The problem
Two questions are often left unanswered when teams evaluate a new AI capability: “What can it do in work like mine?” and “How much time or money does it actually save?” Product marketing can demonstrate possibility, but a decision requires the work itself and a credible measurement trail.
Cowork Recipes closes that gap with a public, end-to-end Turkish proof gallery. Every claim starts with a real Cowork run, and every cost starts with measured credit usage. Anything that cannot be measured is explicitly labelled as an estimate.
From a scenario to publishable evidence
Each recipe follows the same production pipeline:
- Define the role, industry, pain point, delegation brief, and expected output.
- Generate clearly labelled synthetic Turkish data instead of using customer data.
- Run the task in a Turkish demo tenant with human supervision.
- Capture every stage, from setup and planning to approval and output, in 16:9 screenshots.
- Download the actual artifact and measure the run with the
/costcommand. - Convert credits to USD, daily Central Bank of Türkiye rates, loaded human cost, and comparable ROI.
- Produce the article and social drafts, then run a critic loop until every quality gate passes.
More than an ROI headline
A case page exposes the input files, Cowork's plan, approval checkpoints, real downloadable artifacts, and the honest assessment together. The measured Cowork cost remains fixed while visitors can change their own time and hourly cost in a client-side ROI calculator.
The result reframes “what could AI do?” as “what was completed, with which data, through which steps, and at what measured cost?”
Measurement architecture
Measured Cowork credits × $0.01
→ daily USD/TRY exchange rate
→ measured Cowork cost
Human hours × loaded hourly cost
→ human baseline
→ comparable ROI
TRY and ROI values are not frozen inside articles. Each build recalculates them from measured credits, the current exchange rate, and transparent assumptions, preserving the evidence chain as currencies move.
The system
| Layer | Implementation |
|---|---|
| Experience | Static, client-light, fully Turkish Astro site |
| Evidence | Scenario definition, synthetic data, screenshots, artifacts, and editorial narrative |
| Measurement | TypeScript cost engine and daily exchange-rate automation |
| Generation | Azure AI Foundry text/image generation with a critic loop |
| Automation | Human-supervised Playwright + Edge tr-TR runner |
| Delivery | Azure Static Web Apps, Functions, Storage, and Bicep IaC |
The monorepo separates the cost engine, content generator, browser runner, site, and API. A dedicated Turkish copywriting skill and CI-enforced lint:tr gate keep machine-translated language out of the published experience.
Scope and results
Sixteen published studies cover marketing attribution, CX analysis, supply shocks, legal contract review, finance, HR, IT, and strategy. Every study includes a real downloadable artifact. Measured ROI examples range from 132x to 319x.
Real customer data, automated social posting, and regulatory compliance claims are intentionally out of scope. This is an independent personal project, not an official Microsoft property.
My role
I owned the product concept, information architecture, measurement model, content and automation pipeline, Astro frontend, Azure infrastructure, Turkish quality gates, and production release end to end.