← Back to portal

Notes since 2023

How I actually use AI

I'm a QA engineer. Since 2023 I've dragged models into real work: first as search that beat another Google tab, then as an IDE partner, then as a pipeline step. Numbers below are a snapshot, plus what I used and what I still owe myself. Early chats got cleaned, so 2023–2025 is partly reconstructed from practice, not telemetry.

Snapshot 9 September 2026Author: Denis Sviridov

in progressNext: DoR and an artifact between stages

01 · Numbers

Four numbers on the snapshot date. Not live telemetry.

  • Cursor sessions1594Cursor agent chats across projects on this machine
  • Other agents1102Sessions outside Cursor: OpenCode, Hermes, Pi, and the rest
  • Portal commits309Committed work in PersonalBlog
  • Statusin progressHow this snapshot is going

02 · In short

Why this page exists and where the numbers come from.

In short

2023: ChatGPT in the browser so I closed fewer Google/Yandex tabs. Late 2024: DeepSeek plus paste into the IDE, and the first thing other people saw was a Telegram Mini App Advent calendar for the company QA community. 2025: agents, MCP, RAG, a year to map the zoo instead of marrying one tool. 2025/2026 boundary: a company coverage dashboard in Cursor (early skills and a custom MCP). 2026: several agent runtimes next to Cursor, a live portal, test automation, and my own AI projects. Session counts are a lower bound. Some of the history lived only in web chats and then got wiped.

03 · What already holds

What works in production, and what still needs me in the chair.

Strengths

  • A pipeline, not a chatI split work across agents and MCP instead of one endless thread.
  • Agree firstSpec or plan before code. Otherwise the model sprawls.
  • Tool for the jobCursor for long features. CLI/TUI when that's faster. Not a collection.
  • Ships to prodCommit, deploy, live site. A draft in chat does not count.
  • Don't trust the modelReview, tests, security/SEO. The QA habit did not retire.

Growth areas

  • Talk about itCases and talks so someone else can steal the practice, not just stare at KPIs.
  • Short playbooksOne-pager «how to repeat» for a team, not for a blog.
  • English pageThe same notes in en, without translation fog.
  • Live demoRun the pipeline in front of people, not a slide with a screenshot.

04 · How it changed

Since 2023: from a browser chat to several runtimes and a real pipeline. About skill, not website features.

  1. 2023

    Chat instead of the tenth tab

    ChatGPT in the browser. Not «build a product». Queries, checks, fewer Google/Yandex tabs.

  2. late 2024

    First artifact other people saw

    DeepSeek in the browser, paste into the IDE. What reached people: a Telegram Mini App Advent calendar for the company QA community.

  3. 2025

    Map the zoo

    Claude Code, Cursor, VS Code/PyCharm extensions (Kilo Code, Continue, Firebender). MCP in parallel (Playwright was early) and RAG. A year to understand the set, not pick one tool forever.

  4. late 2025 – early 2026

    Skills and a coverage dashboard

    In Cursor with the first skills I built a company coverage dashboard: TestOps, DocHub, GitLab, SonarQube, per-service cards, admin UI, and an MCP so an agent could pull stats.

  5. spring–summer 2026

    Several runtimes on the desk

    OpenCode, Hermes, Pi, Cline and others sat next to Cursor. Since mid-July OpenCode, Hermes, and Pi have been idle. I pick the cockpit for the job: IDE, TUI, desktop, CLI.

  6. 2026-07

    Agent as a pipeline step

    spec → plan → verify, MCP into work systems, Codegraph, review skills, the portal as a proven ship. The agent is not magic. A step with a Definition of Done.

05 · Where the work happened

Top Cursor projects by session count, plus key milestones.

  • Hoff tests (automation)

    141 sessions

    Hoff retail suite: scenarios, fixtures, failure triage. I use the model to write and fix tests and move faster through a large repo.

  • PersonalBlog (personal portal)

    125 sessions

    Blog, admin, client portal, CRM. Agents from feature to deploy. Also the lab for skills, MCP, and agent rules.

  • Hoff app test (mobile)

    124 sessions

    Mobile app automation: structure, fixtures, flakes. Lots of «agent + real test repo» sessions.

  • QA Agent (my own agent)

    103 sessions

    Configurable QA agent: CLI/UI, LLM providers, MCP/tools, flows for cases and checklists. Building one, not only using someone else's.

  • Coverage dashboard (company)

    milestone

    Coverage dashboard: TestOps + DocHub + GitLab + SonarQube, service cards, admin UI, MCP. Cursor and early skills at the 2025/2026 boundary. The bridge from experiments to in-company work.

06 · Tools

Two layers: where the conversation ran, and where the agent fetched data.

Where the agent ran

IDE, TUI, CLI, extensions. Model plus tools.

  • Cursor
  • OpenCode
  • Hermes
  • Pi
  • Claude Code
  • Codex
  • Cline
  • Kilo Code
  • Continue
  • Firebender

MCP into work systems

Not only local files: Jira, TestOps, GitLab, and the rest.

  • Playwright MCP
  • Jira
  • Confluence
  • TestOps
  • Grafana
  • GitLab
  • Codegraph
  • AgentMemory

Skills and rules

Reusable instructions: review, tests, writing.

  • Skills (SKILL.md)
  • AGENTS.md rules
  • Security review
  • SEO audit
  • Playwright checks
  • RAG experiments

How work actually ships

So AI work does not dissolve into another chat.

  • Spec → plan → ship
  • TDD / verify before done
  • Multi-agent review
  • Docker deploy

By tool

How many sessions where, and why the tool is on the list.

  • Cursor1594

    Main IDE. Code, refactors, long features. Almost all of the heavy work lives here.

  • Codex96

    OpenAI Codex CLI. A coding lane next to Cursor, not a replacement.

  • OpenCode172

    TUI/desktop: portal work, MCP, audits. Idle since mid-July.

  • Hermes43

    Desktop runtime: digests, experiments, QA-related flows. Idle since mid-July.

  • Kilo Code24

    VS Code / Windsurf agent. Focused edits when I don't want the whole project in Cursor.

  • Pi23

    CLI: worktrees and fast terminal ops. Idle since mid-July.

  • Cline21

    Extension with task sessions. One task, one pass.

  • Claude Code19

    Anthropic terminal agent. Separate repo sessions.

  • Factory14

    Another runtime for comparing UX and pipelines. Not the daily driver.

  • Crush1

    A light trial. Barely used.

07 · What's next

Steps so another engineer can repeat the practice.

  1. Two or three cases «problem → agents → outcome» with time and quality numbers, not chat counts.
  2. Definition of Ready at stage entry. Do not start while the input is mush.
  3. An artifact between stages: spec, plan, patch, verify. Not «more sessions».
  4. Track lead time, defects, coverage. Session count by itself means nothing.
  5. Walk one engineer through spec → plan → ship with an artifact at each step.
  6. Refresh this snapshot quarterly with current tool numbers.

08 · Snapshot takeaway

Where the practice sits now, and the next step.

What this is, stripped of slogans

From 2023 the arc is short on paper and long in habits: chat as search, a Mini App, an agent/MCP/RAG lab, a company dashboard, then several runtimes and a live portal. Lower bound on saved sessions in other agents is about 1100, plus 1594 in Cursor. The next step is not «more models». Definition of Ready and an artifact between stages, or it is just chat again.

DoR + an artifact between stages

Session counts come from local logs on the snapshot date, not live telemetry. 2023–2025 is partly reconstructed by hand: I cleaned chats, and some work lived in web UIs with no agent transcript. Tools disagree on what a «session» is.