Author.AI · The Operating System

One person. Seven agents.
One bar.

How a PMM-of-one built an operating system , a permanent instruction set and a seven-agent stack , to ship better work than teams ten times the size. Taste stays human. Everything else scales. Proof shipped · Spry case below

Interview Simran.GPT ↗ See the stack ↓
01 · Thesis

The principle

Why most AI marketing work looks like AI marketing work , and how to build it so it doesn't.
Work done faster with AI, not work done by AI.

Human does

  • Strategy, direction, the sharp angle
  • Taste, judgment, what stays and what goes
  • Voice, essence, the human signature
  • Every decision that has a consequence

AI does

  • Execution and formatting at scale
  • Data processing, competitive scans
  • Repetitive synthesis across sources
  • Drafting inside tight, specified constraints

AI is a taste amplifier, not a content factory. Default LLM output is the population mean of the internet. Left alone, it reverts to mush. The job is to starve it of defaults.

Taste is the scarce resource. Anyone can generate. Fewer can direct. One person making every call, seven agents executing downstream.

Constraints beat creativity. Better work comes from removing options, not adding them. The BAN FILE is what not to do. What's left is a narrow channel the model has to be interesting inside.

02 · Instrument

The BAN FILE

Brand guidelines, but for AI. A permanent instruction set that teaches every agent my voice, my standards, and what never ships.

What most people miss about prompting.

Most people prompt an LLM like they brief a freelancer , once, then hope. The freelancer has a resume. The LLM has nothing.

The BAN FILE fixes that. A permanent document , loaded into every agent , that codifies what my voice is, what it isn't, and what gets rejected on sight.

Forbidden words. "Leverage." "Unlock." "Game-changer." The full list of AI tells.

Banned frameworks. No three-point listicles. No "intro / body / conclusion."

Voice anchors. Sentence variance. Em-dashes with restraint. Concrete over abstract.

Every agent reads this file before it writes a word.

BAN_FILE.md · v4.2
# Voice , Simran Raj Bhyana identity: "narrative-first PMM operator" register: "dry · precise · confident · no preamble"   # Forbidden words (full list: 120+) ban:   - "leverage"   - "unlock"   - "game-changer"   - "in today's fast-paced world"   - "at the end of the day"   - "seamless / seamlessly"   - "revolutionise"   - "in conclusion"   - "... and more"   # Frameworks to avoid no_patterns:   - three-point listicles as body   - "intro / body / conclusion"   - opening with the topic restated   # Voice anchors prefer:   - specificity over abstraction   - active verbs   - sentence-length variance   - em-dashes , used with restraint   - italics for emphasis, not decoration   # Output bar done_looks_like: "a sentence that earns its place"
03 · The stack

Seven agents, one director

Specialised agents. Each one narrow. Each one readable. The director is human.
01 / RESEARCH

Research agent

NotebookLM
Ingests 1000+ documents , interviews, reports, transcripts. Extracts patterns, surfaces quotes. Retrieves and synthesises. Never generates.
02 / SCRAPE

Scraping agent

ChatGPT · BrowseAI
Web-scrapes competitors, pricing, LinkedIn. Structures raw HTML into briefing-ready tables downstream.
03 / PROMPT

Prompt meta-agent

Claude
Writes prompts for the other agents. Specifies frame, constraints, success criteria , before a word ships.
04 / POSITION

Positioning agent

Claude + BAN FILE
Trained on 50+ brand briefs. Applies Obviously Awesome, Crossing the Chasm, JTBD frames , consistently.
05 / FORMAT

Formatting agent

Claude
Structures content to Simran standards. Message hierarchy holds. Sentence-length variance lands. Em-dashes stay rare.
06 / DESIGN

Design agent

GenSpark · Figma
Decks, collaterals, one-pagers. Works inside a defined system , never invents new palettes.
07 / WRITE

Writing agent

Claude Opus · BAN FILE
Last mile. Final copy, headline, deck slide. Tight constraints from the previous six. The only output that ships.
00 / DIRECTOR

The human

Simran · always
Sets the frame. Approves direction. Kills what's wrong. Every consequential call. The seven scale. The director never does.

A real workflow , Spry spec assignment

The same stack, 20+ hours of work, compressed into a week.

01
Frame

The sharp angle. What we're actually arguing.

HUMAN
02
Research

Notebook LM across 40+ sources.

AI
03
Scan

7 competitors structured into a landscape.

AI
04
Judge

Which positioning angle wins. What stays, what goes.

HUMAN
05
Draft

Three messaging pillars, proof points under each.

AI
06
Edit

BAN FILE pass. Every line earns its place.

HUMAN
07
Ship

Deck, website critique, sales scenarios.

BOTH
04 · Proof · Spry assignment

Receipts from the stack

The PMM spec assignment for Spry Therapeutics. Same seven-agent stack, public artifacts.
Claude artifact

Competitor table

8 competitors × YouTube + product + observational insights. BrowseAI + NotebookLM + 3-prompt framework.

Open artifact ↗
Google Sheet

ICP matrix

5 × 3 × 4 segmentation of US outpatient rehab. Unit economics and bottleneck scoring. The top ICP falls out of the math.

Open sheet ↗
InVideo AI

AI-generated pitch film

The Spry positioning, in 60 seconds, without a shoot. Generated end-to-end by the stack.

Watch film ↗
Full case study

Spry , the whole thing

Method, positioning, ICP, competitive frame, messaging pillars, narrative architecture. The stack, applied.

Read the case →
05 · The agent you can talk to

Simran.GPT

Most portfolios are static. This one talks back.

Don't take my word for any of this.

Simran.GPT is a trained agent running on the full corpus , 50+ brand briefs, career timeline, positioning frameworks, every interview prep note. Loaded under the BAN FILE. Interview-ready, on-brand, available now.

Ask it why I'd pick HNI buyers over NRI investors. Ask it how I'd position Writesonic for the GEO shift. Ask it what's in the SPRY competitive landscape. It'll answer in my voice, with my receipts.

Full career corpus BAN FILE loaded Positioning frameworks Interview mode Case-study recall
Open Simran.GPT ↗
06 · The manifesto

Three things I believe

The operating assumptions behind everything above.
01

AI should make you pickier, not faster.

The speed gain is a trap. The real unlock is that you can now afford to reject the first eight drafts. Volume is free; taste is expensive. Spend the surplus on standards.

02

Specialist agents beat super-prompts.

One narrow agent doing one narrow job, well, beats a mega-prompt trying to do everything. Split the work. Constrain each role. The seams are where quality comes from.

03

The bar is what doesn't change.

Models update. Tools churn. Agents come and go. The BAN FILE , the standard for what's good , is the only permanent asset. Portable across any stack. Outlasts every tool.