directory-engine/best-practices.md

Grokbot / Agent-Team Best Practices

Extracted from Startup Ideas Podcast (Greg Isenberg + Billy Howell). Adapted for Leonard / Hermes.

Core Design Philosophy

• Constraints are a feature. Limited agents (color + shape combinations) force mission orientation instead of one-agent-per-task sprawl.
• One thread / one box per agent = DM-style interaction. Dramatically lower switching costs than ChatGPT threads or scattered project folders.
• Agents live on cloud VMs. Context is shared within the team but should not be polluted across unrelated businesses.
• Separate accounts / contexts for separate businesses. Running newsletter + Twitter + Gmail receipt sorting in one place creates context bloat and burns tokens.

Mission Alignment

Mission of every business = make money (or clear progress toward it). Chief of Staff (or Leonard acting as COO) is the single point of contact for the human. Other agents report upward.

The 4-Week Cadence (non-negotiable)

Week	Focus	Rule
1	Build the team	Chief of Staff + max 3 agents. Give CoS access to existing tools (Notion, Gmail, Slack, etc.) and have it audit + recommend the first three revenue-driving roles.
2	Execute, no tinkering	Run the damn ball. No new agents, no skill shopping, no tool debates. Human makes the hard decisions (e.g. “we are using Notion, period”).
3	Hire / fire	Expand only for clear gaps discovered in week 2 (inbox monitor, merch shop, etc.).
4	Automate	Ask CoS: “Based on how we ran this week, what routines should move the ball while I sleep?” Add daily/weekly briefs and SOPs.

Repeat the loop. Never start four businesses at once.

Chief of Staff Pattern

1. First agent you create.
2. Day-1 task: audit the business (or idea) and propose the top 3 agents that drive revenue.
3. Perfect a workflow with the CoS first. Only after the output is good do you spin out a dedicated agent for that job.
4. CoS routes tasks, receives 5-line briefs from specialists, and surfaces blockers + decisions to the human.
5. Human can still go direct to a specialist when needed, then tell it to report back to CoS.

Routines (the real leverage)

Every specialist agent should have a short routine (daily / weekly):

• Look at current todos / what you created today
• What blockers exist?
• Write a 5-line report and send to Chief of Staff

5 lines is the hard limit. More burns tokens and context.

CoS aggregates: what is done, what is stuck, what needs human unblock.

Quality & Feedback Loops

• Adversarial / QA panels: After a piece of work, spin a temporary panel (CoS + 1–2 others) for 3 rounds of critique. Moves output from ~50% → ~90% without human micromanagement.
• Turn human feedback into a skill the agent (or reviewer) can reuse: “Review exactly like Billy / Tony would.”
• Automate low-value widget tasks (two-sentence blurbs, formatting) with scripts, Make.com, or cheaper models. Do not waste the high-context agent on them.

Anti-Patterns (kill these)

• Agent creep / idea creep
• Letting agents decide the stack for three weeks
• Open-ended research tasks (“go research X for two weeks”)
• Putting generated artifacts on the human desktop then hunting for them later — keep work on the agent computer, surface the final artifact in chat / Notion after review
• Cold-emailing the ocean. Instead: Monday = add 5 prospects; later = pick top 3 + custom package + draft → CoS → human tweak

Three Proven Business Shapes

1. Niche / Local Newsletter (lowest barrier)

• Beehiiv + Notion (or equivalent)
• Research agent fills source cards
• Blurbs can be automated once the format is locked
• Sales / inbox agent monitors for sponsors, prices packages, drafts outreach (quality > quantity for local)

2. Shopify / Physical Product

• Separate Grokbot / agent context
• Sourcing agent + developer agent + email agent
• Use ImportYeti-style research + CLI tools
• Ask for ranked recommendations with costs and pricing, never open-ended exploration

3. Site Builder Stack → Directories & Landing Pages

• Stack: Cloudflare + Porkbun (domains) + Astro + GitHub
• Excellent for static directories and quick landing pages
• Research agent writes Markdown pages daily → engineer (Leonard) commits → deploy
• Once traffic exists: referral-link collector + outbound deals agent
• Directory becomes the foundation that feeds a newsletter, Shopify, SaaS, or app

Directory-Specific Notes (our current focus)

A directory is curated data. In an AI-answer world (ChatGPT, Google AI Overviews) curated sources become more valuable, not less.

Play:

1. Pick niche with taste
2. Astro site
3. Daily research → polished Markdown page
4. Engineer publishes
5. Traffic first
6. Only then affiliate / outbound / newsletter cross-sell

Never solicit deals before traffic.

Operating Rule for Leonard

Treat every run like a real company with a real team. Small number of high-value agents. Mission-critical only. Human decides the hard things. Automate only after the process is proven.