Proposal Engine

By Luther Roberts · Business Development, Sprint Couriers, Gaborone, Botswana

Somebody forwards you a quote a competitor sent them. Thirty minutes later a personalised eight page proposal is sitting in your drafts folder with the PDF attached, ready to read and send.

It never sends by itself. It never calls a language model. Everything in the loop is deterministic, which is why it can run every thirty minutes forever and cost nothing.

Built for a courier business in Gaborone, Botswana. The engine is industry agnostic.

---

Why I built it

I kept losing work to quotes I never saw, and winning it when I happened to know something about how the customer's week actually ran. That knowledge does not scale by working harder. It scales by being written down once per industry and reused.

So the expensive thinking happens once, for a trade, and every company in that trade after that costs nothing. Everything else is plumbing.

The first version was five airy pages and it was rubbish. It looked designed and said nothing. What is here now is dense on purpose, because the people who sign these read documents for a living.

— Luther Roberts

---

The idea

Most proposals lose because they answer a price with a price.

This answers a price with an understanding of the buyer's week. It holds one industry playbook per trade, and a playbook knows four things a supplier almost never knows:

1. The recurring calendar. Dates that repeat whether anybody feels like it or not, each with the published consequence of missing it.
2. The clock. The deadline inside that profession that outsiders have never heard of. For external auditors it is the standard that caps final audit file assembly at sixty days after the report date, with a five year retention rule behind it and a regulator that inspects how it was done. Every regulated trade has one. Naming it is what earns the meeting.
3. Where documents and goods actually travel. One hub and six destinations.
4. Five things that go wrong, what each costs, and what replaces it.

Write that once per industry. Every company in that industry afterwards assembles for free.

Where the money goes

Step	Cost
Watching the mailbox every 30 minutes	free
Working out who the quote is from and which industry they are in	free
Building the eight page proposal and rendering the PDF	free
Staging the reply as a draft with the PDF attached	free
Writing a new industry playbook	one sitting, once per industry, forever

The document

Eight A4 pages, built as a consulting exhibit pack rather than a brochure. Every page carries a kicker, an action title, dense content and a sourced footer.

Page	Comes from
Cover, with a document control block and contents	the company file
Their recurring calendar, drawn as bars	the industry playbook
Their clock, the deadline inside their trade	the industry playbook
Document movement, hub and spoke	the industry playbook
Five failure modes and what each costs	the industry playbook
What the supplier is, tested the way a buyer would test it	constant
Custody, confidentiality and data protection	constant
Engagement, three steps and one question	constant

An earlier version of this was five sparse, airy pages. It was rejected for looking like it had nothing in it. Density plus domain fluency is the whole point, so the design is fixed in style.css and build.py and is never re-decided per client.

Rules the build enforces

These are not style preferences. The build fails or refuses on each one.

• No tariff, ever. A price printed before the runs are measured is a number the buyer cannot rely on, and they find that out at the worst moment.
• Every figure carries its source on the page. Nothing is typed from memory.
• No claim about the buyer's own operation. You have not been inside it, and the document says so.
• Zero dash characters in anything a person reads. The renderer fails the build on one.
• The logo position is asserted out of the finished PDF, not trusted from the code.
• The old PDF is deleted before every render, so a file that exists afterwards was written by this run and cannot be a stale one you review by mistake.

Running it

pip install pymupdf pillow
python build.py example/company.example.json

Writes out/<slug>/proposal.html, the PDF, and a proof image of every page. Needs a Chromium browser present for headless printing; it looks for Edge, then Chrome.

Then edit supplier.json so the proposal comes from you rather than from a placeholder, and install the watcher:

powershell -ExecutionPolicy Bypass -File Install-Task.ps1

Every 30 minutes. The script stands down at weekends, outside 07:30 to 17:30, on Fridays after 16:30, and if a PAUSED.flag file exists.

What it refuses to do

• It will not send an email. Drafts only, stamped for the next working morning.
• It will not guess a company name it cannot read. That goes to NEEDS-A-HUMAN.md.
• It will not build for an industry with no playbook. Same file.
• It will not treat an internal colleague forwarding a rate card as a sales lead.
• It will not build more than five in one run, so a backlog cannot flood the drafts folder.

Layout

build.py            the design, in code
style.css           the look, fixed once
render-pdf.py       render, prove the PDF is current, fail on a dash, proof every page
supplier.json       who the proposal is FROM. Edit this first
sectors/            one industry playbook per trade. This is the asset
  _TEMPLATE.json    copy this and fill it in
example/            a sample company file
Watch-Quotes.ps1    the mailbox watcher, Outlook on Windows
Install-Task.ps1    the scheduled task

Real industry playbooks and all client data live outside this repository.

Status

Last pushed 15 September 2026. This is the engine as it runs in production. The industry playbooks that drive it improve continuously and are not public, so treat the sectors/_TEMPLATE.json here as the shape, not the state of the art.

Security posture and known limitations are in SECURITY.md.

Honest limits

• The watcher is Outlook on Windows only. The builder is cross platform.
• Company names are read out of forwarded mail headers with a regular expression, which fails often enough that anything unreadable is handed to a human rather than guessed.
• The staleness check on a render samples ten passages, so a single edited sentence can slip past it. The real protection is that the old PDF is deleted before every run.
• There is no company level research. A playbook is industry level by design, because that is what makes it free.

Author

Luther Roberts, Gaborone, Botswana · github.com/94luther

Built for Sprint Couriers. Written with AI assistance, directed, specified and corrected by me over several rounds. The industry playbooks, which are the part that makes it work, are mine and are not in this repository.

Licence

MIT, Copyright 2026 Luther Roberts. See LICENSE.