Self-Improving Prompt Agent

A prompt optimization loop that automatically improves prompts through iterative mutation and evaluation — the same structure used in RLHF, AutoML, and production LLM fine-tuning pipelines, implemented in ~100 lines of Python.

Problem & Why

Prompt engineering is manual and intuition-driven. Most teams iterate by hand with no systematic way to know if a change actually improved output quality. This project asks: what if the system could evaluate its own prompt and keep only improvements?

What This Is

A minimal edit → evaluate → keep/reject → repeat loop that optimizes a prompt against a scoring function, with full run history logged.

Architecture

prompt.txt  →  loop.py (mutate)  →  judge.py (score)  →  keep if score > best
    ↑                                                           |
    └─────────────────────── iterate ←──────────────────────────────────────────────────────────┘

• prompt.txt — the artifact being optimized
• loop.py — generates candidate mutations (add constraint, clarify audience, specify tone, etc.)
• judge.py — scoring function; returns 0.0–1.0
• results.log — complete run history with mutation, score, and KEEP/REJECT decision

Evaluation & Results

10 optimization rounds on a fintech landing page prompt:

Round	Mutation	Score	Decision
1	Ask for rationale behind design choices	0.10	KEEP
2	Add pricing preview	0.20	KEEP
3	Add pricing preview (duplicate)	0.20	REJECT
4	Include a feature section	0.30	KEEP
5	Ask for rationale (duplicate)	0.30	REJECT
6	Emphasize hierarchy and spacing	0.50	KEEP
7	Define the target audience clearly	0.60	KEEP
8	Include a feature section (duplicate)	0.60	REJECT
9	Specify a strong CTA	0.70	KEEP
10	Mention visual style clearly	0.80	KEEP

Score: 0.10 → 0.80 across 10 rounds. 3 mutations rejected (no regression).

Before: "Act as a designer and design a modern landing page for a fintech startup."

After: Same base + target audience definition, hero section, pricing preview, feature section, hierarchy/spacing, strong CTA, visual style, rationale requirement.

How to Use

git clone https://github.com/Ruthwik-Data/self-improving-prompt-agent
cd self-improving-prompt-agent
python3 loop.py
# Results written to results.log

Lessons Learned

1. Optimization is bounded by your evaluation function. The system gets as good as judge.py allows — not better. This is the fundamental constraint in RLHF, AutoML, and every DSPy-style optimizer. Getting the eval right matters more than the loop.
2. Rejection is signal, not failure. 3 of 10 mutations were rejected. That's the system working — duplicate mutations don't compound, and the loop maintains the best known state without regression.
3. The loop structure scales. Replace judge.py with DeepEval metrics and loop.py mutations with an LLM, and this becomes DSPy. The architecture is the insight, not the implementation.

Known Limitations

• Scoring function is heuristic, not grounded in user preference or task performance
• Mutation strategy is rule-based; LLM-generated mutations would explore the space better
• 10 rounds is illustrative, not statistically significant
• No baseline comparison against human-optimized prompts