Grigoriy Dobryakov – IT+AI Blog
Grigoriy Dobryakov's blog: management, development and testing
Hype says “hire AI.” Job descriptions buy something else: production agents with evals and guardrails, Applied AI leadership, partner architecture, evaluation harness as a must-have. Method: read role clusters as an org chart under pressure.
Hype says “hire AI.” Real job descriptions buy people who own production agents, evals, Applied AI, and partner delivery. Below is a method for reading vacancy clusters as an org chart under pressure, plus five role clusters from a parsed corpus of open roles.
Continue reading “AI Job Demand in 2026: What Hiring Signals Show in Job Postings”
Demo-grade eval greenlights a feature on the same examples as the pitch. Production then shows the long tail. Three layers — incident-born regression, distribution diff vs the last snapshot, human spot-check — plus one named owner. Minimal harness: github.com/dobryakov/eval-harness.
Treat AI-output quality as a release criterion: a fixed regression set built from past incidents, a distribution diff against the previous release snapshot (20–50 real inputs), a human spot-check before the first production deploy of a new output type — and one person who signs off. Below: the demo-grade anti-pattern, the three-layer method, and a minimal harness that fails CI with a readable exit code.
Without a domain substrate, an agent ships FTP and CSV and calls it done. Book-as-context puts a book into the project as a wiki — next to LLM Wiki and book-to-skill — so a short prompt yields outbox, queues, idempotency, and fault tolerance.
Give the agent an authoritative book on the subject as a wiki in the repo — indexed and cross-linked, not a one-off file drop into the chat. Then even a short prompt designs in the source’s terms: queues, delivery guarantees, idempotency, fault tolerance. Below: the book-as-context method, a shop→ERP walkthrough, and the method’s limits.
Three skills are fine. Thirty — style drifts, shared state gets overwritten, behavior changes with no stack and no log line. Dropping markdown into `.claude/skills/` is a hire with no interview. Onboarding exists so you catch the conflict before it hits daily flow — not because process looks nice.
Because copying markdown into a folder is not installing a package. It is a hire: a new slice of behavior enters the assistant and will collide with its neighbors tomorrow. A skill here is an instruction pack for an AI assistant (Claude Code / Cursor): frontmatter, rules, declared reads/writes/calls. While the intake path is empty, every new file under .claude/skills/ is a silent hire with no interview and no record in the file.
Continue reading “Why AI skills need onboarding if you can just copy them into the folder”
“Don’t delete prod” in a system prompt is not a security control. The real boundary is where the model physically cannot act: API surface, isolation, operation filter, audit. Working reference: enterprise-code-bastion.
A prompt is a wish. A perimeter is when the model physically cannot reach files or a shell: native tools off, an MCP intermediary, Docker per project+task, a filter before execution, an audit trail. Below is a working reference — and where the design gives way.
Continue reading “How to secure an AI coding agent without relying on the system prompt”
