The master plan: building a consulting business in public

I'm going to lay out the entire plan for building Black Hills Labs into a real consulting business. Not the polished version. The actual plan, with the numbers, the timelines, and the things that haven't worked yet.

Why? Because I think the most useful thing I can share right now isn't another tutorial. It's what the process of building a business with AI tools actually looks like from the inside.

In this article:

• The goal
• The four workstreams (outbound, inbound, sales ops, delivery)
• What's actually working
• What's coming next
• Follow along

The goal

Five paid consulting engagements by March 31, 2026. Revenue target: $10K-$40K total, at $2K-$8K per engagement.

That's seven weeks from when we started. We're in week two now.

The business is operations automation and sales acceleration for startups and SMEs in the GCC region. We use a simple stack that we've tested with clients and with our own business (Lexacon, BlackHills): Claude Code + n8n + NocoDB + email tracking + CrewAI. We help small teams build the operational infrastructure that enterprise companies take for granted, but at 10% of the cost. Discovery workshops, rapid prototyping sprints, environment setup, ongoing advisory.

Everything runs on a $5/month server. That's not marketing. That's the actual infrastructure and software bill (LLM credits separate and depend on your usage).

The four workstreams

We broke the launch into four parallel tracks. Each one builds on the others, but they can progress independently.

Outbound sales engine

The problem: we can research prospects and draft messages, but if we're doing it manually, we max out at maybe 10 per session. That doesn't generate enough pipeline.

What we've built so far:

• Five autonomous n8n workflows that handle lead discovery, contact enrichment, message drafting, email sending, and reply monitoring. All tested end-to-end.
• A policy table in NocoDB that controls daily email caps, market-specific rules (UAE requires human approval for outreach), and feature flags. Change a number in the database, the agent's behavior changes. No code edits.
• A validator agent that checks every drafted message before it goes out. Word count, formatting, case study relevance, CTA quality, AI-tell detection. It catches about 30% of first-draft issues.
• Reply classification via IMAP monitoring. When someone responds, the system classifies it (interested, question, objection, out of office, bounce), extracts BANT signals, updates the pipeline stage, and alerts me if it's urgent.

The whole thing runs on the same $5/month VPS. LLM costs via OpenRouter add $15-30/month depending on volume. Compare that to $350-10,000/month for SaaS alternatives like Instantly, Apollo, or Clay.

The real test: seeing if the first wave of outreach converts to discovery calls.

Inbound and value hooks

The problem: every outreach message we send, the recipient checks our website. If they land on a generic page with a "contact us" form, they bounce. We need immediate value.

What we've built:

• An AI readiness quiz that scores visitors on automation maturity, captures their email, and routes the data to NocoDB for lead scoring. Three result tiers with personalized service recommendations.
• A gated starter kit with a full Claude Code walkthrough, companion guide, and sample project. Take the quiz, provide your email, get the download.
• Persona-specific landing pages: /for-startups and /for-business-owners with different case studies, pain points, and CTAs. Research showed persona-matched pages convert significantly better than one-size-fits-all.
• A persona evaluation system that tests our website copy against synthetic personas before we deploy. Three personas (a brand strategist, a pre-PMF startup founder, and a non-VC SMB owner) score each page on seven dimensions. We use this to iterate on copy before it goes live. Our scores went from 17/35 average at baseline to 23/35 on the persona pages.
• Two blog posts (this is the third) documenting the build process.
• n8n webhooks that capture every quiz completion and booking form submission, create records, and score the lead automatically.

All text on the website is CMS-editable via Keystatic. I can change any headline without touching code.

What's next: nurture email sequences (warm and cold), a recorded getting-started webinar, an e-invoicing readiness checklist (UAE compliance starts July 2026), consistent LinkedIn content (3-5 posts/week), and a paid community with exclusive pre-made tools and content to support clients and nurture new leads.

Sales operations and closing

The problem: no systematic follow-up, no proposal process, no pipeline visibility.

What we've built:

• A proposal generator that produces branded PPTX decks and WBS spreadsheets. Our first real proposal (Dominican National Election Board support app, ~$25k) was built in a single afternoon using Claude Code.
• NocoDB pipeline tracking across six stages from research through closed-won, with temperature scoring and deal values.
• Reusable templates for each service tier (Discovery Workshop, Rapid Sprint, GTM Sprint, Environment Setup, Monthly Advisory).

What's next: automated stale-deal alerts, weekly pipeline reports, Cal.com for self-service booking, and a notetaking tool for discovery calls to auto-generate requirements and validate the roadmap. The real gap here is follow-up automation. Right now, follow-ups are manual.

Delivery infrastructure

The problem: when we close a deal, we need to onboard and deliver without scrambling.

What we've built:

• A discovery session framework distilled from four real sessions. Five-step process with setup guides, guided prompts, and follow-up templates.
• Pre-session setup guides for Mac and Windows.
• A session package skill in Claude Code that automates building the client delivery package.
• A generic starter kit that's session-resilient (works regardless of what we cover in the call).

What's next: onboarding questionnaire, engagement contracts, invoicing, per-tier delivery templates, and satisfaction tracking. Most of this can be built proactively before the first deal closes.

What's actually working

Two weeks in, here's what I can say honestly:

The autonomous workflows are real. Our outbound discovery workflow discovered and classified three new companies on its first real run. Our reply classification workflow correctly classified an "interested" reply, extracted BANT signals, updated the pipeline, and sent me an alert, all in 2.5 seconds. The policy table pattern means I can tune behavior without touching code.

The persona evaluation system changed how we write copy. Instead of guessing whether a landing page resonates with a pre-PMF founder, we test it. The evaluation score jumped from 19/35 to 23/35 from the persona agent for this ICP after building the persona page. That's still not where we want it, but the feedback loop is tight.

The CMS was the right call. Keystatic + Astro means content changes are lighting fast without a heavy CMS behind it. This matters when you're iterating on messaging weekly.

The $5/month infrastructure holds up. Caddy + Postgres + NocoDB + n8n + CrewAI + the Astro website, all on one Hetzner VPS. We hardened it with SSH key-only auth, UFW, Tailscale for admin tools. It hasn't needed attention since setup.

What's coming next

Here's what we're planning to write about. Some of these are almost done, others are weeks out:

• "Building a $30/month autonomous sales agent with n8n": the full Phase 1 story. Five workflows, policy tables, cap gates, LLM classification, reply monitoring. The most technically interesting thing we've built so far.
• "How I built a $25K proposal in one afternoon": mockup generation, PPTX automation, Playwright screenshots. The JCE story.
• "Testing your marketing with AI personas before you ship": the evaluation system, how to avoid sycophantic feedback, score-driven iteration.
• Guardian.AI deep dive: 10 sales, 30% conversion, 1 person, $30/month in tools. The case study in full.
• "The $100K consulting gap": why SMEs are underserved by AI consultants and what it takes to serve them at $2K-$8K.
• UAE e-invoicing automation: what July 2026 means for SMBs and how to prepare.

If any of those are particularly interesting to you, let us know. We'll prioritize based on what people want to read.

Follow along

This is episode three of the "Building with Claude" series. Episode one was the announcement and Lexacon story. Episode two covered Claude Desktop vs Claude Code vs Cowork (now updated with a comparison to Cursor, Gemini, Codex, and Aider).

We're posting the build process as it happens. Not retrospectives. Not polished case studies written six months later. The actual decisions, mistakes, and metrics as they unfold.

Take the AI readiness quiz to see where your team stands. Or grab the free starter kit and try the tools yourself.

If you want to skip ahead and get hands-on help, book a discovery workshop. Half a day, $1,500, and you'll leave with a prioritized automation roadmap and your first working prototype.