AI systems in daily use.

The company’s display advertising sat in a library of 69 ads made by an outside agency in 2025. None had been updated or tested against each other since, and every new message meant a separate design job for each ad size, so the ads rarely changed. I built a system that turns one approved message into a complete ad set: the message is written once, and the system produces it in all 9 sizes AdRoll recommends, checks every file against AdRoll’s format requirements automatically, and uploads the set into audience groups matched to the pages each visitor viewed. Ad claims are drawn only from statements the company has already published, and a person approves the copy before anything is generated.

1. Write the message
2. →
3. Generate all sizes
4. →
5. Format checks
6. →
7. Approval
8. →
9. Upload and test

• One message, all 9 AdRoll sizes, including three native formats
• A/B tests run messages side by side, and weaker ads are replaced from the same templates every 6 to 8 weeks
• Clicks and costs are read from AdRoll, and leads are traced in HubSpot through tagged links, so results are measured as enquiries rather than clicks alone
• The first set is 5 designs produced as 45 files across two audience groups, with no change to the advertising budget
• Design rules, decisions, source files and the generator are kept together in one project, so the next change follows the same steps as the last

The company had no video team and no studio budget, yet its new systems needed training material so people could learn them without one-to-one sessions. I built a video production system on Claude that turns a brief and source documents into a finished, on-brand video with a matching PDF guide, and a Video Studio page where every video is filed by number and series.

The problem

Training, explainer and promotional video would otherwise mean an outside studio for every piece, so most of it never got made.

What I built

The system writes the script and a scene plan for approval, records the software on screen where a walkthrough is needed, generates AI footage and images where the story needs them, and assembles the video in the house format: branded open and close, music and word-timed sync between voice and picture. It then builds the matching guide and files both in the Video Studio. A person records the voiceover from the approved script, and the lessons from each build go into a checklist so the next video needs no rework.

The result

25 videos in the first 12 days, including two complete training series for the company’s registers, service explainers and short promotional pieces. The whole system is packaged as a shareable Claude plugin, so a colleague can make a video by describing it and supplying the documents.

1. Brief and documents
2. →
3. Script approved
4. →
5. Capture and AI footage
6. →
7. Assembly and guide
8. →
9. Filed in the studio

Specialist agents anyone in the company can open and ask. Each has one job and draws on the shared knowledge base, so its answers use the company’s real services and approved facts.

• Bid response assistant
• Marketing expert
• Google Ads expert
• Cybersecurity advisor
• Sales sequence builder
• 3D image prompt generator

Finds new public-sector opportunities, including bids the standard reports miss, and delivers them as one short digest every weekday morning.

1. 18 sources screened
2. →
3. New bids matched
4. →
5. Weekday digest
6. →
7. Bid decision makers

• Public procurement portals
• Municipal and regional buyers
• Provincial bid listings

A complete bid management app for every pursuit, current and historical, built so leaders and subject experts can see where each bid stands and check the history before a go/no-go call.

The problem

Decades of bid history sat in dozens of spreadsheets, entered by hand. Many outcomes were missing or recorded wrong, so the history was hard to rely on.

What I built

A single register where every pursuit carries a file code and one of 13 defined states, each with a plain-language definition. A state change takes one click and is written back to the master tracker, with a visible flag until the write-back is done. Historical records can be flagged and corrected, a no-bid stays open until someone records the reason, and every change is versioned. In October 2026 I rebuilt it on the Shipley bid lifecycle, with six stages from Identify to Closed, a named bid lead on every record and a win/loss reason with a debrief field. A six-criteria scorecard scores each opportunity on fit, relationship, competition, capacity, strategic value and risk, and returns a Bid, Discuss or Pass recommendation.

The result

1,800 records across 684 buyer organisations now sit in one traceable register. Every correction improves the data behind the next go/no-go decision. A dashboard shows win rate by year, top buyers and contracts coming up for renewal, and a recompete radar flags contracts before they expire.

1. Flag a record
2. →
3. Correct the state
4. →
5. Written back to the tracker
6. →
7. Better go/no-go data

MYRA’s visual identity, brand rules and past work, held in one central system in Claude Design, so every new piece of design work starts on brand instead of from a blank page.

The problem

MYRA’s brand material was spread across more than 16,000 files in many formats, mixed with drafts and archives, much of it cloud-only and readable only in the software that created it.

What I built

I converted that material into a structured design system in one working session: 547 files across four packages covering logos, fonts, colour rules, finished work in 12 categories and a vector library converted from the original artwork. The logo was rebuilt as layered vector files in five approved colourways, with named layers for the arc, each letter and the tagline, so it can be animated piece by piece. Two off-palette blues found in older files were ruled out in favour of the approved palette.

The result

From the system, Claude Design generated design tokens, 24 foundation cards, 19 reusable components, 7 slide types and 2 UI kits, and its automated design system check reported no issues. Social posts, presentations, brochures, event material, print, advertising, email, interactive pages and animation can now start from one brand foundation.

1. Brand files
2. →
3. Four packages
4. →
5. Tokens and components
6. →
7. On-brand work

Claude projects I built for my own work and use every day. Each one knows its task, the house style and the company’s facts, so a first draft takes minutes rather than hours.

• Connected to the shared knowledge base with defined permissions
• Recurring procedures saved as reusable skills
• Every draft is reviewed and edited by a person

A sales sheet starts as a short brief. My one-pager project drafts it in house style from approved company facts in the shared knowledge base, and a person checks every claim before it goes to a seller. The mock-up below follows one sheet from brief to approval.

• Drafts from the shared knowledge base, not from memory
• One angle per sheet, from the same framework I use for social posts
• Every sheet is reviewed and approved by a person

I run the company’s Google Ads and AdRoll accounts in-house as one programme, with AI working the accounts under a written playbook that names me as owner. The AI can act on keywords, negative keywords, ad copy and bids. Any increase in total spend, any new campaign type and any change to forms, landing pages or site tags waits for my approval.

1. Weekly checks
2. →
3. AI acts inside the playbook
4. →
5. Spend changes held
6. →
7. A person approves

• Weekly checks on spend pacing, search terms, negative keywords, tracking and disapprovals
• Runs in place of an agency retainer the company was quoted and deferred
• A staged plan for leadership that restructures the account by service and ties any budget increase to verified leads (a proposal, awaiting a decision)