An assistant that drafts, tracks — and waits for my yes
One assistant, two workflows: my job search and my painting sales. It organises and drafts — and never sends, submits or publishes without my approval.
- Type
- Personal project
- Status
- Concept in development
- My role
- Everything — research, IA, flows, agent rules, UI
- Built so far
- Nothing yet — all features are proposed
Concept screens are illustrative. All names, numbers and data shown are placeholders.
The problem
Two parts of my life run on scattered notes, spreadsheets and inboxes: applying for jobs, and selling my paintings.
Both involve the same work — organising information, matching it to what I have, writing replies, and following up on time. Both also involve messages and listings that carry my name, so I can’t hand them to an AI that acts on its own.
Concept screens
MVP scope
Job applications
- Organise opportunities and job descriptions
- Compare requirements with my verified experience
- Recommend relevant portfolio projects
- Draft tailored application materials
- Track applications, interviews and follow-ups
- Ask my approval before submitting or sending
Painting sales
- Organise artwork images and details
- Track prices, availability, inquiries and commissions
- Draft listings and customer replies
- Track orders, payments, packaging and delivery
- Ask my approval before publishing or sending
Information architecture
Both workflows share one home and one approval queue. Each keeps its own records, and both draw on a single source of truth about me: verified experience, portfolio projects and artwork details.
User flows
Job application
- 01Save job post
- 02Agent compares to verified experience
- 03Suggests projects + drafts
- 04I approve, edit or reject
- 05Submitted & tracked
- 06Follow-up reminder
Painting inquiry
- 01Inquiry arrives
- 02Agent matches painting record
- 03Drafts reply (flags missing price)
- 04I approve, edit or reject
- 05Reply sent & logged
- 06Order, packaging, delivery tasks
Agent behaviour
Can do on its own
Read and organise job posts and inquiries. Compare, summarise, draft and set reminders.
Must ask first
Submit an application, send any message, publish or change a listing, confirm a price.
Never does
Invent experience, prices or availability. Take payment. Delete records.
Every draft shows its sources — which requirement matched which project, or which inventory record a price came from — so I can check it in seconds.
Approval steps
- 01
Gather
A job post or inquiry arrives.
- 02
Draft
The agent prepares a response and cites its sources.
- 03
Approve
I approve, edit or reject in the queue.
- 04
Act & track
Only then is it sent or published, logged, and a follow-up set.
Missing or uncertain information
Missing facts
Flags the gap and asks. It never fills in a price, size or skill by itself.
Low confidence
Says why — which requirement didn’t match, or which detail it couldn’t find.
Conflicts
If a painting is marked sold but has an open inquiry, it pauses and asks.
Built vs proposed
Designed so far
- Problem framing and MVP scope
- Information architecture
- Core flows and approval rules
- Concept screens for four key views
Proposed, not built
- Every agent capability described here
- Integrations with email, job boards or marketplaces
- Payment and delivery tracking
Planned validation
Use it for my own search and sales over a few weeks, logging every draft I edit or reject. Those edits will show where the agent’s judgement falls short — and where the approval steps feel like friction rather than safety.
- How often do I approve a draft without edits?
- Which kinds of missing information come up most?
- Does the approval queue get cleared, or does it pile up?
Learnings so far
- Designing the “never does” list first made every other decision easier.
- One approval queue for both workflows is simpler than two — the question is always the same: is this OK to send?
- Showing sources is the cheapest way to make an AI draft trustworthy.
Next case study
BizClik brand websites