Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

A 1C developer distilled a month of working with AI into 7 steps for a safe start on a database copy

A 1C developer who has worked with the platform since 2006 spent a month trying AI on real tasks every day and boiled his start down to 7 steps. Instead of telling the model to "rework this data processor," he first makes a copy of the database with an export of the source code. He then asks the model only to explain the mechanism and to name the file and method behind each claim. Edits on the copy come only after that, and the result is checked by hand and by a second model. The practical takeaway: a safe process matters more than the choice of model, and the human is responsible for the code.

AI-processed from Habr AI; edited by Hamidun News
A 1C developer distilled a month of working with AI into 7 steps for a safe start on a database copy
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The main model is Claude Opus 5.5; Gemini critiques the solution via MCP
Key metric #2The production database is not connected to AI even read-only, because of the risk of load and locks
The first four taskstwo read-only, one solution without writing, one edit

Many people start using AI in development with the phrase "rework this data processor" and get a result they cannot make sense of: it is unclear why the code is written the way it is, or whether something broke along the way. In 1C the risk is higher, because live customer data and the production database sit right next to the code. A 1C developer who has worked with the platform since 2006 applied AI to real tasks every day for a month and described a 7-step route that starts with a database copy, not with choosing a model.

In practice it looks like this: instead of "fix the error in the document," you give the model an export of the source code from the copy and ask, "explain how this mechanism works, and name the file and method for each claim." If the places it names exist, you can trust the tool a little more. If they don't, you lost five minutes, not a production configuration.

This is worth reading for anyone planning to connect AI to closed corporate code: the approach is easy to carry over beyond 1C.

Seven steps: from reading to editing

The author advises treating AI as a fast but unverified contractor. It reads code well, finds connections and prepares options, but the developer is still responsible for the result. The order is as follows:

  1. Prepare a separate copy of the database and an export of the source code.
  2. Decide which data must not be sent to an external service.
  3. Choose a tool that can work with files and external utilities.
  4. Connect the 1C context read-only.
  5. Give the model a task to analyze an existing mechanism.
  6. Move on to edits on the copy.
  7. Verify the result yourself and through a second opinion.

The author orders the first four tasks by increasing risk. First, explain a mechanism. Then find the cause of an error, separating facts from hypotheses. Next, prepare a query or procedure without writing any data. Only after that make an edit in an extension on the copy, having first made a backup and a Git commit. In one case, the cause of the "bug" turned out not to be in a module: after the database was restored, the available cash register equipment had disappeared.

Tools: an agent, MCP and a second model

The author chose not a chat but an agentic tool that reads files, calls external utilities and carries a task through to a verifiable result. At the time of publication that is Claude Desktop, and the author stresses that this is a personal choice, not a ranking of the market. For EDT he advises looking at 1С:Напарник first; for VS Code or Cursor, starting with those; and if the code cannot leave the company, using only solutions for a closed perimeter.

The main model is Claude Opus 5.5, released on 22 September 2026. The author uses it to analyze other people's code, find the causes of errors and make multi-step edits; for small tasks a lighter model may be enough, and limits are worth monitoring. Gemini (currently gemini-3.1-pro-preview) is connected as a reviewer through a small MCP server; MCP is an open protocol for connecting external data and tools. Claude passes over the question and a draft solution, and Gemini sends back a critique.

The second opinion makes mistakes too. When fixing an error in a list form, Gemini suggested switching to a query schema and nested queries. The author rejected this: the standard code depended on the exact text of the query, and nested queries violated the project's rules. The author's rule: a disputed claim is confirmed by code, documentation or a test. Agreement between two models proves nothing.

Context and security

You cannot simply hand AI a .cf or .dt file: it needs text files. The procedure is as follows: unpack the .dt into a separate copy, open it in the Configurator and export the configuration to files, then export the extensions separately. The export can be automated with the /DumpConfigToFiles parameter.

The author did not find a universal MCP server for 1C. An index over the export searches quickly but goes stale: in one case it did not show the cause of an error in an extension, while a direct file comparison found an extra space in the text of the controlled method. The platform's help reference reduces the number of invented methods, but its version may not match yours. BSL checking catches technical errors but does not verify business logic. So the index is for searching, and any conclusion is confirmed against the source.

Minimum security rules:

  • do not send personal data, commercial terms, passwords, connection strings or API keys;
  • remember that the second opinion goes to yet another provider;
  • give tools rights only to the folder they need, and read-only at first;
  • read every action confirmation;
  • do not connect the production database at all, even read-only: a bad query can load the server, affect locks or expose extra data.

If live data is still needed, connect only a copy with anonymized data, a separate user with minimal rights, query limits and logging.

The limitations are worth keeping in mind: this is the experience of one developer on his own tasks, and product names and pricing plans change quickly. But the main skill, in the author's words, is defining the task and verifying the result, and the decision to adopt remains with the human.

Story

Key Questions Answered

What are the key details on Seven steps: from reading to editing?

The author advises treating AI as a fast but unverified contractor. It reads code well, finds connections and prepares options, but the developer is still responsible for the result. The order is as follows: Prepare a separate copy of the database and an export of the source code. Decide which data must not be sent to an external service. Choose a tool that can work with files and external utilities. Connect the 1C context read-only. Give the model a task to analyze an existing mechanism.

What are the key details on Tools: an agent, MCP and a second model?

The author chose not a chat but an agentic tool that reads files, calls external utilities and carries a task through to a verifiable result. At the time of publication that is Claude Desktop, and the author stresses that this is a personal choice, not a ranking of the market. For EDT he advises looking at 1С:Напарник first; for VS Code or Cursor, starting with those; and if the code cannot leave the company, using only solutions for a closed perimeter.

What are the key details on Context and security?

You cannot simply hand AI a .cf or .dt file: it needs text files. The procedure is as follows: unpack the .dt into a separate copy, open it in the Configurator and export the configuration to files, then export the extensions separately. The export can be automated with the /DumpConfigToFiles parameter. The author did not find a universal MCP server for 1C.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…