Vibe Coding Workshop — Submission PR Draft

Name: Prachi Sangaonkar
City / Group: Pune / Workshop
Date: 2026-04-17
AI tool(s) used: Antigravity (Gemini-based Coding Assistant)

Checklist — Complete Before Opening This PR

[x] agents.md committed for all 4 UCs
[x] skills.md committed for all 4 UCs
[x] classifier.py executes on test_[city].csv without errors
[x] results_[city].csv available in uc-0a/
[x] app.py for UC-0B, UC-0C, UC-X runs successfully
[x] summary_hr_leave.txt included in uc-0b/
[x] growth_output.csv included in uc-0c/
[x] Minimum 4 meaningful commits following the required format
[x] All sections below are completed

UC-0A — Complaint Classifier

Initial failure mode encountered:
Severity blindness. The early version failed to prioritize safety-related keywords like "injury" or "child" when they appeared alongside routine complaints such as potholes or streetlight issues.

Rule that resolved the issue (from agents.md):
"Urgent triggers: Flag as Urgent if the complaint mentions 'injury', 'blood', 'hospital', 'child', or 'doctor' regardless of other keywords."

Accuracy against answer key:
15 out of 15 rows matched.

Handling of severity signals:
Yes. All entries containing terms like "child," "hazard," "injury," or "fell" were correctly classified as Urgent.

Git commit message:
UC-0A Fix severity blindness and taxonomy drift: initial implementation lacked safety keywords and strict categories → implemented rule-based enforcement in classifier.py

UC-0B — Summary That Changes Meaning

Failure mode observed:
Clause omission and obligation softening. For example, Clause 5.2 (dual approval requirement) was reduced to a vague “requires approval.”

Clauses affected in naive output:

*Clause 5.2 (Approval required from BOTH Dept Head and HR Director)
*Clause 2.6 (Leave exceeding 5 days is forfeited)

Post-fix validation:
Yes. All 10 critical clauses are accurately captured with conditions preserved.

Scope bleed in naive prompt:
Yes. It introduced unsupported statements such as “standard leave practices apply.”

Git commit message:
UC-0B Fix clause omission and obligation softening: initial summarizer omitted dual-approval and strict deadlines → implemented strict clause mapping and multi-condition preservation in app.py

UC-0C — Number That Looks Right

Naive output result:
A single overall percentage representing total city-level growth.

Issues observed:

• Aggregated across all wards
• Ignored 5 null rows without reporting them

Fix implementation results:

• Global aggregation is blocked
• Requires explicit --ward and --category inputs

Handling of null values:
Yes. All null rows are logged and flagged with reasons such as “Audit freeze.”

Validation against reference values:
Yes. Outputs match expected values (Ward 1 Roads: +33.1% in July, −34.8% in October).

Git commit message:
UC-0C Fix aggregation level and silent nulls: initial version aggregated global totals and hid missing data → implemented ward-wise segmentation and mandatory null auditing in app.py

UC-X — Ask My Documents

Naive system response:
"Yes, you can use your personal phone for approved remote tools and work files."
(This incorrectly merged HR and IT policies.)

Issue identified:
Cross-document blending between IT and HR policies.

Corrected system response:
"Answer: Personal devices may be used to access CMC email and the CMC employee self-service portal only. [policy_it_acceptable_use.txt, Section 3.1]"

Use of hedging language:
No. All uncertain responses now follow a strict refusal template.

Test question outcomes:
Yes. All 7 questions resulted in either:

• A properly cited single-source answer, or
• The exact refusal template

Git commit message:
[UC-X] Fix cross-document blending and hedged hallucination: implemented single-source retrieval and strict refusal template in app.py

CRAFT Loop Reflection

Most challenging step:
Analyze — It required careful inspection to identify subtle issues like softened legal language or incorrect aggregation logic.

Most important manual addition to agents.md:
The exact Refusal Template for UC-X. Without it, the AI tended to produce vague or misleading responses.

Planned real-world application:
Improving internal technical documentation and onboarding materials by ensuring strict citation of security policies and avoiding unsupported assumptions.