Vibe-:     GPT  LLM
  


  ,    ,  ,     ,    ?       Vibe-: ,  GPT/LLM    , ,   .        ,  workflow Prompt ? Code ? Review,   IDE (, Cursor)  - ,     ,    ,  PR  CI/CD.    :   ,    ,     .      ,    ,         .







 

Vibe-:     GPT  LLM





 1.  Vibe-


1.1 Ҹ   pipeline:    

  () pipeline   :  ?  ?  ?  ? .          ,  ,    .



Ҹpipeline Vibe-  :    GPT-     ,    .       ,     .







Ҹ



 

 

 35   LLM-



 

  

,    





  

  -



.   :    LTV.        SQL-,      ;   GPT-4o  15    ,     .

 .,  LLM-   . :  prompt   -.         ( hallucination).

 .    draft-&-review:

 -   .

  .

   .

,      .

1.2 GPT    

  LLM    :      ,     ,     . GPT   ,    ,   .

 McKinsey (2024) : ,  LLM-   ,   idea ? prototype   37 %.       22 ..     .

.      ,      :   ,       .      ,    .

 .  GPT  - (  ), ,         .   廠   BI-,  KPI.     .

 .

  prompt-:  ?  ?   ?   .

   23 ,         .

1.3      

-     ,   ,  . Vibe-  :   -.

 Stack Overflow Labs (Q1 2025) : ,   60 % pull-request-  GPT- (Cursor, Copilot Workspace),   OKR   45 %,        27 %.

  ?

      ( Google Code Health, 2023).

     .

  :   .

  2.  - (   LLM),     :       API.    -: 400 $  GPT-4o       3000 $.

 .     (  40 %   !).    -  DORA-,     .

   .

KPI Time-to-First-Feedback Story Points .

 (, NPS)  API- LLM.

   :            .

 

Vibe-       ,  GPT-    .        ,    -. Ҹ pipeline,             ,      .






 2.  GPT- 


(   ,   -     )

:    

   JetBrains AI Tools Survey (IV . 2024) 78 %      LLM-    ;    31 %       .   :     ,     Notion,      .    -   ,   .

      ,    ,     ,     .

2.1  :, , 







 

 





  , , SQL-

Prompt- You are Architect

 ,  





 ,  ,  

Pull Request comment ? diff-patch

-,  





      

Chat  IDE / CLI-assistance

,  legacy-



 .-   :

 (Claude 4 Sonnet)  45      FastAPI   .

 (GPT-4o)        .env.

 (Gemini 2.5 Flash)             Confluence.

 ࠖ    .   :       ,   :   .  ,    .

.   ,     :      .

2.2 IDE / CLI- :     

Cursor Agent-Rewrite:  57 , ?K ?I,   ,  diff.

Copilot Workspace  PR:  ,      ;    .

JetBrains AI Assistant    -   Llama-3-Instruct Q4 (  dev-  ).

CLI-(bash+ollamallama-cpp)      .

.   Tinkoff Tech ( 2025) ,   diff- (Cursor  WS)    MR  42 %.

 

   as-is -    . :  human-  autogen-.

 IDE  :     ,   .

 .     ;      ,      .

2.3   : Git  LLM

/prompts .  prompt   Markdown  YAML-:

title: risk_scoring_generator role: generator model: claude-4-sonnet version: 1.3.0 last_test: 2025-04-12

Semantic Versioning:MAJOR  ;MINOR  ;PATCH  .

Unit-  prompts.  Prompt Layer   :    ? ,      JSON.

CI- Prompt ? Code:  PR,  ,    - .

 .

v1.0   SQL-    LTV.

v1.1   cutoff_date.

v2.0    Pydantic-,   JSON-schema.

 ࠖ       Git.     ,    :   prompt    Slack.

.     /prompts,       :       .

 .

KPI Prompt Coverage:  -,    (  ? 85 %).

   Prompt Grooming  34        .

 

 GPT-     :  ,    IDE/CLI    .

,    , :

    (MR-  40  50 %),

    ,

  ,     -  prompt-     .

  ,    ,           .






 3.  


 prompt       :   ,   .  Stanford HAI Prompt Engineering Report 2024       32 %,      ,    .

3.1 SVO-:  ?  ? 

S (subject).   :  Python-.



V (verb).   : .



O (object/ ).  : Django 5.0,   ORM,     .py.







  SVO



S

 

  senior-backend



V

 

 stack-trace   



O

 

Django >= 5.0,  PEP 8, max 80 



 . -:

  <ROLE>. : <BUSINESS GOAL>.

: <ACTION VERB>.

: <OBJECT CONSTRAINTS>.

 : <FORMAT>.

.



   BI. :  LTV. :  SQL under Postgres 15. :   180 ,   user_id. : code-block SQL +   .

 Microsoft Prompt Patterns 2025  SVO-     1,6    .

3.2   ? 

LLM-     .        :

Hard-Spec (  ). bash-,   uptime    exit- 1  < 99 %.

Semi-Creative.        10 k RPS.

Open-Creative.     NFT-.

  .

(be concise vs brainstorm 10 ideas).

  / n-.

 (  ,  5     ).

.  prompt    ,    :         .

3.3    prompt:  

 . REST-,       .











1

 -  

MAU > 10 k, SLA ? 100 



2

 

  senior-backend Go



3

 

 Go 1.22, Gin,   TimescaleDB



4

 

 full- handler.go + docker-snippet



5

  

 go vetgo test ./ 



 prompt

  senior-backend Go. : REST- /forecast

    (MAU>10k, SLA<=100).

 : Go 1.22, Gin, TimescaleDB.

:   handler.go  docker-compose.yml.

   go vet  go test ./  .

:  code-,  Go,  YML.

 OpenAI Function Calling Benchmark 2025         70 %.





3.4  :   

        ,  prompt    :,     -  󅻠     ,       .

 

  ,   .

   (  Redis,     ).

  > 4k  .

 

   SVO,   .

  /      .

    code-block,  黠    .

 Alibaba DAMO 2024: 60 % -,  LLM-,   RCA     .





 - Vibe-prompter

 SVO:  ?  ? .

  :exact,semi,open.

  : code-block, JSON, .

 : ,   .

: unit- prompt    .

   ,                 ,      pipeline,       .






 4.  :   


     :      ,   .  Time-to-Prototype Index, ,  2025.





4.1 Claude 4 Sonnet   Architect

 -       :  ,  ,  ,  Docker-compose.  70 %     .    ArchitectClaude 4 Sonnet.

 .

  senior- JS. : SPA-  React 18

 SSR  Next 14, CI  GitHub Actions,   Postgres 15.

 baseline-: package.json, tsconfig,

Prettier, E2E- CodeceptJS. : zip- + README.

    (Sonnet  zip,    URL,  24 ).

  GitHub   auto-pipeline,  Sonnet     .

. Skyeng Tech ( 2025):       14,6   ,  Architect-prompt  -.

.     prompt,  zip: Sonnet  ,   ,   .  boilerplate- :     -.





4.2 -SPA- :      

LLM-   ,    .      -,   ,   (Single Point of Attention, SPA).



Checkpoint

 

 





SPA-0Bootstrap

  

,  build  

npm ci && npm run build



SPA-1Style

  

Code style, , dead-code

ESLint + Prettier CI step



SPA-2Security

 merge main

SCA, secret-scan

Trivy, GitLeaks



SPA-3Value

 demo-

 Time-to-First-Feedback

Jira + custom API



 Avito Engineering Metrics Q1-2025, ,  SPA-0  SPA-1 ,          ( 12  6 ).

 SPA-0:

# .github/workflows/bootstrap.yml name: SPA-0 Bootstrap on: push: branches: [init/** ] jobs: build: runs-on: ubuntu-latest steps: uses: actions/checkout@v4  name: Use Node 20 uses: actions/setup-node@v4 with: {node-version: '20' }  run: npm ci  run: npm run lint && npm run build

LLM   YAML,   .   ,    ,  Claude  GPT-4o         .

 . SPA :   pipeline,     .        :  incremental gatekeeping.





4.3    ,      

    ,        .    :  ,  ,  environment-.

 .

  README.LLM-     .

  .  ,  -,  shared-.

 .   ,  o  .

  LLM Onboarding Audit,  2025:        11 %   - .

:   

Auto-MAP.md.  prompt       :

/src

/core  -

/ui   

/infra    , 

 /docs/glossary.md: 12   .

-.  PR-     /infra?     .           .

.   ,    ,     . ,  ?   .





  

 Claude 4 Sonnet Architect,   SPA-0 pipeline     .

 . init/yy-mm-projectname ,   baseline-.

   MAP.md   , go.mod.        .

  Time-to-First-Feedback:  < 2       .







    Vibe-     Generate,   :LLM-   , SPA-   ,         .    ,                 .









5.Workflow Prompt ? Code ? Review

      LLM:       ,    .





5.1 Chain-of-Thought:   

:     ,      .



 GPT      (chain-of-thought)  ,      :      .

  .     X         .    Google AI ( 2025)            28 %.

 :

- 1:

,          API ,    .

 : ,   .

- 2:

       Go,   :handler,service,repository.

.,      ,  .         ,        CI   5  12  MR (   Tinkoff Tech,  2025).





5.2 Self-review :  -

:           .

 :

 .

 :

 self-review:  ,  ,  .

  (diff-   unified      git-).

 堖  IDE  Cursor  Copilot Workspace.

:

    .

  XSS/SQL-,    .

   (  Style Guide).

 .  self-review     .      ,     CI.  :

    git-,     .





5.3 : XSS-   

:     ,      XSS.

 Chain-of-Thought:

,   ,   HTML-  .



:

 whitelist-  .

 sanitize().

  .

   .



 :

  Node.js sanitize.js  DOMPurify.

const createDOMPurify = require('dompurify'); const { JSDOM } = require('jsdom'); const window = (new JSDOM('')).window; const DOMPurify = createDOMPurify(window); function sanitize(input) { return DOMPurify.sanitize(input, { ALLOWED_TAGS: ['b','i','em','strong','a'], ALLOWED_ATTR: ['href'] }); }module.exports = { sanitize };

Self-review :

 self-review  ,        -.

  :

 sanitize.js +++ sanitize.js @@ -1,3 +1,5 @@ -const createDOMPurify = require('dompurify'); +const createDOMPurify = require('dompurify'); +//      +if (typeof input !== 'string') throw new TypeError('Expected input to be a string'); const { JSDOM } = require('jsdom');

  :

 unit- sanitize() Jest,  <script> <b>.

const { sanitize } = require('./sanitize'); test('escapes script tags', () => {expect(sanitize('<script>alert(1)</script>')).not.toContain('<script>'); }); test('keeps b tags', () => {expect(sanitize('<b>bold</b>')).toContain('<b>bold</b>'); });

:       , MR        3 -    .





  

    .Chain-of-Thought =  .

  .Git-    CI.

     . sanitize,       .

  Prompt ? Code ? Review,   GPT-   -,     .   ,    ,          2025 .






 6. TDD + Vibe-:  ,   


  TDD   ,    Vibe-        ,       .





6.1  Claude 4,   Gemini Flash

  Vibe-TDD??      .

 -.

  QA-. :   .

:  unit-  Jest   login(), logout(), isAuthenticated().

: mock   API,     .

: code-block  .

  .



 (Claude 4)  auth.test.js   .

   .

  senior-backend  Node.js. :   login(), logout(), isAuthenticated() ,     auth.test.js .

: Express 5, JWT,  API auth-service.

:  code-:    package.json.

Gemini Flash     .

:

 :     ,   ?? .

  -:   ,    .

 :  PR    -.





6.2   ?  prompt

  ,   ,     prompt    .

 .



 CI Jenkins  GitHub Actions    failed tests.

 prompt.

   ,  login()  undefined  JWT.

  login() ,     { token: <string> }     Error('Invalid credentials')   .

 .



    ,    .

.   ,           . ,      ,    prompt??      .

 :

?? .    -,    -,        assert. :     ??    -      .





6.3    

    Vibe-TDD??   ,  -.              -  ,  -   .

  :

//  ,  isAuthenticated()   "OK" expect(isAuthenticated('token')).toEqual('OK');

    true,  ,        .

 :

 .    ,     ,    .

expect(typeof isAuthenticated('token')).toBe('boolean');

 JSON-  .Prompt  :

 ,     JSON-schema:

type: "object",

properties: {

token: {type: "string", minLength: 20 }

required: ["token"]

  Prompt-Grooming??   -     API.





  

   -(/tests/templates).   unit-   .

     prompt (must??error??success),       .

 -  pre-commit hook.  ,      -,   CI.

        IDE (Cursor  , Copilot  )??   .







 TDD  Vibe-,    prompt ?  ?  ? ,       . Claude 4   , Gemini Flash  ,       CI/CD.       15 %,   MR  30 %      -.






 7.  Legacy-


 Vibe-         . Legacy-        :       ,       .     ,          : Reverse-prompt, GPT-      code-smells.





7.1 Reverse-prompt:     

 ࠖ             ,   .

  .



 ,   : ,      .

,         .



 Reverse-prompt.

    Node.js.     processOrder(orderData). :

1)     ?

2)      ?

3)     ?

:      .

 .



  GPT   .

 : ,     feature-,    .

. -         . Reverse-prompt ,  validate() crypto.createHash('md5'),       SHA-256.

 .          .    :   ,   ,   ,        .

 .     Reverse-prompt     /docs/legacy-analysis.           .





7.2 GPT    

     ,   -. GPT     ,  -,      :

 .

role: documentator model: gpt-4o prompt: |  senior- TypeScript.       JSDoc-:    .   .   .

2. **  Cursor  Copilot Workspace.**

   `orderProcessor.ts`.

   JSDoc.

   .



3.**  .**

,      (,   ).

 ,  GPT   .



> **.**     ,     .     ,  ,   ,    .



** .**    .  GPT    .          JSDoc,      .



** .**   JSDoc-  ESLint- `eslint-plugin-jsdoc`.  PR    red-build,      .



### 7.3     code-smells

     -:           .



1.**  .**

  - 15- ,   .    ,   code-smells:

 

 

 



:        .

2. **     200300 .**

 legacy-,    .

     -  Jira  GitHub Issues.



3.**.**

 :  ,  , race-condition.

:  ,  .

: cosmetic-smells   , -.



> **.**      GPT    `bufferSize = 8192`.               `maxBufferSize`,   OutOfMemoryError   .



** .**   code-smells    :  .        .



** .**   code-smells    :        High- .          .



##  

 Legacy-        ,    :

1. **Reverse-prompt**     .

2. **GPT-**  ,     .

3. ** **      backlog .



  ,      ,          ,     -.






 8.   


   Vibe-     ,      : , , .





8.1 Explain ? Refactor ? Validate:  

 1. Explain.      .

  senior-  JavaScript. ,    processData(data)   ,   ,      .

   :

   forEach( 1227).

   .

temp      .

 2. Refactor.     :

       refactor- processData.js:   `forEach`  `forof`        formatDate()   try/catch     : unified diff-.

 3. Validate.,    :

 self-review    git diff,    ESLint-   .

.   18 ,           -.

:       3040 %   .

 :  diff-  Cursor  Copilot,    as-is.

 . refactor      Explain:        .

 . - Explain?Refactor?Validate  /prompts/refactor/   semantic versioning.





8.2 :  jQuery ? Vanilla JS

   jQuery:

$('.btn').on('click', function() { const $item = $(this).closest('.item'); $item.find('.value').text(parseInt($item.find('.value').text()) + 1); });

Prompt:

  -.      ES6-JavaScript  jQuery,  event delegation   API DOM. :    diff-    eventHandler.js.

 diff:

 before.js +++ eventHandler.js @@ -1,4 +1,12 @@ -$('.btn').on('click', function() {  const $item = $(this).closest('.item'); $item.find('.value').text(parseInt($item.find('.value').text()) + 1); -}); +document.addEventListener('click', event => { + if (!event.target.matches('.btn')) return; +const item = event.target.closest('.item'); +const valueEl = item.querySelector('.value'); +const current = Number(valueEl.textContent); +valueEl.textContent = current + 1; +}); + +export default eventHandler;

.

   jQuery,    15 .

 npm run build .

   4      .

 .  - : event.target.closest.

 . prompt  :  closest-  IE11,     .





8.3  1 % diff  99 % 

 .  (1 % )       ,        99 %    . ,      ,    ,         50 %.

.  LLM  2     formatDate() Intl.DateTimeFormat,

 -   8  1  ,

   PR   4   1,8 .

 .

   -: 510   sprint.

  Jira  refactor: extract helper, refactor: improve naming   .

   DORA-: lead time  change failure rate.







Explain ? Refactor ? Validate    :   ,       .    jQuery ? Vanilla JS ,      diff,    .   1 % diff  99 %  :          .      ,  Vibe-       .






 9.  & auto-README


README      ,         .

        :       , ,  ,    .  README     ,    .       ,       . Vibe-        :   Readme  LLM,             .





9.1 TL;DR + Installation + Contribution templates

  README     :

TL;DR   ,      23 .

Installation       ( ,  ,  ).

Contribution     : , , , PR-,  .

 (Markdown):

#   ## TL;DR :           ClickHouse. ## Installation ```bash git clone https:///repo.git cd repo docker-compose up -d npm run migrate npm start

Usage

 http://localhost:3000

 GET/metrics?from=&to=

Contribution

 

 feature/-

   (npm test)

 Pull Request  :



:    

 :   

Checklist:   lint, , 



**-  GPT:**

```text

  technical writer.  README   service-metrics:

TL;DR (< 50 )

Installation    Docker   

   

Contribution  - PR

  Markdown.

   ,      /README.md    CI.

.  ,   : ,  README   ,     ,   .





9.2 : README = 

  README

       .

     .

     .

      2 ,   ,   ,      .

 .        ,  npm installENV-.

 .

   .  Installation         ,    .

 Usage   HTTP-, CLI-  .        5 .

Contribution  ,    :  PR-,      -    CI.

 . CI-:   Pull Request , README.md ## Installation## Contribution.        ,  PR   .





 

 README   . PR    ,   .

   pre-commit hook. ,   LLM   README   API.

 badges.  ,  ,  NPM,        .

  Doc-Grooming.   ,    ,    .





 

README       ,       . Vibe-         LLM-:   README,      CI   Doc-Grooming.   ,       ,         .






 10. Code-review 2025:  + LLM


  Vibe- -   ,       ,        .  GitHub Code Review Survey Q1 2025, ,      LLM,     PR-  38 %       24 %    .                .





10.1 - 

            .      ,   - :

  



   Style Guide (ESLint, Prettier).

  ,  ,  .







SQL-  XSS-   LLM-prompt.

        .







    .

 resource-leaks ( ,  ).



 



  unit-      .

     ( ? 80 %).







   API  README.

    JSDoc/Swagger-.



 



  (God Object,  ).

    .

 :     PR     100 .      ,     .

 :      .    MR-    1520 :  ,  ,  .





10.2   

     ,    .  Microsoft Developer Velocity Index 2024, PR-         22 % ,  PR,     .

:     .

-:   ( ,   ).

-:   (  ,  ).

:           ,       .

 :    system prompt  You are a code reviewer   .    ,    .

 :   /prompts/review_tone/  :

tone_friendly.md   ;

tone_neutral.md  ;

tone_strict.md   .

      ,     .





10.3 :     

   LLM,         . JetBrains AI Tools Survey IV . 2024: 12 % ,       ,     .

  -:

 :       -.

 :    ,    .

 :    PR-    .

 :     Copilot Agents,   merge on successful tests and zero critical alerts.        .

 : 堫-໠  ,             .    ,       .





 

  code-review  + LLM     .  -  ,     ,          -.   ,         ,    2025 .









11.



IDE: Cursor, Copilot Workspace, JetBrains AI

 2025           ,       .  IDE         ,         Pull Request.       :Cursor,Copilot WorkspaceJetBrains AI Assistant,   ,     .





11.1  AI- (, )

Cursor



 :     ,    .

:  diff   ,  /  .

:  100 000        .

:  $30/   ,    $50/   on-premise.



GitHub Copilot Workspace



- PR: -       PR   .

: Unit-, doc-, CI-     .

:    GitHub,   .

: $20/  ,    $40    .



JetBrains AI Assistant



 :   Llama-3-Instruct Q4     -.

 :   ,     Style Guide.

UI-: inline-,       quick-fix.

:   All Products Pack (?$249/),   .

:     AI-,     ,    ,  .        .





11.2 Cursor: Agent-, multi-edit, smart-rewrites  off-line-

Agent- Cursor     ,    :

Generator     prompt-.

Reviewer diff-   .

Navigator    Confluence-.

Multi-edit            .     :  package-   .

Smart-rewrites   :      ,  Cursor   map/filter  for-loop.

Off-line-       :

   Cursor      (?200 ).

   Cache-   ( <50 ).

  lifelike-:  on-premise ?     .

 :



       Mocha  Jest:

test/**/*.spec.js.

 AgentConvert Mocha to Jest.

  diff-  describe ? test,before ? beforeEach  .

    ,   90 %    Cursor   .





11.3 :  Cursor ? PR auto-draft

 



 feature/add-payment-gateway.

   handler.js  Cursor-:

  senior-backend  Node.js.   Stripe API:

 endpoint POST /payment

  

  Winston



  self-review



Agent-Generator   handler.js.

Agent-Reviewer    :     ,  git-.




  .


   .

   ,     (https://www.litres.ru/pages/biblio_book/?art=73395873)  .

      Visa, MasterCard, Maestro,    ,   ,     ,  PayPal, WebMoney, ., QIWI ,       .


