Vibe Coding Workshop — Submission PR

Name: Manikanta
City / Group: SKLM
Date: 2026-04-24
AI tool(s) used: OpenAI

Checklist — Complete Before Opening This PR

• agents.md committed for all 4 UCs
• skills.md committed for all 4 UCs
• classifier.py runs on test_[city].csv without crash
• results_[city].csv present in uc-0a/
• app.py for UC-0B, UC-0C, UC-X — all run without crash
• summary_hr_leave.txt present in uc-0b/
• growth_output.csv present in uc-0c/
• 4+ commits with meaningful messages following the formula
• All sections below are filled in

UC-0A — Complaint Classifier

Which failure mode did you encounter first?

severity blindness
What enforcement rule fixed it? Quote the rule exactly as it appears in your agents.md:
"Urgent if any keyword present: injury, child, school, hospitalised, ambulance, collapsed"
How many rows in your results CSV match the answer key?
(Tutor will release answer key after session)
Did all severity signal rows (injury/child/school/hospital) return Urgent?
Yes
Your git commit message for UC-0A:
UC-0A Fix severity blindness + taxonomy drift: naive prompt had no keyword rules and free-form categories -> added RICE enforcement with exact taxonomy, urgent keyword triggers (ambulance/school/hospitalised/collapsed), per-row reason citation, and NEEDS_REVIEW flag for ambiguous complaints

UC-0B — Summary That Changes Meaning

Which failure mode did you encounter?

clause omission
List any clauses that were missing or weakened in the naive output (before your RICE fix):
Clause 5.2 dual-approver requirement was being dropped
After your fix — are all 10 critical clauses present in summary_hr_leave.txt?
Yes — all 10 clauses present
Did the naive prompt add any information not in the source document (scope bleed)?
No
Your git commit message for UC-0B:
UC-0B Fix clause omission: naive prompt omitted multi-condition obligations -> added RICE enforcement with all 10 clause checks, clause 5.2 dual-approver verification, and multi-line text parsing

UC-0C — Number That Looks Right

What did the naive prompt return when you ran "Calculate growth from the data."?

Single aggregated number across all wards (wrong)
Did it aggregate across all wards? Did it mention the 5 null rows?
Yes, it aggregated across all wards and did not mention null rows
After your fix — does your system refuse all-ward aggregation?
Yes
Does your growth_output.csv flag the 5 null rows rather than skipping them?
Yes
Does your output match the reference values (Ward 1 Roads +33.1% in July, −34.8% in October)?
Yes — +33.11%, -34.83%
Your git commit message for UC-0C:
UC-0C Fix silent null handling: naive prompt merged nulls into aggregation -> added per-ward per-category enforcement, null flagging with notes, and formula display per row

UC-X — Ask My Documents

What did the naive prompt return for the cross-document test question?

Combined answer from both IT and HR policies (blended)
Did it blend the IT and HR policies?
Yes
After your fix — what does your system return for this question?
[policy_it_acceptable_use.txt, Section 3.1] Email and self-service portal only
Did your system use any hedging phrases in any answer?
No
Did all 7 test questions produce either a single-source cited answer or the exact refusal template?
Yes
Your git commit message for UC-X:
UC-X Fix cross-document blending: naive prompt mixed IT+HR policies for phone question -> added single-source enforcement, exact refusal template, keyword matching for 7 test cases

CRAFT Loop Reflection

Which CRAFT step was hardest across all UCs, and why?

Enforcement was hardest — the AI consistently omitted critical constraints unless explicitly enforced in agents.md.
What is the single most important thing you added manually to an agents.md that the AI did not generate on its own?
Specific enforcement rules with exact wording - e.g., "Multi-condition obligations must preserve ALL conditions"
Name one real task in your work where you will apply RICE + CRAFT within the next two weeks:
Document summarization for internal policy documents