A working thesis from

You can't beat the model.
You can build the workflow that wins around it.

Frontier labs are years ahead, and they're still accelerating. The opportunity for small and mid-sized teams isn't to compete with the model — it's to wrap it in workflows so specific, so well-instructed, that the result is repeatable, defensible, and yours.

01The Challenge & The Opportunity

No one is going to out-build OpenAI, Anthropic, or Google. The win is in how you wrap them.

Frontier models are self-improving on a curve no team can match. But the model alone is not the product. The result is the product — and the result depends on the context, the instructions, and the workflow you build on top.

The Trap

Treating the chat box as the product

Most teams paste a request into ChatGPT or Claude and accept whatever comes back. The output isn't curated — it's the model's best guess at what you might have wanted. Quality, voice, and structure drift every time.

The Unlock

Reverse-engineering from a proven win

Start from a result you'd put your name behind. Decompose every step that produced it. Document each step's required skill, context, and inputs. Now you have a workflow a model can actually run — repeatedly, accurately, on rails.

Speed
Workflows compress days into minutes.
Accuracy
Documented rails replace hopeful prompts.
Leverage
One spec, infinite runs, scaling output without scaling headcount.
02Three Tiers of AI Interface

The model is the engine. The interface decides what you can build.

Every major lab now ships at three levels. Each tier trades setup for control. Knowing where you sit — and where you should sit for any given workflow — is half the battle.

What it looks like
New chat
Ask anything…
Chat · A single conversation
Tier 01 · Chat

The Consumer Layer

A single conversation window. You type, it answers. No project memory beyond the chat, no tool execution beyond what the provider exposes in-app.

A brilliant intern with a fresh notebook every morning.
Who plays here
OpenAIChatGPT
AnthropicClaude.ai
GoogleGemini
PerplexityPerplexity Search
MicrosoftCopilot
Pros
  • Zero setup — open the tab, start working.
  • Fastest path from question to answer.
  • Great for one-off thinking, drafting, brainstorming.
  • Cheapest tier; usually a flat consumer subscription.
  • Newest models land here first.
Cons
  • Output quality is whatever the model guesses you wanted.
  • Limited persistent context — no real org memory.
  • Hard to enforce voice, format, or quality standards.
  • Not repeatable — every run drifts.
  • No real workflow logic; no handoffs to other agents.
Best fit
Individuals exploring, drafting, or thinking out loud.
Graduate when
You find yourself pasting the same instructions, examples, or files over and over.
Comparison
Dimension
Tier 1 · Chat
Tier 2 · Co-work
Setup time
Seconds
Minutes
Customization depth
Low
Medium
Repeatability
Drifts every run
Stable within a project
Cost model
Flat subscription
Subscription + light usage
Best for
Solo thinking
Small teams
Workflow automation
None
Light, in-app
Learning curve
None
Light
A note on Anthropic

I'm bullish on Anthropic specifically because all three tiers — Claude.ai, Claude Projects/Skills, and Claude Code — share the same model family and integrate cleanly. You can prototype a workflow in a Project, then graduate the same instructions and skills into Claude Code without rebuilding. Most other vendors force a context break between tiers.

03The Method · Workflow Development

A workflow is just an SOP that a model can run. Build the SOP first.

The mistake most people make is starting with the model. Start with the win. Then the steps. Then the skills. The model is the last thing you add — because by then, you know exactly what to ask it for.

01

Start from a result you'd put your name behind

Don't ask the model what good looks like. Find — or produce by hand — one example of the exact output you want to ship. Voice, format, depth, structure. This is your north star.

02

Reverse-engineer the steps that produced it

Walk backward from the win. What context was needed at each step? What decisions were made? What inputs fed which outputs? Document like you're writing an SOP for a sharp new hire.

03

Identify the skill required at each step

Each step has a discipline behind it — research, copy, analysis, design judgement. Name it. Granular skills are what you instruct the model on.

04

Write the rails — instructions, examples, constraints

Per step, codify: what context to pull, what to do with it, the format of the handoff to the next step, and the failure modes to avoid. This is the rail your agent runs on.

05

Run it as a workflow, not a prompt

The sequence of steps becomes a single workflow that an agent can execute end-to-end. You're no longer prompting — you're operating a process.

06

Compose workflows into orchestration

Once you have niche workflows for the parts of your business that matter, build an orchestration agent on top. You speak to it; it fires the right workflow at the right moment.

The Compounding Effect

Every workflow you ship makes the next one cheaper. Skills become reusable. Context becomes shared. The org's accumulated process becomes a moat.

Inside a Future Workflow

How we build it in Claude Code

A workflow is a folder of markdown — not a tool, not a framework. Claude Code reads instructions at runtime. If something's off, you edit the file. The next run picks up the change.

The four-folder shape
[workflow-name]/
├─ inputs/# raw source material
├─ workspace/# scratchpad — agents write here
├─ outputs/# final, client-facing deliverables
├─ session-log.md# log of changes to the workflow itself
└─ .claude/# the instruction layer
   ├─ CLAUDE.md# job, sequence, gates, rules
   ├─ agents/# specialist instruction sets
   ├─ shared/# style guide, standards, examples
   └─ skills/# user-invokable phase commands
The pipeline
/intake
Phase 1
readsinputs/
writesworkspace/research-brief.md
/research
Phase 2
readsresearch-brief.md
writesworkspace/[research]
/build
Phase 3
readsworkspace/[research]
writesworkspace/[draft]
/export
Phase 4
readsworkspace/[draft]
writesoutputs/[final]
Skill / the verb
What the operator invokes.

One slash command per phase. The conductor of the run. /intake, /research, /build, /export.

Agent / the expert
What the skill consults.

Specialist instruction sets. Never user-invoked. A /research skill might pull in a guest-researcher and a transcript-analyst.

Source of truth
Everything is markdown.

No code, no compile step. Edit the file to change behavior.

Human cadence
Operator drives phases.

A checkpoint between every phase. Quality drift gets caught early.

Loud failure
Gates fail loudly.

If a check fails, the skill stops and points to the phase to rerun.

Anti-pattern

“Hey ChatGPT, do this for me.”

You'll get something. It might even be good. But it won't be repeatable, and it won't be yours. Service businesses live and die on consistency — that's exactly what undirected chat can't give you.

The Right Pattern

“Run the onboarding workflow for this client.”

One trigger fires a documented sequence: profile lookup, competitor research, agenda generation, project setup, kickoff email. The output is the same caliber every time because the rails are the same every time.

04The Company Brain

Workflows are useless without the data they need to fire on.

If your information lives in twelve disconnected tools, your workflows can't reach it. The single biggest unlock — bigger than any model upgrade — is consolidating your context into a structured, queryable spine. Everything plugs into it. Everything writes back to it.

The Company Brain · Schema
Profiles
Transcripts
Strategy
Projects
Assets
Comms
Research
Outputs
Company
Brain
One source of truth · Every workflow reads from it · Every workflow writes back to it
Worked example · Onboarding a new client

Most workflows start with a trigger. The trigger needs context. The context lives in the brain.

Imagine the moment a new founder signs. The trigger fires the onboarding workflow. The workflow needs an accurate business profile, a voice profile, the call transcript, the avatars, the competitor set. If those documents are scaffolded in a known place for every client, the workflow runs cleanly. If they aren't, it stalls before it starts.

Get the scaffold right once. Standardize the folder structure, the documents that must exist, the format they take. Now any workflow you build downstream can assume that data exists, in that shape, in that place.

Trigger flow
01
Trigger

Founder signs the agreement.

02
Read context

Pull profile, transcripts, voice doc, avatars from the brain.

03
Run sub-workflows

Competitor research · agenda draft · project setup · kickoff email.

04
Write back

Update the brain with new artifacts so the next workflow inherits them.

Per-client scaffolding · A starting checklist
Business profileWho they are, what they sell, who they sell to
Voice profileTone, banned phrases, reference samples
Avatar / ICP docDetailed buyer personas with motivations
Competitor setNamed, profiled, refreshed on a cadence
Call transcriptsEvery call, dated, in one folder
Asset libraryLogos, brand kit, prior deliverables
Strategy docLiving document, single source of truth
Project trackerStatus, owners, due dates, links
Comms logKey threads pulled in, tagged, searchable
One Last Thing

All of this is moving fast. The principles aren't.

New tools will land every quarter. Tiers will blur. Capabilities I haven't seen yet will change the math on specific decisions. But the core sequence — start from the result, decompose it, write the rails, centralize the context, orchestrate workflows — is the same playbook regardless of which model is on top.

The teams that win in the agentic era aren't the ones with the cleverest prompts. They're the ones with the most disciplined process and the cleanest company brain.

Don't
Compete with the model. You can't and won't.
Do
Wrap the model in workflows that produce your result, every time.
Then
Compound — orchestrate workflows together, sync everything to the brain.