AI Agent in VS Code Drafts a Technical Spec in 15–20 Minutes Instead of 1.5 Days
Analysts have to rewrite a customer's requirements and meeting notes by hand into a project-specific technical specification (ЧТЗ), which takes a day to a day and a half. One team put an AI agent in VS Code and gave it skills: Markdown files with rules like “don't make things up” and “cite your source.” A draft ЧТЗ is now ready in 15–20 minutes, or in half a day including the analyst's review. The takeaway: the gain comes not from the model but from strict constraints and analyst verification.
AI-processed from Habr AI; edited by Hamidun News
An analyst receives a document with business and functional requirements (БФТ) from the customer, adds notes from a couple of meetings, and has to turn all of it into a project-specific technical specification (ЧТЗ), a document developers can work from. Done by hand, this takes a day to a day and a half: you open the template, transfer the requirements, pick the wording, and format everything. The team that writes requirements for development handed this routine to an AI agent in VS Code.
A draft ЧТЗ is now assembled in 15–20 minutes, and with the analyst's review the document is ready in half a day instead of a day and a half. These figures are the team's internal estimates, not independent measurements. Still, the team's experience shows how to configure an agent so that it helps rather than invents.
How the pipeline works
The flow is as follows: a meeting or a БФТ is turned into a summary in Markdown, the agent assembles a draft, the analyst edits it, and the document goes to Confluence or to development. There are several agents. Some run on internal servers with local models, and they can be given source code without any approvals. The others connect through corporate cloud access and receive only documents permitted by company policy.
They work in the built-in VS Code chat (GitHub Copilot Chat). Local models connect via BYOK (“bring your own key”): you specify a provider (Ollama, Azure, or any OpenAI-compatible endpoint) and a key, and the model appears in the list next to the cloud ones. The team chose VS Code because the repositories are already open there, a separate “sandbox” folder can sit alongside them, and files are attached to the chat by reference, such as #file:requirements.md, without pasting text into the prompt. The same approach works in Cursor, IntelliJ IDEA, and opencode.
There is an economic side as well. Analyzing a large БФТ together with the code of three services burns a lot of tokens, so a local model on your own hardware often turns out cheaper than a cloud one.
Skills instead of long prompts
The agent is configured through skills. A skill is an ordinary Markdown file with instructions. In VS Code it lives at .github/skills/<name>/SKILL.md, and the agent loads it on its own when the description matches the task. A skill has three parts: the structure of the output with hard constraints (“don't make things up,” “cite your source,” “highlight discrepancies”), a digest of the project context (domain, stack, roles, regulations), and links to reference materials such as a glossary and an architecture description. As a result, the request itself is two lines: the files and the task.
The agent has a minimal set of tools: reading files, keyword search, and writing only to the output/ folder. The authors warn that with write access the agent can overwrite the wrong file, so boundaries are mandatory.
Skills are kept in a shared git repository, one per type of work: ЧТЗ, Confluence article, code analysis, test cases, meeting summary, explanatory note for a technical design, function points for calculating the НМЦК (initial maximum contract price), and requirements for development. A newcomer clones the repository and gets a ready-made set. If the agent “acts up,” the team compares the skill against a reference version. They chose skills over Cursor Rules and system prompts because a skill is a portable file that can be reviewed and versioned.
The context is strictly limited: 2–4 Markdown files with documents and 1–3 code services, with a specific method named, for example, “look at the commission calculation in CommissionService.java.” When the team tried to feed the agent everything at once, it got swamped and started making things up. Hence the rule: no more than 3–4 files at a time, and only with the key terms of the task.
What worked, and where the agent lies
Three cases from practice. Analyzing a commission calculation spread across three microservices with no proper documentation took about 40 minutes (20 for the agent's work and 15 for verification, the rest for setting up the task) instead of two to three hours, and no developer had to be pulled away. A draft ЧТЗ is assembled in 15–20 minutes, plus an hour or two for review.
A Confluence article follows the same logic: a draft in 10 minutes instead of half a day. Here MCP servers are connected, and the agent creates the page and pulls in attachments on its own. For a small change a couple of edits are enough, while for a large feature with a dozen scenarios, half the text gets rewritten.
The errors are typical. When context is lacking, the agent fills in the gaps: from the phrase “we discussed changes to the status model,” it may infer statuses that do not exist in the system. It rarely admits ignorance and confidently produces plausible but wrong information. In Russian-language projects, the terms “заявка,” “обращение,” and “запрос” (all roughly “request”) get mixed up, so a glossary in the skill is mandatory. If the БФТ and the meeting notes contradict each other, the agent may pick the convenient version and not flag the discrepancy, so the “Discrepancies” block is checked against the sources by hand.
Verification has several stages: the analyst checks the draft against the sources, the logic, and the format; a document for the customer is agreed with the customer; and requirements for development are reviewed by the development team. Systematic errors are easier to fix with a single instruction (“in section 3 you mixed up ‘Rejected’ and ‘Withdrawn,’ fix it everywhere”), while minor ones are corrected by hand.
What to take away
Start with one БФТ and one meeting record in an empty folder, not with a large feature. Create a separate skill for each type of document. Restrict the agent to a single folder for writing. A skill is worth writing if the task recurs, such as a weekly meeting summary; for a one-off task, it is faster to do it yourself. Treat the agent like a capable intern: it will gather the material and highlight the unclear spots, but the analyst is responsible for the final document.
- Nov 20, 2026Anthropic, Maker of Claude, Raises $81.3 Billion in IPO at a $1.5 Trillion Valuation
- Nov 20, 2026Anthropic raises $81.3 billion in IPO, beating SpaceX's $75 billion record
- Nov 20, 2026Anthropic goes public and raises $81.3 billion — shares jump 17% on day one
- Nov 20, 2026Hugging Face and llama.cpp Release 1 MB Llama for Windows: AI Without a Terminal
Key Questions Answered
How the pipeline works?
The flow is as follows: a meeting or a БФТ is turned into a summary in Markdown, the agent assembles a draft, the analyst edits it, and the document goes to Confluence or to development. There are several agents. Some run on internal servers with local models, and they can be given source code without any approvals. The others connect through corporate cloud access and receive only documents permitted by company policy. They work in the built-in VS Code chat (GitHub Copilot Chat).
What are the key details on Skills instead of long prompts?
The agent is configured through skills. A skill is an ordinary Markdown file with instructions. In VS Code it lives at .github/skills/<name>/SKILL.md, and the agent loads it on its own when the description matches the task.
What worked, and where the agent lies?
Three cases from practice. Analyzing a commission calculation spread across three microservices with no proper documentation took about 40 minutes (20 for the agent's work and 15 for verification, the rest for setting up the task) instead of two to three hours, and no developer had to be pulled away. A draft ЧТЗ is assembled in 15–20 minutes, plus an hour or two for review. A Confluence article follows the same logic: a draft in 10 minutes instead of half a day.
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.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.