Behind the scenes

How the work actually gets done.

The technical side, for anyone who wants it: how I set up AI agents, test models and check every result. You don't need any of this to work with me.

Approach

Out-of-the-box AI does it the generic way. I make it do it yours.

Your systems, terms and standards, written down once as context files and skills that every agent follows.

  1. Break the work into deliverables

    Each recurring output gets an owner, a cadence and a quality bar.

  2. Write your context file

    Systems, data sources, terminology and rules that every agent reads before it starts.

  3. Build a skill for each workflow

    A tested, written procedure: inputs, steps, checks and output format.

  4. Route each task to the right model

    Chosen from my own head-to-head evaluations on quality, cost and time.

  5. Review and deliver

    I check every output against the agreed bar before it reaches you.

Run my workflowPick a task
  1. Context
  2. Skill
  3. Model
  4. Run
  5. Review
0tokens
0%from cache
0:00agent time
Illustrative runs at typical sizes, following my real routing rules.
Model evaluations

I test models before I trust them.

Same tasks, several models, graders written first. Each kind of work goes to the model that earns it.

Real-work suite, 12 frozen cases
Top tierlargest model0/12$0.00
Mid tierhalf the price0/12$0.00
Small tiera few cents0/12$0.00
Passed first try Missed Fixed by mid tier

What came out of it

Work typeRouted toWhy
Writing, design, judgment callsTop tierWhere quality is the product
Bounded build work with checksMid tierSame pass rate, about half the cost
Extraction with full validatorsSmall tierCents per run when a grader catches misses
Anything the small model missesMid tier repairAll five fixed in one attempt each
Agentic workflows

My own business already runs this way.

Agents with defined roles and handoffs, and a review step before anything goes out.

Agent team, one task from scope to shipped
Lead scopes the task, research gathers facts, creative owns the copy, engineering builds, QA verifies independently, the lead accepts and Evan signs off. Leadscopes the task Researchgathers facts Creativeowns the copy Engineeringbuilds, releases QAverifies alone Leadaccepts Evanfinal sign-off
43custom skills across Claude and Codex, each a written, tested procedure
140+project context files holding each project's rules, accounts and quality bar
16scheduled automations, from booking smoke tests to email triage
2MCP servers I built: a CRM agents work in, and a DaVinci Resolve controller
AI usage, by the numbers

80 billion tokens. And counting.

Real numbers from my Codex profile and my Claude Code logs. Here is how that work is split between models, and the skills every agent works from.

80B+lifetime tokensCodex 48B · Claude Code 32B
113day streakdays in a row
104hlongest single taskone agent, one job
611Claude Code sessions143,627 model responses
97%served from cachecontext, re-read cheaply

Which model gets which job

  • Claude OpusThe defaultLeads every task: planning, writing, design decisions and anything where judgment is the product.
  • Claude SonnetRepetitive workBulk edits, data cleanup, template fills and repeat runs that a checker can verify.
  • Claude HaikuExplorationSearches codebases and documents, reads widely and brings back the facts, fast and cheap.
  • GPTAdvisory reviewAn independent second opinion on tasks with high ambiguity, before I commit to a direction.
  • GPT image generationImagesConcept art, product shots and design studies.
  • GPT voiceVoiceNarration, like the intro video, and real-time voice agents that answer the phone.

Whichever model does the work, I review the result before it goes out.

Skills the agents work from

Each skill is a written, tested procedure: inputs, steps, checks and output format.

Show all 37 skills

Websites

  • ShipRuns spec, build, review and deploy for a website change from one prompt.
  • Website creation and deployBuilds or updates a site with design research, local checks and post-deploy verification.
  • Website replicationRebuilds an existing site or screenshot route by route, matched at each screen size.
  • Landing pageThe section-by-section message order for a service business homepage.
  • Design systemLocked type scale, spacing, color and components so every site stays consistent.
  • ImpeccableDesigns, critiques and polishes interfaces: hierarchy, layout, accessibility and detail.
  • UI reviewScores rendered pages against a rubric and returns ship or fix.
  • Watermelon UIA library of polished animated components to build from instead of hand-writing them.
  • Web motion and scroll revealPurposeful animation and one-time scroll reveals that stay fast.
  • Taste referencesWritten references for a specific brand look, used when a brief calls for it.
  • Local SEOBusiness profile, schema, service pages, indexing and reviews for local search.

Words, strategy and sales

  • PositioningA positioning document for each client before any copy or design exists.
  • Copy rubricReviews and repairs marketing copy against a quality bar and a banned-word list.
  • Human writingRewrites prose so it reads natural and specific, not machine-made.
  • One-page marketing planPlans and audits marketing against the nine-block 1-Page Marketing Plan.
  • Customer job evidenceFinds out why customers choose, delay or switch, using Jobs to Be Done interviews.
  • Business strategySparring on pricing, packaging, niching and client acquisition, with a committed answer.
  • Authority salesFounder-led selling through specialization, diagnosis and proof.
  • Prospect fit gateQualifies a prospect before any custom demo work starts.
  • Prospect demosBuilds and deploys tailored demo sites from a lead list and approved templates.
  • Seasonal revenuePricing and demand planning for a seasonal rental and retail business.
  • Marketplace pilotDesigns and measures a local two-sided marketplace pilot.

Video, voice and 3D

  • Explainer videosNarrated business explainers built from real brand assets and site screenshots.
  • Training videosScreen-recorded walkthroughs and onboarding videos for software and processes.
  • Short-form studioScripts, directs and edits TikToks, Reels and Shorts for a business.
  • Video remixAdapts an idea, reference video or existing footage into short promos.
  • Text to speechGenerates natural voiceover audio from a script.
  • 3D property modelsAn editable 3D model and walk-through viewer from floor plans and photos.
  • Town-scale 3DRealistic whole-town Blender scenes and flyovers at a sensible render cost.
  • 3D knowledge graphTurns a body of notes into an interactive 3D map.
  • Animated mascotsGenerates, animates and packages a character from brand art.

Agents and operations

  • Work orchestratorPicks the model for each piece of work, delegates it and arranges independent checks.
  • External councilGets independent advice from other models on a plan or decision, then synthesizes it.
  • System auditReviews the reliability of the agent setup itself: skills, routes and manuals.
  • Task control planeCoordinates assignments, status and handoffs between agents.
  • Release safetyDeploys, verifies in a real browser and records each production release.
  • Account routerSends every deploy and hosting command to the right account for that project.
Creative lab

Music, animation and 3D, made with the same agents.

Hallucination Hits, my AI music channel: Suno songs, animated entirely in code.

Claude vs ChatGPT: Number One (For Now)

An AI rap battle between two assistants. My most-watched video.

  • Suno
  • Code-drawn animation
  • ffmpeg
Cutaway of the Rundle Bike and Ski shop model: rental ski racks, boot shelves, checkout and coffee counter

Rundle Bike & Ski in 3D

A concept model of the Canmore ski shop, built in Blender from the floor plans and photos of the space: rental ski racks, boot wall, checkout with coffee, and upstairs storage. Dimensions are approximate.

  • Blender
  • three.js
  • glTF
All videos on the Hallucination Hits channel

Jingle ads for local trades

Thirty-second spots with an original jingle and a 3D marble animation, made for Sophon Studio's social channels. Around fifty produced so far.

  • Suno
  • Blender
  • Code animation
Career, skills and builds

Everything connects. Explore it.

My career as one map. Pick a role to see what it led to, or switch to the timeline.

Browse as a list

Have work that needs doing?

Tell me the problem. I'll come back with a plan and a fixed price.