Layer 1
Trigger
Time, event, or on-demand. Same primitive for cron and webhook.
Four builds we run on ourselves. Two live, two dated.
We run our own playbook on ourselves first. Two of these are in production today. Two have a go-live month and are not live yet, and they are named here in the same breath as the ones that shipped. Every live date is on the Still Running board. Each build is a receipt for how we work: AI handles judgement, code handles repetition, a human stays the editor where the stakes are highest.
A note on these
These are Core Nova-internal builds.
Two are in production. Two are built and dated but have not reached their go-live month, and they are marked that way on every mention. They are described here to show how we work, not as benchmarks for client engagements. Client builds carry different scope, different data boundaries, and different governance, all named in the SOW. The cost figures are our own operating costs, not client-outcome claims. Where we have built equivalent workflows for clients, we describe those qualitatively only and only with consent.
Build 01 · Skill
The branded-PDF skill.
Live since August 2026
The problem
Asking a chat model to generate a branded PDF directly burns credits, drifts logos and fonts, and breaks tables. Same prompt, different output every run. The model is doing two jobs: thinking about the content and pretending to be a typesetting engine. It is good at the first and bad at the second.
The fix
The model writes markdown, which is its native language and consistent every run. Pandoc plus the Core Nova stylesheet renders the markdown into a branded PDF. Same result every run. The skill is one file that documents the template, the output path, and the rendering call. Available in every project, on every machine, called by name.
Where it shows up
The same skill renders our Snapshot briefs, our Scaffold specs, our Retainer monthly reports, our internal weekly plans, our capability decks, and the proposals built by Build 02 below. Build it once. Six other workflows inherit it for free.
The pattern
AI for words. Code for layout.
AI is good at reading and writing. Code is good at fetching, scheduling, and formatting. Combine the two and the work that used to eat your evenings runs while you sleep.
Concept covered: Skill, a saved recipe the AI can call by name.
Build 02 · MCP + Skill
One-line brief to branded proposal in five minutes.
Not yet live · October 2026
An MCP, short for Model Context Protocol, is the standard plug that lets a model read from your business systems with the buyer's permission. USB-C for AI: one plug, many tools. With MCPs the assistant goes from "a well-read teammate in another building" to "a teammate who can actually open the cabinet". Same model, different value.
The chain, end to end
- MCP to the CRM. Look up the client. Past engagements, named contacts, current pipeline state.
- MCP to the file system. Read the matching service spec. Scope, deliverables, terms.
- Deterministic code. Apply pricing rules, workflow-band logic, GST treatment.
- AI reasoning. Draft the narrative tailored to the client and the conversation history.
- The branded-PDF skill from Build 01. Render the markdown into the branded proposal PDF.
The whole chain is one prompt. The operator stays the editor: read the draft, tweak, send. Around five minutes including the read-through, versus most of an hour writing one cold.
Equivalent shapes
The pattern, in other businesses
| Business type | Skill chain |
|---|---|
| Consultancy | Capability statement to tailored proposal |
| Trades | Site notes to branded quote |
| Professional services | Engagement brief to letter of engagement |
| Agencies | Creative brief to scope of work |
| Coaches | Discovery notes to programme proposal |
The pattern is the same. The names change.
Build 03 · Agent
Inbox triage at 6:30am. Day plan at 7:00am.
Not yet live · September 2026
Two scheduled agents, built and due to start running every weekday before coffee from September 2026. The inbox agent reads overnight mail, classifies into "needs you", "can wait", and "drafted reply ready", and lands a one-page summary in the inbox by 6:30am. Drafted replies are staged in the Drafts folder, not sent. The operator is the one who hits send.
The day-plan agent runs thirty minutes later. It compares the calendar with standing weekly priorities, flags conflicts before the day starts (childcare pickup against an evening meeting, a deep-work block colliding with a call), and produces the day's shape with focus blocks already named. Zero decisions to make at 7am. The synthesis is already on the page.
What it costs to run
Measured in testing, not in production: the inbox agent classifies and drafts for around twenty-five emails per morning at roughly five cents per run, and both agents together come in under five Australian dollars per month, all-in. The figure gets confirmed against a full month once the build goes live. They were built across one weekend. Once the plumbing is right, the same shape composes for daily standups, weekly summaries, board prep, or any other "synthesise this for me before the day starts" job.
Sample output
Your morning brief 4 need you 18 can wait 2 drafts staged Needs you today 1. Inbound RFP, decide bid/no-bid 2. Coffee with $partner, pick a time 3. Doodle poll due Friday 4. Industry event tomorrow, confirm Drafted replies (review and send) -> coffee time confirmation -> event availability FYI (8 newsletters, 5 auto-notifications)
Sample. Real briefs name people and deals. Those details stay private and never appear on the website.
Build 04 · Agent at scale
A daily LinkedIn newsroom that holds the operator's voice.
Live since May 2026
Founder-led companies are told to post daily. The reality: writing kills evenings, delegating produces generic AI mush, and a dedicated content writer is two thousand dollars a month for someone who does not know the business. This build solves all three.
What it does
- Reads fourteen RSS feeds at 9am AWST (regulator briefs, security press, industry publications).
- Scores each item against positioning to find what is worth saying.
- Drafts three short posts in the operator's voice, anchored to past published writing.
- Emails the drafts to a human reviewer.
A human reviews, picks, posts. The operator stays the editor. Voice does not collapse into generic AI mush because the writing system is anchored to a corpus of the operator's own prior published work.
Design decisions that matter
- Two models, one cheap one capable. A smaller model handles the scoring pass, a larger model handles the drafting pass. Cuts cost without hurting voice.
- Voice anchors. Past operator-written posts are loaded as style references so the draft does not drift into generic AI mush.
- Prompt caching. Seventy percent input-token saving by caching the positioning brief and voice corpus.
- S3-stored config. Edit voice or feed list without redeploying code.
- Human in the loop. Posts go to a Drafts folder, not a publishing API.
In production
AU$35 / month
All-in operating cost, every cost line included.
| AI inference | ~$23 |
| Secrets storage | $0.40 |
| Compute, queue, storage | <$1 |
| Email, scheduling | <$0.10 |
Running every weekday since May 2026.
The pattern under all four
One stack. Different jobs.
The four builds use the same five layers. The AI sits in one layer. The other four are commodity plumbing. Build the plumbing right and a cheaper or better model swaps in later with no other change.
Layer 2
Connectors (MCPs)
Read from email, calendar, CRM, file system, accounting system, whatever the workflow needs.
Layer 3
AI reasoning
Classify, summarise, draft. The only layer where the model sits. Cost is measured in cents.
Layer 4
Code orchestration
Schedule, chain steps, apply rules, format output. Deterministic. Effectively free to run.
Layer 5
Delivery
Email, drafts folder, branded PDF, dashboard. Human review baked in where it matters.
Governance
Built to align
NIST AI RMF aligned and oriented to ISO/IEC 42001. Governance, traceability, and human review at layer 5 by design.
Find one workflow worth automating. Ship it in four days.
The First Workflow engagement is fixed price, four working days, A$6,000 + GST. We map your business processes, name the workflow with the largest unrecovered hours, and ship the first improvement before the second week ends. The engagement ends there unless we both agree there is more value to capture.