Skip to content

CUSTOMIZATION.md

Latest commit

 

History

History
135 lines (101 loc) · 4.79 KB

CUSTOMIZATION.md

File metadata and controls

135 lines (101 loc) · 4.79 KB

Customization Guide

The power of AI Cofounder comes from deep customization to your company. A generic setup is 10% as useful as a properly configured one.

Step 1: IDENTITY.md — Name Your Agent

Give your AI Chief of Staff an identity:

- **Name:** Atlas
- **Creature:** AI Chief of Staff for Acme CEO
- **Vibe:** Direct, data-driven, no-BS.
- **Emoji:** [pick one that fits]
- **Role:** CEO Operating System — helps Jane work as CEO, not as engineer.

Tips:

  • Pick a name you'll enjoy seeing in chat every day
  • The vibe sets the tone — "honest and direct" vs "supportive and encouraging"
  • The emoji appears in the agent's identity display

Step 2: USER.md — Tell It About Yourself

This is how the agent calibrates its communication:

- **Name:** Jane Smith
- **Timezone:** America/New_York
- **Language:** English

Key sections to fill:

  • Communication style — how direct/brief do you want responses?
  • CEO transformation context — what shifts are you working through? (engineer->CEO, consultant->CEO, etc.)
  • Output preferences — platform (Telegram/WhatsApp), max message length, language

Step 3: MEMORY.md — Your Company Context (Most Important)

This is what makes every recommendation specific to YOUR situation. Fill in every section:

Company

Basic info: name, what you do, team, stage.

Current Wedge

This drives everything. Be specific:

  • ICP: Not "companies" but "Engineering leads at 200-2000 dev companies with multi-repo legacy codebases"
  • Pain: Not "they need our product" but "AI coding tools lose context in large codebases, onboarding takes weeks"
  • Promise: Your one-line value prop for this specific ICP
  • Competitors: Top 3-5 alternatives they actually consider (including "do nothing")

PMF Stage

Be honest about where you are:

  • 0-1 (Search): Still looking for repeatable pain + value
  • 1-2 (Prove): Found pain, proving value is repeatable
  • 2-3 (Scale): Repeatable, now scaling

System Constraint

The single biggest bottleneck in your system. Common pre-PMF constraints:

  • "Rate of validated market learning" (not enough customer contact)
  • "Time to first value" (product too complex to onboard)
  • "Positioning clarity" (people don't understand what we do)
  • "Pipeline quality" (talking to wrong people)

The agent uses this to filter every recommendation: "Does this exploit the constraint or distract from it?"

Step 4: TOOLS.md — External Tools (Optional)

If you use PostHog, Exa, or other tools via mcporter, configure them here. Otherwise, leave TOOLS.md minimal — the CEO OS works without external tools.

Step 5: HEARTBEAT.md — Proactive Cadence

The default heartbeat sends:

  • Type B (Insight + constraint nudge): max 2/day, min 4h apart
  • Type C (CEO coaching): alternates with B
  • Type D (Daily recap): once in the evening window

Adjust timing to your schedule:

  • Change the morning quiet period (default: before 13:00)
  • Change the evening recap window (default: 21:00-22:00)
  • Increase/decrease the max messages per day

Step 6: AGENTS.md — Skill Routing

Usually no changes needed. But if you:

  • Removed some skills → remove them from the skill lists
  • Added custom skills → add them to the appropriate category
  • Want different routing → update the Routing section

Step 7: Cron Jobs (Optional)

Set up recurring tasks:

# Morning briefing every day at 9 AM
openclaw cron create \
  --cron "0 9 * * *" \
  --timezone "America/New_York" \
  --session isolated \
  --deliver --channel telegram --to YOUR_CHAT_ID \
  --prompt "Run the morning CEO briefing. Check metrics, flag anomalies, suggest priorities."

# Weekly review on Sundays at 6 PM
openclaw cron create \
  --cron "0 18 * * 0" \
  --timezone "America/New_York" \
  --session isolated \
  --deliver --channel telegram --to YOUR_CHAT_ID \
  --prompt "Run ceo-weekly-review"

After Changes

  1. Sync to server (if remote): copy files to ~/.openclaw/workspace/
  2. Fix permissions: chown -R openclaw:openclaw ~/.openclaw/workspace/
  3. Reload: send /new in chat

Context Budget

OpenClaw has a bootstrap context limit (~15,000 chars). Keep within:

  • SOUL.md < 3KB
  • AGENTS.md < 2KB
  • MEMORY.md < 3KB
  • IDENTITY.md < 500B
  • USER.md < 1KB
  • TOOLS.md < 1KB
  • HEARTBEAT.md < 1.5KB

Skills load on-demand and don't count toward this budget.

Tips

  • Update MEMORY.md weekly — after your weekly review, update metrics, hypotheses, and learnings
  • Don't over-fill MEMORY.md — keep it to confirmed facts and current state. Use daily logs for ephemeral data.
  • Be specific in your constraint — "growth is slow" is useless. "0 qualified conversations this week, target is 8-12" is actionable.
  • Let the agent push back — the anti-pattern checks and self-deception flags are features, not bugs