Iurie Plesu

Human-directed AI engineering

AI in the workflow.
Accountability stays with me.

I use agents for investigation, implementation and verification. I define the scope, review the changes and make the release decision. Architecture and access boundaries remain engineering responsibilities.

How I work

Define the check before the change.

I inspect the source and current system state, then define the intended result and how to verify it. Skills provide a repeatable process. I accept the work only after reviewing the diff, test results and deployed behavior.

  1. 01
    Scope

    Define the objective, constraints, ownership boundaries and acceptance criteria.

  2. 02
    Select

    Choose focused skills and provide only the context and permissions required.

  3. 03
    Execute

    Compose repository, terminal, browser and delivery tools into one controlled workflow.

  4. 04
    Verify

    Inspect the diff, test the behavior and confirm the deployed result before acceptance.

Reusable capabilities

The practices behind the workflow.

Repeatable tasks become focused capabilities with defined inputs, allowed tools and completion checks. I keep the context small enough to review and the result concrete enough to test.

01

Focused skills

I use specialized instructions and completion checks for repository analysis, infrastructure, browser QA, technical documents and visual production.

02

Grounded context

I provide relevant source files, current system state and architecture constraints, with privacy boundaries that limit what the agent can access.

03

Efficient retrieval

Targeted search, staged file loading and reusable summaries keep useful context available without repeatedly loading the entire repository.

04

Ordered tool use

Filesystem, terminal, browser and cloud tools share one workflow. Independent inspection can run in parallel; dependent or consequential changes run in order with controlled permissions.

05

Test and correction

I plan, execute, inspect the result and correct errors in short cycles. Tests, acceptance checks and explicit stop conditions determine when to continue or stop.

06

Human review

I challenge assumptions and correct scope or implementation errors. Architecture, security and consequential actions require evidence and my approval before release.

Engineering control

What stays under my control.

Context boundaryThe agent receives the files and system details required for the task, not unrestricted private context.

Decision boundaryArchitecture, credentials, destructive actions and release decisions remain explicitly reviewed.

Evidence boundaryA generated change is unfinished until the real system, browser or deployment confirms the intended behavior.

Automated delivery example

FitBack Coach

A compact application built to exercise the complete delivery path from a GitHub change to a live Cloudflare Pages release.

FitBack Coach is the deployable workload for a Git-based CI/CD flow. A push to the main branch triggers the Cloudflare Pages pipeline, runs the TypeScript and Vite production build, publishes the dist artifact and makes the release available for browser verification.

Trigger
Push to GitHub main
Build
TypeScript check and Vite bundle
Publish
Cloudflare Pages from dist
Verify
Live PWA and offline behavior

Delivery focus: The application provides a real release target for validating automated builds, hosted delivery, service-worker caching and local IndexedDB persistence.

Open live deployment