source: Built from scratch by Atlas + Mink. Not templated.

The 3-Layer Brand Voice System

Real Data Only -- Zero AI Generation

• brand-profile.md

  Revenue, customers, AOV, LTV, product lines, margins, account details.

  Source: memory/shopify_data.md + memory/meta_ads_data.md + memory/klaviyo_data.md
• benchmarks.md

  KPI targets, alert thresholds, breakeven points. Must show the math.

  Source: Calculated from real financials (55% margin = 1.82x breakeven ROAS)
• playbooks/

  Domain-specific decision trees. Actionable, not theoretical.

  Source: Real account data + practitioner frameworks Mink approves

Data Sources by Agent Type

Continuous And Never-ending Improvement

Every agent has a CANI system that allows it to grow smarter over time. The baseline is set at launch from real data. CANI sessions add to it.

• sources.md

  Curate 5-10 practitioners, blogs, thought leaders for this agent's domain.

  Atlas + Mink curate together. Not AI-generated.
• best_practices.md

  Baseline proven patterns from verified data. CANI sessions add new ones over time.

  Initial baseline from real analysis only.
• queue.md, study_log.md, changelog.md

  Start empty. Fill over time as the agent learns.

• permissions.yaml

  Read-only by default. List exactly what this agent can and cannot do.
• authorized_contacts.md

  Mink, Spencer, Regina + any agent-specific contacts.
• manifest.json

  Define tools: generate_html_report, any domain-specific tools.

Telegram Bot Code Requirements

• .env loading at top of script

  Manual file parsing. No python-dotenv. Parse BEFORE any third-party imports.
• max_tokens=8192 on all API calls

  4096 is too low for HTML reports. 8192 is the standard.
• Correct model assigned

  claude-opus-4-6 for Atlas ONLY. claude-sonnet-4-6 for all specialists.
• try/except with logger.error in every handler

  Text, photo, voice, document -- all wrapped. User-facing error message on failure.
• max_tokens fallback

  If Claude hits max_tokens during tool_use, retry without tools. Never silently fail.
• System prompt loads ALL agent folders

  identity/, knowledge/, references/, memory/, cani/, data/ -- all loaded.
• Tool use loop with safety limit (loop_count < 5)

  Process tool calls, send results back, loop until end_turn.
• generate_html_report tool

  Uses lb_report_template.py. LB branded design.
• Vercel push for reports

  push_report_to_vercel() returns public URL.
• Data summaries, not raw JSON

  Summarize to ~2KB. NEVER dump 200KB of raw JSON into the system prompt.
• Conversation history (last 10 messages)

  Maintain context within a session.

Telegram Bot Setup

• Create bot via @BotFather

  Name: {Agent Name} - Olympus. Username: {agent}_olympus_bot
• Save credentials to .env

  TELEGRAM_BOT_TOKEN, ANTHROPIC_API_KEY, MINK_CHAT_ID
• Test locally on laptop first

  Send a test message. Verify response. Check logs. Fix issues before deploying.

Data Pipeline (If Applicable)

• Build data pull script (Python, no LLM calls = $0)

  Saves to data/reports/YYYY-MM-DD.json. Pure API calls.
• Credentials in ~/.openclaw/agents/{name}/credentials.json

  Never in repo. Never in .env. Separate credential file.
• Run manually and verify output

  Check the JSON. Verify data is real. All fields populated.

Deploy Checklist

• Rsync brain files to Olympus

  rsync -avz atlas/agents/{agent}/ olympus:~/olympus/team/{agent}/
• Verify files arrived

  SSH in. head -5 on each critical file. Content matches.
• Rsync bot script

  To ~/olympus/infrastructure/telegram-{agent}/. Include .env separately.
• Create LaunchAgent plist

  RunAtLoad=true, KeepAlive=true. NEVER nohup.
• Load LaunchAgent

  launchctl load ~/Library/LaunchAgents/com.{agent}.telegram-bot.plist
• Confirm process running

  ps aux | grep {agent}. Must show the Python process.
• Check logs for clean startup

  No errors. Bot logs "Started polling" or equivalent.
• Set up cron (if data pipeline)

  crontab -e. Verify with crontab -l. Run manually once.

10-Point Verification Protocol

• Send test message in Telegram

  Simple greeting. Verify bot receives it (check logs).
• Verify Claude API responds

  Logs show API response. No errors. Correct model.
• Verify reply is coherent and on-brand

  Does it say "brother" not "customer"? Does it sound like LB?
• Ask a domain-specific question

  Answer must reflect verified data, not AI training data.
• Test the 3-step system knowledge

  Must say: Step 1 Wash & Condition, Step 2 Oil, Step 3 Butter. 4 products. Never skip conditioner.
• Ask about its own capabilities

  Must know its files, what it can do, what it can't. No "I don't have X" when X exists.
• Request an HTML report

  HTML generates. Vercel URL returned. Report renders correctly. LB branding present.
• Verify CANI system is wired in

  Agent knows its CANI files exist and what's in them.
• Test error handling

  Send something that might cause an error. Verify graceful handling, not a crash.
• Run for 24 hours without intervention

  No crashes, no memory leaks, no log errors. LaunchAgent keeps it alive.

• Tell Mink the agent is live

  "[Name] is live and tested. Here's what it can do."
• Show a sample interaction

  Screenshot or live demo. Real question, real answer, real data.
• Walk through capabilities

  What it does, what it doesn't, how to trigger reports, how to run CANI.
• Mink sends first real message

  Watch it work. If anything breaks, fix it right there.