Enforce document voice consistency

LEANDERANTONY · LEANDERANTONY · commit e149d8d33a8b · 2026-03-18T14:11:32.000+05:30
diff --git a/src/agents/cover_letter_agent.py b/src/agents/cover_letter_agent.py
@@ -1,3 +1,5 @@
+import re
+
 from src.config import get_openai_max_completion_tokens_for_task
 from src.prompts import build_cover_letter_agent_prompt
 from src.schemas import (
@@ -15,6 +17,10 @@
 from .common import coerce_string, coerce_string_list, unique_strings
 
 
+_THIRD_PERSON_SELF_REFERENCE_RE = re.compile(r"\b(he|his|him|she|her)\b", re.IGNORECASE)
+_CANDIDATE_LABEL_RE = re.compile(r"\b(the candidate|this candidate)\b", re.IGNORECASE)
+
+
 class CoverLetterAgent:
     def __init__(self, openai_service=None):
         self._openai_service = openai_service
@@ -49,14 +55,25 @@ def run(
                 task_name="cover_letter",
                 metadata=prompt.get("metadata"),
             )
-            return CoverLetterAgentOutput(
+            output = CoverLetterAgentOutput(
                 greeting=coerce_string(payload.get("greeting"), default="Dear Hiring Team"),
                 opening_paragraph=coerce_string(payload.get("opening_paragraph")),
                 body_paragraphs=coerce_string_list(payload.get("body_paragraphs"), limit=3),
                 closing_paragraph=coerce_string(payload.get("closing_paragraph")),
                 signoff=coerce_string(payload.get("signoff"), default="Sincerely"),
                 signature_name=coerce_string(payload.get("signature_name"), default=candidate_profile.full_name or "Candidate"),
             )
+            if self._contains_third_person_self_reference(candidate_profile, output):
+                return self._fallback(
+                    candidate_profile,
+                    job_description,
+                    fit_analysis,
+                    tailored_draft,
+                    tailoring_output,
+                    strategy_output,
+                    resume_generation_output,
+                )
+            return output
         return self._fallback(
             candidate_profile,
             job_description,
@@ -67,6 +84,26 @@ def run(
             resume_generation_output,
         )
 
+    @staticmethod
+    def _contains_third_person_self_reference(
+        candidate_profile: CandidateProfile,
+        cover_letter_output: CoverLetterAgentOutput,
+    ) -> bool:
+        text_blocks = [
+            cover_letter_output.opening_paragraph,
+            *cover_letter_output.body_paragraphs,
+            cover_letter_output.closing_paragraph,
+        ]
+        combined_text = " ".join(str(block or "").strip() for block in text_blocks if str(block or "").strip())
+        if not combined_text:
+            return False
+        candidate_name = str(candidate_profile.full_name or "").strip()
+        if candidate_name and candidate_name.lower() in combined_text.lower():
+            return True
+        if _CANDIDATE_LABEL_RE.search(combined_text):
+            return True
+        return _THIRD_PERSON_SELF_REFERENCE_RE.search(combined_text) is not None
+
     @staticmethod
     def _fallback(
         candidate_profile: CandidateProfile,
diff --git a/src/agents/resume_generation_agent.py b/src/agents/resume_generation_agent.py
@@ -1,3 +1,5 @@
+import re
+
 from src.config import get_openai_max_completion_tokens_for_task
 from src.prompts import build_resume_generation_agent_prompt
 from src.schemas import (
@@ -14,6 +16,12 @@
 from .common import coerce_string, coerce_string_list, unique_strings
 
 
+_RESUME_SELF_REFERENCE_RE = re.compile(
+    r"\b(i|me|my|mine|myself|he|his|him|himself|she|her|hers|herself|the candidate|this candidate)\b",
+    re.IGNORECASE,
+)
+
+
 class ResumeGenerationAgent:
     def __init__(self, openai_service=None):
         self._openai_service = openai_service
@@ -48,15 +56,32 @@ def run(
                 task_name="resume_generation",
                 metadata=prompt.get("metadata"),
             )
-            return ResumeGenerationAgentOutput(
+            output = ResumeGenerationAgentOutput(
                 professional_summary=coerce_string(payload.get("professional_summary")),
                 highlighted_skills=coerce_string_list(payload.get("highlighted_skills"), limit=8),
                 experience_bullets=coerce_string_list(payload.get("experience_bullets"), limit=6),
                 section_order=coerce_string_list(payload.get("section_order"), limit=6),
                 template_hint=coerce_string(payload.get("template_hint"), default="classic_ats"),
             )
+            if self._contains_self_reference(candidate_profile, output):
+                return self._fallback(fit_analysis, tailored_draft, tailoring_output)
+            return output
         return self._fallback(fit_analysis, tailored_draft, tailoring_output)
 
+    @staticmethod
+    def _contains_self_reference(
+        candidate_profile: CandidateProfile,
+        resume_output: ResumeGenerationAgentOutput,
+    ) -> bool:
+        text_blocks = [resume_output.professional_summary, *resume_output.experience_bullets]
+        combined_text = " ".join(str(block or "").strip() for block in text_blocks if str(block or "").strip())
+        if not combined_text:
+            return False
+        candidate_name = str(candidate_profile.full_name or "").strip()
+        if candidate_name and candidate_name.lower() in combined_text.lower():
+            return True
+        return _RESUME_SELF_REFERENCE_RE.search(combined_text) is not None
+
     @staticmethod
     def _fallback(
         fit_analysis: FitAnalysis,
diff --git a/src/prompts.py b/src/prompts.py
@@ -347,6 +347,7 @@ def build_resume_generation_agent_prompt(
         "system": (
             "You are the Resume Generation Agent. Produce the final tailored resume content from grounded upstream analysis. "
             "You may rewrite, reorder, and emphasize, but you must not invent employers, achievements, dates, metrics, or unsupported skills. "
+            "Write in standard resume style: no first-person or third-person pronouns, no full-name self-reference inside the summary or bullets, and no cover-letter phrasing. "
             "Keep the output ATS-safe and recruiter-readable. "
             + _build_contract(contract)
         ),
@@ -391,6 +392,8 @@ def build_cover_letter_agent_prompt(
     return {
         "system": (
             "You are the Cover Letter Agent. Write a recruiter-facing cover letter only after the review stage has approved or corrected the upstream outputs. "
+            "Write entirely in first person from the candidate's perspective. "
+            "Do not describe the candidate as he, she, him, his, her, or by full name anywhere in the letter body; reserve the candidate name for the signature line only. "
             "Use the approved tailoring, strategy, review, and resume-generation context as the source of truth. "
             "Do not invent employers, metrics, projects, technologies, or direct experience that are not supported by the provided inputs. "
             "Keep the result specific to the role, grounded, and ready for packaging into the final artifact. "
diff --git a/tests/test_cover_letter_agent.py b/tests/test_cover_letter_agent.py
@@ -0,0 +1,71 @@
+from src.agents.cover_letter_agent import CoverLetterAgent
+from src.schemas import CandidateProfile, FitAnalysis, JobDescription, JobRequirements, TailoredResumeDraft, TailoringAgentOutput, WorkExperience
+
+
+class FakeThirdPersonOpenAIService:
+    model = "fake-model"
+
+    @staticmethod
+    def is_available():
+        return True
+
+    @staticmethod
+    def run_json_prompt(system_prompt, user_prompt, expected_keys=None, **kwargs):
+        return {
+            "greeting": "Dear Hiring Team",
+            "opening_paragraph": "I am excited to apply for the Data Scientist role. Leander Antony is a project-based machine-learning candidate with hands-on Python experience.",
+            "body_paragraphs": [
+                "His portfolio work includes predictive modeling and validation projects.",
+            ],
+            "closing_paragraph": "He would welcome the opportunity to discuss the role further.",
+            "signoff": "Sincerely",
+            "signature_name": "Leander Antony",
+        }
+
+
+def test_cover_letter_agent_falls_back_when_ai_uses_third_person_self_reference():
+    agent = CoverLetterAgent(openai_service=FakeThirdPersonOpenAIService())
+    candidate_profile = CandidateProfile(
+        full_name="Leander Antony",
+        experience=[
+            WorkExperience(
+                title="AI Engineer",
+                organization="Example Labs",
+            )
+        ],
+    )
+    job_description = JobDescription(
+        title="Data Scientist",
+        raw_text="",
+        cleaned_text="",
+        requirements=JobRequirements(hard_skills=["Python", "SQL"]),
+    )
+    fit_analysis = FitAnalysis(
+        target_role="Data Scientist",
+        overall_score=82,
+        readiness_label="Strong",
+        matched_hard_skills=["Python", "SQL"],
+        experience_signal="I have built predictive models and validation workflows in portfolio projects.",
+    )
+    tailored_draft = TailoredResumeDraft(
+        target_role="Data Scientist",
+        professional_summary="Project-based machine learning candidate with grounded Python experience.",
+        highlighted_skills=["Python", "SQL"],
+    )
+    tailoring_output = TailoringAgentOutput(
+        professional_summary="Project-based machine learning candidate with grounded Python experience.",
+        cover_letter_themes=["Highlight predictive modeling and validation work."],
+    )
+
+    result = agent.run(
+        candidate_profile,
+        job_description,
+        fit_analysis,
+        tailored_draft,
+        tailoring_output,
+    )
+
+    assert "Leander Antony is" not in result.opening_paragraph
+    assert "His portfolio work" not in " ".join(result.body_paragraphs)
+    assert result.opening_paragraph.startswith("I am excited to apply for the Data Scientist role.")
+    assert result.closing_paragraph.startswith("I would welcome the opportunity to discuss")
diff --git a/tests/test_prompts.py b/tests/test_prompts.py
@@ -1,7 +1,9 @@
 from src.prompts import (
     build_assistant_prompt,
     build_application_qa_assistant_prompt,
+    build_cover_letter_agent_prompt,
     build_fit_agent_prompt,
+    build_resume_generation_agent_prompt,
     build_review_agent_prompt,
     build_strategy_agent_prompt,
 )
@@ -94,4 +96,30 @@ def test_review_prompt_allows_null_corrections_when_no_rewrite_is_needed():
     assert "null when no tailoring changes are needed" in prompt["system"]
     assert "null when no strategy changes are needed" in prompt["system"]
     assert "unresolved_issues" in prompt["system"]
-    assert "Approve when the final corrected wording stays grounded" in prompt["system"]
+    assert "Approve when the final corrected wording stays grounded" in prompt["system"]
+
+
+def test_cover_letter_prompt_requires_first_person_voice():
+    prompt = build_cover_letter_agent_prompt(
+        candidate_profile={"full_name": "Leander Antony"},
+        job_description={"title": "Data Scientist"},
+        fit_analysis={"experience_signal": "Grounded ML project experience."},
+        tailored_draft={"professional_summary": "Project-based ML candidate."},
+        tailoring_output={"professional_summary": "Grounded summary."},
+    )
+
+    assert "Write entirely in first person from the candidate's perspective" in prompt["system"]
+    assert "Do not describe the candidate as he, she, him, his, her, or by full name" in prompt["system"]
+
+
+def test_resume_generation_prompt_requires_pronoun_free_resume_style():
+    prompt = build_resume_generation_agent_prompt(
+        candidate_profile={"full_name": "Leander Antony"},
+        job_description={"title": "Data Scientist"},
+        fit_analysis={"matched_hard_skills": ["Python"]},
+        tailored_draft={"professional_summary": "Grounded draft summary."},
+        tailoring_output={"professional_summary": "Grounded summary."},
+    )
+
+    assert "no first-person or third-person pronouns" in prompt["system"]
+    assert "no full-name self-reference inside the summary or bullets" in prompt["system"]
diff --git a/tests/test_resume_generation_agent.py b/tests/test_resume_generation_agent.py
@@ -0,0 +1,58 @@
+from src.agents.resume_generation_agent import ResumeGenerationAgent
+from src.schemas import CandidateProfile, FitAnalysis, ResumeGenerationAgentOutput, TailoredResumeDraft, TailoringAgentOutput
+
+
+class FakePronounResumeOpenAIService:
+    model = "fake-model"
+
+    @staticmethod
+    def is_available():
+        return True
+
+    @staticmethod
+    def run_json_prompt(system_prompt, user_prompt, expected_keys=None, **kwargs):
+        return {
+            "professional_summary": "I am a project-based machine learning candidate with strong predictive modeling experience.",
+            "highlighted_skills": ["Python", "SQL", "XGBoost"],
+            "experience_bullets": [
+                "I built predictive models for fraud detection.",
+                "Leander Antony developed validation workflows for ML projects.",
+            ],
+            "section_order": ["Professional Summary", "Core Skills", "Professional Experience", "Education"],
+            "template_hint": "classic_ats",
+        }
+
+
+def test_resume_generation_agent_falls_back_when_ai_uses_self_referential_resume_voice():
+    agent = ResumeGenerationAgent(openai_service=FakePronounResumeOpenAIService())
+    candidate_profile = CandidateProfile(full_name="Leander Antony")
+    fit_analysis = FitAnalysis(
+        target_role="Data Scientist",
+        overall_score=84,
+        readiness_label="Strong",
+        matched_hard_skills=["Python", "SQL", "XGBoost"],
+    )
+    tailored_draft = TailoredResumeDraft(
+        target_role="Data Scientist",
+        professional_summary="Candidate profile aligned to Data Scientist with grounded evidence around Python, SQL, XGBoost.",
+        highlighted_skills=["Python", "SQL", "XGBoost"],
+        priority_bullets=["Built predictive modeling workflows for fraud detection use cases."],
+    )
+    tailoring_output = TailoringAgentOutput(
+        professional_summary="Candidate profile aligned to Data Scientist with grounded evidence around Python, SQL, XGBoost.",
+        rewritten_bullets=["Built predictive modeling workflows for fraud detection use cases."],
+        highlighted_skills=["Python", "SQL", "XGBoost"],
+    )
+
+    result = agent.run(
+        candidate_profile,
+        job_description={"title": "Data Scientist"},
+        fit_analysis=fit_analysis,
+        tailored_draft=tailored_draft,
+        tailoring_output=tailoring_output,
+    )
+
+    assert isinstance(result, ResumeGenerationAgentOutput)
+    assert result.professional_summary == tailoring_output.professional_summary
+    assert result.experience_bullets == ["Built predictive modeling workflows for fraud detection use cases."]
+    assert all("I " not in bullet for bullet in result.experience_bullets)
