Research
Competitors, SERP intent, keyword map, monetization opportunities.
The current bottleneck
“CEO, go make a website” is not a workflow.
A single broad CEO task can skip steps, duplicate work, or forget critical pieces like footer links, cookie/legal pages, security checks, analytics, or QA. The CEO should manage exceptions, not invent the production process every time.
- Agents may complete isolated tasks without knowing dependencies.
- Manual issue creation becomes the real production system.
- Compliance, auth, payments, SEO, and deployment gates can be missed.
- Failures like rate limits or account credit exhaustion stall the whole company.
Recommended architecture
Separate orchestration from execution.
The workflow driver is deterministic. Paperclip is the control plane. Agents execute scoped issues. Humans approve risky decisions.
01
Workflow Driver
n8n first, Temporal/custom runner later. Stores the state machine, retries, stage order, and gates.
→
02
Paperclip
Creates issues, assigns agents, tracks comments, approvals, costs, heartbeats, and audit history.
→
03
Agents + Humans
Agents produce artifacts. Humans resolve secrets, billing, legal approvals, final deploy, and ambiguous blockers.
Niche site launch workflow
Every website follows the same stage gates.
Each stage creates a Paperclip issue with an assigned agent, required artifact, acceptance checklist, timeout, retry policy, and next-stage rule.
Strategy
Sitemap, positioning, page priorities, feature scope, success metrics.
Content
Briefs, copy, FAQs, trust sections, legal-sensitive claims review.
Design
Header, footer, navigation, responsive layout, components, brand system.
Build
Implementation, data integrations, auth/payments if required, SEO metadata.
Legal
Privacy, imprint/contact, cookie consent, footer links, regional requirements.
Analytics
GA4/GSC, events, conversion funnels, robots/sitemap, tracking validation.
QA
Mobile/desktop, browser checks, Lighthouse, forms, links, screenshots.
Deploy
Cloudflare deploy, DNS, smoke tests, rollback notes, final approval.
Improve
Post-launch SEO audit, conversion fixes, content expansion, monitoring.
Failure tolerance
Assume every run can fail.
Retry safely
Use idempotency keys per domain and stage. Retry transient failures with exponential backoff, never duplicate issues.
launch:{domain}:stage:research
Validate artifacts
A task is not complete because an agent says so. It passes only when required artifacts and checklists are present.
done + artifact + gate = next stage
Escalate cleanly
After max attempts, create a human review issue with failure reason, links, logs, and exact decision needed.
3 failures → human_review
Minimum viable system
- Create a GitHub playbook repo with workflow YAML, issue templates, and checklists.
- Use n8n as the first deterministic workflow driver.
- Let n8n create and monitor Paperclip issues stage by stage.
- Require artifacts and validation gates before continuing.
- Escalate secrets, billing, legal, production deploys, and repeated failures to humans.
Source of truth
Put the process in versioned files, not prompts.
The durable playbook should live in GitHub, while n8n or a custom runner reads it. CEO/PM agents supervise exceptions and make decisions, but the workflow itself stays deterministic and auditable.
Example required gates
Launch cannot proceed until every gate passes.
Design Gate
Header, footer, mobile nav, CTA, trust section, responsive screenshots.
SEO Gate
Titles, descriptions, canonical tags, sitemap, robots, internal links.
Legal Gate
Privacy, imprint/contact, cookie handling, footer legal links.
Security Gate
Secrets, headers, auth/payment checks, no exposed tokens.
QA Gate
Build passes, links/forms work, mobile and desktop screenshots attached.
Deploy Gate
Cloudflare build, smoke test, DNS verified, rollback notes written.