docs(prompts): expand all template instructions with detailed extraction rules and validation

[ThinkerDesigns]

Summary

Expands all prompt templates with significantly more detailed instructions, reducing LLM ambiguity. Each template grew from ~12 lines (a single sentence + schema) to 80–130+ lines of structured guidance.

What changed

Template	Before	After	Key additions
system_message.jinja	1-line directive	~75 lines	Parser persona, date normalization, title standardization, dedup rules, error handling
basics.jinja	2-line directive + URL hints	~80 lines	Per-field extraction (name/email/phone/url/summary/location/profiles), validation checks, canonical naming examples
education.jinja	1-line directive	~95 lines	6-field schema with inference rules, reverse-chronological ordering, multi-degree handling, what to skip
work.jinja	Basic work extraction + date rules	~105 lines	Employment type classification (FT/intern/contract/fellowship/founder), role segmentation for promotions, remote detection
projects.jinja	1-line directive	~100 lines	Project boundary detection, shared-technology propagation, technology mapping table, complexity signal guidance
skills.jinja	1-line directive	~105 lines	Category classification logic, canonical naming table, cross-resume crawling rules (don't limit to skills section), level consistency
awards.jinja	1-line directive	~85 lines	Award vs non-award classification, tier/rank handling, competition/scholarship detection, deduplication
github_project_selection.jinja	Detailed but unstructured	Restructured	Tighter commit-tier tiers with explicit scoring bands, project-type classification rules (open_source vs self_project), stronger reasoning requirements per selection
resume_evaluation_system_message.jinja	Dense 50-line wall of text	~130 lines	Clear sectioned layout, per-category band guidance with evidence standards, bonus/audit logic organized step-by-step

Net change: 9 files, +878 / -248 lines. Total prompt instructions grew from ~12 lines to ~670+ lines.

Why this matters

The old prompts were essentially "extract X and return JSON" with no guardrails on how — leaving everything to LLM inference, which is the source of inconsistent parsing and scoring variance. These expanded instructions give the model explicit rules for:

• Date normalization (YYYY-MM format, partial dates, season → month)
• Canonical naming (JS → JavaScript, Node.js not nodejs, etc.)
• Edge-case handling (empty fields → null vs empty string, deduplication, remote work indicators)
• What to skip (classroom demos, placeholder text, implicit soft skills)

Files changed

prompts/templates/awards.jinja
prompts/templates/basics.jinja
prompts/templates/education.jinja
prompts/templates/github_project_selection.jinja
prompts/templates/projects.jinja
prompts/templates/resume_evaluation_system_message.jinja
prompts/templates/skills.jinja
prompts/templates/system_message.jinja
prompts/templates/work.jinja

How to test

Run a couple of real resumes through the pipeline:

python score.py ./resume/sample.pdf --role software_engineering_intern

Check that section extraction JSON matches expected output and scoring is consistent across multiple runs.

[Copilot]

Pull request overview

This PR expands the resume parsing and evaluation prompt templates with detailed, structured extraction rules (dates, canonical naming, validation, deduplication, scoring bands) to reduce LLM ambiguity and improve output consistency across the pipeline.

Changes:

• Added per-field extraction requirements and validation checks to section templates (basics/work/education/skills/projects/awards).
• Restructured GitHub project selection prompt with explicit filtering, ranking tiers, and reasoning requirements.
• Reworked resume evaluation system message into sectioned scoring guidance with bonuses/deductions and strict output constraints.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
prompts/templates/awards.jinja	Expands award extraction rules, sorting/dedup guidance, and validation.
prompts/templates/basics.jinja	Adds detailed per-field contact/profile extraction and normalization rules.
prompts/templates/education.jinja	Adds detailed education schema mapping, ordering, and date handling guidance.
prompts/templates/github_project_selection.jinja	Adds structured selection methodology, commit-tiering, and reasoning requirements.
prompts/templates/projects.jinja	Adds project boundary detection, URL/tech normalization, and validation guidance.
prompts/templates/resume_evaluation_system_message.jinja	Adds detailed scoring bands, bonuses/deductions, and strict JSON output constraints.
prompts/templates/skills.jinja	Adds category logic, canonical naming table, and whole-resume skill crawling rules.
prompts/templates/system_message.jinja	Adds global extraction rules (dates, dedup, mapping, formatting, error handling).
prompts/templates/work.jinja	Adds structured work schema mapping, date rules, ordering, and validation checks.

Suppressed comments (4)

prompts/templates/work.jinja:31

• The template instructs endDate="present" (lowercase) and includes an education-specific phrase ("graduations") in a work context. The codebase’s existing transformations emit "Present" (capitalized), so this will likely create inconsistent endDate values across the pipeline.

5. **endDate**:
   - Use the format `YYYY-MM` or `YYYY`.
   - For ongoing/current employment: `"present"` — always lowercase.
   - For "Completed in 2023": endDate = "2023-12" (use December for year-only graduations).
   - If no end date is given and the role is clearly past, infer from surrounding context or set to `null`.

prompts/templates/work.jinja:72

• Validation references the literal string "present" (lowercase), but other parts of the pipeline output "Present" (capitalized). Keeping this aligned avoids inconsistent endDate values downstream.

- All dates must be valid YYYY-MM format strings or the literal string "present".

prompts/templates/system_message.jinja:19

• This section uses lowercase "present" while other parts of the pipeline (e.g., transform.py) use "Present". Keeping a single canonical casing avoids inconsistent values across extracted sections.

3. **DATE RANGE HANDLING**:
   - For education: use the full duration of the degree program (enrollment to graduation, or enrollment to present).
   - For work experience: use employment start date to end date (or "present" if currently employed).

prompts/templates/education.jinja:36

• This template uses endDate="present" (lowercase). Elsewhere in the pipeline ongoing dates are represented as "Present" (capital P), so aligning casing here avoids inconsistent outputs.

   - For ongoing education: `"present"`.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.