Prompt optimization

Make prompts clearer, compact verbose input before LLM calls, and preserve exact user intent.

arka prompt_optimize "summarize this research paper" --rounds 3 --role "research assistant"
arka "optimize this prompt: explain the experiment and cite assumptions"

Arka applies bounded refinement passes that clarify objectives, constraints, missing-information behavior, verification, and output structure. It is an original implementation inspired by the evaluation-oriented workflow described by Prompt Optimizer, whose repository is licensed AGPL-3.0-only. No source code is copied or bundled.

Compact intermediate form

For verbose natural-language inputs, Arka also converts the user text into a compact intermediate representation before sending it to the LLM. The original text is still used for symbolic routing, security checks, exact-command detection, and audit metadata.

The compact form keeps:

• high-signal requirements
• a small amount of context
• literal URLs, file paths, code blocks, JSON, and shell commands
• anti-hallucination constraints

It removes or compresses:

• repeated sentences
• low-signal filler
• excess whitespace

Configuration:

ARKA_PROMPT_OPTIMIZE=1
ARKA_PROMPT_COMPACT=1
ARKA_PROMPT_COMPACT_MIN_CHARS=900
ARKA_PROMPT_COMPACT_MAX_ITEMS=24

Opt out when exact wording must reach the model:

ARKA_PROMPT_COMPACT=0 arka ask "..."
arka llm complete --system "You are Arka" --user "..." --no-compact

Compaction is skipped for routing prompts and already-compact Task IR: prompts to avoid recursive rewriting.