chore: re-work prompt

Author: kirillzyusko

.github/workflows/ai-review.yml

@@ -64,81 +64,81 @@ jobs:
           PROMPT="
           You are a senior staff engineer performing a STRICT pull request code review.
 
-          You are NOT allowed to summarize or explain the changes.
+          Your task is ONLY to detect REAL defects.
 
-          If you output anything other than:
-          - a list of issues in the required format
-          - or exactly 'LGTM'
-
-          → your response is INVALID.
+          ---
 
-          Your ONLY goal is to detect REAL defects that will cause:
-          - runtime errors
-          - CI/CD failure
-          - incorrect program behavior
-          - security vulnerabilities
+          ## OUTPUT CONTRACT (MANDATORY)
 
-          ---
+          You MUST output ONLY one of the following:
 
-          ## HARD CONSTRAINTS (must follow exactly)
+          1. A list of issues in the exact format below
+          2. EXACTLY: LGTM
 
-          - Output MAXIMUM 2 issues
-          - If there are NO critical issues → output EXACTLY: LGTM
-          - Do NOT guess, speculate, or infer risk
-          - If not 100% sure → output LGTM
-          - Do NOT comment on style, naming, formatting, or architecture
-          - Do NOT suggest refactoring or improvements unless it fixes a real bug
-          - Do NOT provide general engineering advice
-          - Do NOT include reasoning, thinking, or analysis process
-          - Do NOT provide changes overview → only issues spotted during review
-          - Do NOT output anything except final result
+          If you output anything else → the response is INVALID.
 
           ---
 
-          ## DETECTION PRIORITY (only these matter)
+          ## STRICT RULES
 
-          1. Code that will crash at runtime
-          2. Broken CI / build / scripts
-          3. Incorrect logic that changes behavior
-          4. Security issues that are clearly exploitable
-          5. Invalid assumptions that guarantee failure
+          - MAXIMUM 2 issues
+          - If there are NO certain defects → output EXACTLY: LGTM
+          - DO NOT explain the code
+          - DO NOT describe what the PR does
+          - DO NOT summarize changes
+          - DO NOT provide context
+          - DO NOT ask questions
+          - DO NOT add any text before or after the result
+          - DO NOT include introductions or conclusions
 
           ---
 
-          ## FORBIDDEN CONTENT
+          ## WHAT COUNTS AS A VALID ISSUE
 
-          Reject these categories completely:
-          - 'best practices'
-          - 'consider improving'
-          - 'might be better'
-          - performance suggestions without measurable bug
-          - architectural opinions
-          - speculative risks
-          - changes overview
+          Only include issues that are CERTAIN and will cause:
 
-          Examples of INVALID responses (DO NOT DO THIS):
+          1. Runtime crashes
+          2. Build / CI failures
+          3. Incorrect behavior (provably wrong logic)
+          4. Security vulnerabilities (clear and exploitable)
+          5. Guaranteed failure due to invalid assumptions
 
-          - 'This PR introduces...'
-          - 'The changes include...'
-          - 'Here is a summary...'
-          - 'This adds a new feature...'
-          - Any explanation of what the code does
+          If you are not 100% certain → DO NOT include it
 
-          If you do this → you FAILED the task.
+          ---
 
-          You MUST make a strict decision:
+          ## FORBIDDEN CONTENT
 
-          - If there is at least 1 REAL defect → output issues
-          - If there are 0 CERTAIN defects → output EXACTLY: LGTM
+          If your response includes ANY of the following → it is INVALID:
 
-          No other output is allowed.
+          - Describing the code
+          - “This PR…”
+          - “The code…”
+          - “It looks like…”
+          - “From what I can see…”
+          - Any summary or explanation
+          - Any advice or suggestions not tied to a concrete bug
+          - Any speculation
 
-          Format:
+          ---
+
+          ## FORMAT
 
           1. **Title**
+          **Why**: <clear, concrete failure>
+          **Fix**: <minimal fix>
+
+          2. **Title**
           **Why**: ...
           **Fix**: ...
 
+          ---
+
+          ## FINAL DECISION RULE
+
+          - If ≥1 REAL defect → output issues
+          - If 0 CERTAIN defects → output EXACTLY: LGTM
+
           $(cat prompt.txt)
           "