Vibe coding, multi-model workflows, CLI agent orchestration, and what it actually looks like to use Claude, Mistral, and Codex as real development tools — not demos.
Vibe coding is a real thing and a stupid term at the same time. The real thing: using AI as a pair programmer — typing in intent, iterating fast, letting the model handle boilerplate so you can stay at the architecture and logic level. The stupid part: when people think "vibe coding" means letting the AI write everything while they watch.
The difference is whether you understand what's being generated. If you can't read the output, debug it, or explain why it works, you don't have a workflow — you have a liability. Every piece of AI-generated code in my projects gets read, tested, and signed off by me before it ships.
With that constraint in place, the speed gains are real. Schema generation for 30 pages. Resource summarization pipelines. Boilerplate PHP includes. Test scaffolding. Things that used to eat an afternoon now take 20 minutes — and the result is the same quality because the judgment layer didn't get outsourced.
AI handles the mechanical parts — boilerplate, repetitive structure, format conversion. The developer handles architecture, logic, and review. Both parts are necessary. Neither replaces the other.
Every AI-generated line gets read before it goes into a project. Not skimmed — read. This is the discipline that separates a reliable AI workflow from a black box that works until it doesn't.
Short loops beat long prompts. Ask for one thing, test it, redirect. The model's third attempt informed by two rounds of feedback is better than its first attempt at a 10-part prompt.
Different models have different strengths. A good AI workflow uses each one where it performs best — not just the latest, flashiest option for everything.
My primary AI — used via API, via the Claude Code CLI, and via agent swarms. Strongest for: reasoning through complex problems, generating structured JSON/schema output, long-context document work, and code that needs to be explained not just produced.
Mistral models (Mistral Large, Mistral Small, Codestral) are excellent for fast turnaround on structured tasks — classification, quick extraction, lightweight code generation, and batch jobs where API cost and latency matter. Codestral specifically for code-focused completions.
GPT-4o (successor to Codex) is in the stack for cross-referencing, getting a second opinion on complex logic, and interop with OpenAI-native tooling. Not my primary — but a useful validation layer when Claude's output needs a sanity check from a different model.
Claude Code is Anthropic's CLI tool that puts Claude directly in the terminal as a coding agent. It reads your files, writes and edits code, runs commands, and works in the context of your actual project — not a sandboxed chat window.
This is not "ask Claude a question and copy the answer." The agent operates in the working directory. It can read any file, make edits, run bash commands, check output, and iterate. It's closer to a pair programmer with terminal access than a chatbot.
Every action the agent takes is visible before execution. Bash commands, file writes, and edits are shown with a confirmation step. This is the safety layer. An agent with unrestricted bash access and no review step is a liability — with review, it's a multiplier.
A single Claude agent is powerful. Multiple Claude agents running in parallel — each with a specific task, a defined scope, and their own context — is an agent swarm. This is how large or multi-part tasks get done without a single context window becoming a bottleneck.
Spawn multiple agents simultaneously for independent tasks — one researching, one writing, one building. Each operates in its own context. Results come back in parallel, not sequentially. A 4-hour single-agent job becomes a 1-hour multi-agent job.
Each agent gets a specific, bounded task: "audit these 10 pages for schema errors," "generate JSON-LD for this page type," "find all PHP files that use this function." Tight scope = predictable, reviewable output.
Agents that write code run in isolated git worktrees — separate branches that don't touch the main codebase until reviewed. Multiple agents can work on different features simultaneously without conflict. The main branch stays clean.
One orchestrator agent manages the overall task, breaks it into subtasks, and delegates to worker agents. The orchestrator synthesizes results. This mirrors how a project manager works — except everything runs in the terminal and the agents don't need standups.
Agent swarms use built-in task tools to create, update, and complete discrete units of work. Each subtask has a status (pending, in-progress, complete). The orchestrator checks task state before delegating. Nothing falls through the cracks.
Claude Code maintains file-based project memory across sessions and across agents. Architecture decisions, deploy commands, project conventions — stored and recalled automatically. Agents starting cold still have full project context.
This site was rebuilt from scratch — 31 PHP files, full PHP architecture, 10 project pages, CSS, JS, schema on every page, sitemap, robots.txt, llms.txt, and deploy — using Claude Code CLI in a single working session. The agent read the existing site, built the staging structure, ran PHP syntax checks, created a backup, and deployed. Every file was reviewed before going live.
14 category pages, each needing custom JSON-LD with the right schema type, breadcrumb, and entity references. Claude API pipeline: page structure in, valid schema out. Batch-processed with incremental saves so a failure at page 8 picks up at page 9. Total time for schema generation: under 20 minutes. Manual: half a day.
Dozens of Denver homeless resource providers — each needing address, hours, eligibility, phone, and service type verified from government and nonprofit sites. Python script fetches each page, extracts text, sends to Claude with a structured prompt, gets back a JSON object that feeds directly into the data file. 15 minutes of manual work per provider → 90 seconds.
One agent audits PHP files for security issues (output escaping, SQL patterns, file exposure). A second agent simultaneously audits schema markup across all pages. A third searches for broken internal links. All three run in parallel. Results synthesized by the orchestrator. One task that would take an hour sequentially — done in 15 minutes.
If you have repetitive, structured work that currently eats hours — content pipelines, data extraction, schema generation, batch processing, code scaffolding — there's a good chance it can be turned into an AI-assisted workflow that runs in minutes.
I build these workflows using the same stack described on this page: Claude API for reasoning and structured output, Mistral for fast batch tasks, Python for orchestration, and Claude Code CLI for development itself.