Routine into a skill, mistakes into a rule
Рутину — в скилл, ошибку — в правило. Как я сделал QA-конвейер из скиллов
B · Useful, limited Detailed first-person practitioner case from a QA engineer at Banki.ru with an open-source skill set and concrete rules, but no measured outcomes such as time saved or defect rates.
Summary based on the full text·Original: about 9 minutes to read
In short
A Habr post (7 October 2026) by Konstantin, a QA engineer at the Russian financial marketplace Banki.ru, describing how he turned his daily testing routine into a pipeline of Claude Code skills. Each repeated request became a skill, and each agent mistake became a written rule inside that skill, added only after a human review. He tried a multi-agent setup and went back to a step-by-step pipeline where he checks every stage himself. The skills are published on GitHub.
Key numbers
- 10Claude Code skills built over about two months, covering the full QA workflow
- 19rules accumulated in the task-testing skill, each tied to a past agent error
As reported by the source.
Russian·Web page·opens habr.com
What it reports
- 10 skills built in about two months with Anthropic's skill-creator: acceptance criteria, Bamboo deployment, WireMock stubs, task testing, OpenSearch logs, bug reports, Allure test cases, daily test report and others.
- Stack: Jira, Confluence, Bitbucket, Bamboo, Kibana, OpenSearch, Jaeger, Allure TestOps.
- The task-testing skill holds 19 rules; most carry a short "lesson" recording the error that produced it.
- Guardrails written into the skills: test environments only, production logs only on request, no deploy to prod or stable (environment names checked against a deny list), data changes only by agreed plan and rolled back, nothing posted to Jira or Allure without the engineer's "yes".
- The testing agent only reads until the engineer types "go"; every check must carry evidence (API response, database query or trace), otherwise it is labelled a hypothesis.
- Rules that carry over to other stacks: a negative check counts only next to a passing positive one; claiming "this does not exist" needs more evidence than "this exists"; a bug is filed with its scale; expected results come from the spec, never from observed behaviour.
- Skills are shared with the team through a Claude Code plugin marketplace split by role (QA, analysts, developers, common); omitting the version field makes every commit reach the team.
How they got it
A single-author case study built from the author's own work over roughly two months, illustrated with a fictional task (refund on cancelling a paid order) walked through all eight pipeline stages. No before-and-after measurements are given.
Limits
- No numbers on time saved, defects found or false positives.
- The multi-agent "autopilot" in the public repo is marked experimental and not fully tested by the author.
- Tooling-specific rules apply only to a matching stack.
Why it matters for companies
The post shows a workable pattern for capturing tacit know-how: every recurring instruction becomes a reusable skill, and every agent error becomes a rule with its origin written down. The human approval points (plan before action, "yes" before publishing) are concrete and easy to copy into any function where agents touch live systems. Distribution through a versioned team repository means one person's correction reaches everyone. The author rejects self-written agent rules on the grounds that they can lock in wrong conclusions.
Key passages in translation
Russian original
Цель я поставил такую: перестать объяснять агенту одно и то же, каждое такое действие завернуть в скилл, а каждую ошибку агента — в правило внутри скилла.
I set myself this goal: stop explaining the same thing to the agent, wrap each such action in a skill, and turn each agent mistake into a rule inside the skill.
Russian original
Скилл не учится сам, правила в него дописываю я.
The skill does not learn by itself; I am the one who adds the rules.
Russian original
Так быстрее, но агент, который сам себе пишет правила, может закрепить и неверный вывод.
That is faster, but an agent that writes its own rules can also lock in a wrong conclusion.
Russian original
Наблюдаемое поведение не баг, пока не доказано расхождение с требованием. Порядок: факт → требование → расхождение → BUG.
Observed behaviour is not a bug until a mismatch with the requirement is proven. Order: fact, requirement, mismatch, BUG.
Russian original
Скиллы забирают рутину... Решать, что проверять и баг ли это, остаётся мне.
The skills take over the routine... Deciding what to check and whether something is a bug stays with me.
Summary based on the full text. Read the original for the full argument and data.Read the original on habr.com