Optimize Prompt

optimize-prompt -- one-pass prompt rewrite

Grounded in platform.claude.com prompting best-practices + prompting-tools (the Console prompt improver's 4-step transform) + the model-specific pages (Fable 5, Sonnet 5, Opus 4.8); community: arxiv 2601.22025 (generic rewrites measured net-negative), Decagon GEPA production notes (length caps = regularization), DreamHost 25-technique test (emphasis devices dead; XML/few-shot/ordering alive). When maintaining this skill, re-verify against the LIVE docs, not memory.

Step 1 -- target + context

• Get the draft: pasted text, @-mentioned file, or a described location (find it, Read to EOF).
• If context is absent AND material, ask ONCE, one message: target surface (Claude Code

SKILL / API system prompt / scheduled-task prompt / chat one-off) and latency-vs-accuracy preference. Defaults: accuracy; Claude 4.6+/Fable-era models.

• --check flag → run Step 2 only, report the defect list, stop. No rewrite.
• SKILL.md detection: if the target is a Claude Code SKILL.md, additionally apply the

skill-authoring conventions: description in third person carrying the WHEN-triggers (discovery lives in YAML, execution in the body); body is a procedure to run, not reference prose; , , , ); variable parts → {{template_variables}}; long data ABOVE instructions/query (20k+ context: query-at-end is worth up to ~30%); multi-doc inputs → nested ` shape.

• Instructions: negative constraints → positive equivalents ("write flowing prose" not

"no markdown"); add the WHY to rules so the model generalizes; vague scope → explicit ("every section, not just the first"); strip MUST/CRITICAL/ALL-CAPS emphasis (4.5+ models overtrigger on it); explicit output-format spec; imperative verbs for tool actions.

• Examples: wrap in `` tags; flag if 5; ensure at least one edge case;

reasoning tasks get a `` block inside examples; flag examples written for older model generations as stale.

• Reasoning: complex tasks get numbered analysis steps + a self-check line ("before

finishing, verify against "); prefer "consider/evaluate" over "think" when extended thinking is off.

• Model-currency sweep (4.6+/Fable): remove assistant prefills (400 error now);

budget_tokens → adaptive thinking + effort; remove "reproduce your reasoning in the response" asks (Fable refusal trigger); dial back anti-laziness/over-prescriptive language carried from 3.x-era prompts — newer models DEGRADE under it, so cutting old instructions is often the optimization.

• Anti-bloat regulator: compress examples aggressively, instructions barely (asymmetric

compression); if the rewrite exceeds 1.5x the input length, cut before delivering.

Step 4 -- report

Output to chat:

1. The optimized prompt in a fenced block.
2. Bulleted "what changed and why" — each bullet names the rule applied.
3. Deliberate non-changes ("kept your negative constraint because it carries a why").

If the source is a file and the user asks for a file, write a SIBLING .optimized.md. Never overwrite the source; the user applies it.

Never (failed approaches, pre-encoded)

• No eval loops, golden sets, or LLM-as-judge iteration — judges carry systematic biases

and loops degrade past iteration 2-3. One pass, human reviews.

• No generic "helpful assistant" framing added to task prompts (measured net-negative).
• No multi-candidate tournaments / token-burning fan-out.
• No blank-page generation from a vague wish — require a draft or a concrete task description.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dtiger1889-ops
• Source: dtiger1889-ops/claude-harness-toolbox
• License: MIT

Install and usage instructions live in the source repository linked above.