I shipped a 20-feature DevOps tool solo in 60 days using Claude Code — the unfiltered breakdown

ressions, automated coverage patterns) |

Key Findings:

• Frontend Acceleration: Vue 3/TS/Tailwind component development dropped from 2–3 weeks to 2–3 days by offloading framework convention cognitive load to Claude while retaining architectural control.
• Scaffolding Velocity: Blank-file-to-working-draft for complex modules (MCP/OAuth2, OCI manifest parsing, CVE risk scoring) compressed from 2–5 days to an afternoon.
• Refactoring Safety: A 600-line main.go with circular dependencies and interface{} event buses was restructured into typed events, App container extraction, and DI patterns in 4 days using a custom go-refactoring skill.
• Verification Cost: Hallucination rate was low (2 incidents over 263 commits), but scope drift/misunderstood requests dominated (56 incidents). Verification overhead remains constant and does not decrease with usage.
• Sweet Spot: LLM-augmented development excels when paired with strict constraint prompting, spec-driven task breakdowns, and custom skill libraries that encode team/product conventions.

Core Solution

Architecture & Runtime

• Single Go binary, ~17 MB RAM at idle, zero external dependencies (no Postgres, Redis, or message queues).
• SQLite with WAL mode for persistent state; Vue 3 + TypeScript + Tailwind frontend embedded via embed.FS.
• One Docker image, one process, zero orchestration overhead.

Observation Layer (Read-Only Principle)

• Auto-discovers Docker, Swarm, and Kubernetes workloads.
• Monitors HTTP/TCP endpoints, cron heartbeats, TLS certificates, resource usage, and image update availability.
• Scans for dangerous network configurations: exposed database ports, 0.0.0.0 bindings, privileged containers.
• Strict observe, never act contract: no restarts, pulls, or mutations. Read-only access ensures security, simplifies the trust model, and reduces blast radius.

AI-Augmented Development Workflow

• Custom Claude Skills: go-refactoring (typed events, DI, test coverage patterns), mobile-first-audit (Vue/Tailwind conventions), linkedin-post-generator (tone/style enforcement), linkedin-lead-capture, linkedin-trend-research.
• Spec-Driven Development: GitHub Spec Kit workflow enforces need → plan → atomic tasks breakdown before any code generation. Prevents output scattering and forces architectural alignment.
• Prompt Discipline: Explicit constraint injection per task ("you touch X, you do not modify Y, and you don't factor out anything that wasn't explicitly requested"). Debug prompts mandate reproduction before correction.
• Business & Licensing: AGPL-3.0 open-core model, Pro tier at €29/month. Ed25519-signed license keys delivered via environment variables. Marketing site built on Hugo + PocketBase handling Stripe checkout and verification.

Pitfall Guide

1. Unbounded Scope Drift: LLMs default to over-engineering and adding unrequested abstractions. Fix: Inject explicit boundary constraints at the start of every task ("modify X, never factor out Y").
2. Feature Hallucination & Verification Fatigue: AI generates plausible but non-existent features (e.g., Telegram/Gotify alerts). Fix: Mandatory README/codebase verification for every output; treat verification cost as constant, not diminishing.
3. Premature Root Cause Diagnosis: LLMs jump to fixes without reproducing bugs or reading logs. Fix: Enforce "reproduce concretely, read logs, test request before proposing correction" in all debug prompts.
4. Strategic Misalignment: LLMs optimize for technical feasibility, not market positioning (e.g., exploring SaaS when target audience demands self-hosted data sovereignty). Fix: Frame business constraints and architectural non-negotiables before technical exploration.
5. Custom Skill Maintenance Overhead: Skills encode conventions but require versioning and upkeep alongside the product. Fix: Treat skills as living code; store them in version control, update them during refactors, and audit them quarterly.
6. Spec-Driven Discipline Requirement: Without structured breakdowns, AI outputs scatter across unrelated files and patterns. Fix: Adopt GitHub Spec Kit workflow; never prompt without a completed need → plan → atomic task breakdown.

Deliverables

• Blueprint: maintenant-architecture-v2.pdf — System diagram showing Go binary runtime, SQLite WAL storage, embed.FS frontend injection, K8s/Docker auto-discovery pipeline, and MCP/OAuth2 integration flow.
• Checklist: llm-dev-verification-checklist.md — Pre-prompt constraint validation, scope boundary confirmation, reproduction-before-fix enforcement, and post-generation README/codebase cross-reference steps.
• Configuration Templates:
  • .claude/skills/ directory with go-refactoring.md, mobile-first-audit.md, and constraint-prompt templates
  • spec-kit-task-breakdown.yaml — GitHub Spec Kit compatible task structure for atomic feature delivery
  • go-embed-fs-scaffold/ — Minimal Go project template demonstrating embed.FS frontend integration, SQLite WAL initialization, and read-only monitoring contract enforcement