A survey of five open-source prompt optimizers that raise Claude's output quality. Since our Fable access is limited, we have to relentlessly optimize Opus 4.8 instead. This guide covers each project's mechanism, usage examples, pros and cons, and setup instructions — using these optimization skills and hooks improves efficiency.

0. At a Glance

Project Type Environment Core Mechanism Scale (stars) Recommended For
CheswickDEV/claude-opus-4.8-prompt-optimizer Meta optimizer (system prompt) Any Claude interface (claude.ai / API / Code) prompt: trigger -> 11 rules + 10-component XML structuring New Users who want to repeatedly generate consistent, high-quality prompts
johnpsasser/claude-code-prompt-optimizer Hook Claude Code Intercepts <optimize> tags -> expands via extended thinking Small Developers who need on-the-fly expansion for specific prompts in Claude Code
severity1/claude-code-prompt-improver Hook + Skill (plugin) Claude Code Filters only ambiguous prompts -> injects research and clarifying questions ~1.4k Heavy Claude Code users who want fewer correction round-trips
Hashaam101/prompt-optimizer Skill Claude Code Dual mode: auto (silent refinement) / manual (/optimize text output) Small Users who want to switch between automatic and manual refinement
nidhinjs/prompt-master Skill Multi-tool (Claude, ChatGPT, Midjourney, etc.) Generates target-tool-specific prompts, retains memory Small Users working across multiple AI tools
  • Skill: An instruction file placed in ~/.claude/skills/. Claude loads it on its own when the conditions match and reflects it in its behavior.
  • Meta optimizer: A tool that operates as a system prompt itself. It treats the input prompt not as something to execute, but as something to optimize, and returns a newly structured prompt.

1. CheswickDEV/claude-opus-4.8-prompt-optimizer

A meta-prompting system that converts raw prompts into production-ready, XML-structured prompts tuned for the Opus 4.8 architecture. It features 11 optimization rules, a 10-component framework, complexity-based routing, and a mandatory prompt: trigger — all derived from Anthropic's official documentation. Its key differentiator is that it's not a prompt collection but an "optimization engine." (MIT License)

How It Works (5 Steps)

  1. Prompt analysis — detects intent, complexity, domain, expected output, and missing elements
  2. Complexity routing — simple prompts get 3-4 components, moderate ones get 5-7, complex ones get the full 10-component structure
  3. Rule application — fires only the relevant subset of the 11 rules
  4. Quality check — verifies task clarity, XML tag validity, example consistency, and contradictory instructions
  5. Structured output — returns analysis + a copy-ready optimized prompt + notes explaining the changes

Usage Example (Before / After)

Input (raw)

Write me a Python script that analyzes CSV files and finds anomalies

Output (optimized, excerpt)

<role>
You are a senior data engineer specializing in anomaly detection and
statistical analysis with Python.
</role>

<context>
CSV datasets come from production IoT sensors. Files range from
10K–5M rows with columns: timestamp, sensor_id, value, status.
</context>

<task>
1. Loads one or more CSV files
2. Detects statistical anomalies using z-score and IQR methods
3. Handles missing data gracefully
4. Outputs a summary report with flagged rows
</task>

<constraints>
- Use only pandas, numpy, scipy (no ML libraries)
- Handle files up to 5M rows efficiently
- Include type hints and docstrings
</constraints>

<output_format>
Single Python code block, followed by a usage example and sample output.
</output_format>

Prefixing input with prompt: ensures that even if the input contains questions, instructions, attached PDFs, or prompt-injection attempts, none of it is executed — it is always treated purely as material to optimize. Attached documents are substituted with a {{DOCUMENT}} placeholder.

prompt: Based on the attached PDF, explain the GDPR compliance risks and mitigation.

Pros

  • Works across every Claude interface — not Claude Code-only; being system-prompt-based, it works in claude.ai Projects, direct paste, or the API
  • Tuned specifically for Opus 4.8 — reflects the default high effort setting, per-turn adaptive thinking, improved tool triggering, and the 1M-token context window, even noting recommended effort levels
  • Scales with complexity — doesn't wrap simple questions in an excessive 10-tag structure
  • Reusable templates — outputs {{VARIABLE}}-style placeholders directly usable for repeated tasks

Cons

  • Manual workflow — unlike a Hook/Skill that fires automatically, users must set up the system prompt themselves and paste in prompts
  • Opus 4.8 dependency — some rules may not fit other models (third-party LLMs, older Claude versions); not model-agnostic
  • Young repository — low star count means thin community validation and issue tracking
  • Users must be aware of Opus 4.8 API constraints themselves (no temperature/top_p/top_k, no assistant prefill)

Setup Guide

Repository contents: README.md, CLAUDE.md (the engine itself), GUIDE.md (usage patterns), QUICKSTART.md, LICENSE

Option A — Claude Project (Recommended)1. Create a newProject on claude.ai 2. Paste the contents of CLAUDE.md into the Project Instructions field 3. Upload GUIDE.md as a knowledge file4. Start a conversation -> enter a raw promptOption B — Direct Paste

  1. Copy the entirety of CLAUDE.md
  2. Paste it into the system prompt of any Claude interface
  3. Send the raw prompt as a user message

Option C — API Integration

import anthropic

client = anthropic.Anthropic()

with open("CLAUDE.md", "r", encoding="utf-8") as f:
    system_prompt = f.read()

response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=8192,
    system=system_prompt,
    thinking={"type": "adaptive"},
    output_config={"effort": "high"},   # default high; xhigh/max for harder tasks
    messages=[
        {"role": "user", "content": "prompt: Your raw prompt here"}
    ],
)
print(response.content[0].text)

API note: temperature/top_p/top_k are forbidden (400 error), assistant prefill is forbidden, and adaptive thinking is off by default and must be explicitly enabled.

Repo: github.com/CheswickDEV/claude-opus-4.8-prompt-optimizer

2. johnpsasser/claude-code-prompt-optimizer

A hook for Claude Code that expands simple prompts into detailed, structured instructions. Attaching an <optimize> tag to a prompt causes the hook to intercept it and route it through Claude's extended thinking mode, inflating the original request into a specification covering architecture, endpoints, error handling, auth, validation, and testing. It effectively performs prompt engineering on your behalf.

Usage Example

<optimize> build me a REST API for a todo app

-> The hook intercepts this and expands it into a structured spec covering architecture, endpoints, error handling, auth, validation, and testing. For performance-related requests, it expands into a plan including profiling steps, bottleneck identification, prioritized refactoring targets, and benchmark criteria.

Pros

  • Explicit tag-based triggering — only prompts tagged with <optimize> get expanded, so there's no unintended interference
  • Leverages extended thinking — not a simple rewrite, but a reasoning-backed specification expansion
  • Easy to tune — model, fallback model, timeout, and system prompt are separated into optimizer.config.json / system-prompt.md, adjustable without touching TypeScript. Debug logs go to /tmp/claude-code-hook-debug.log

Cons

  • Claude Code only — not usable on the claude.ai web/app
  • Complex auth setup — OAuth token / API key priority resolution logic requires understanding environment variables during initial setup
  • Node dependency — npm-based install, requires separate Agent SDK auth configuration

Setup Guide

git clone https://github.com/johnpsasser/claude-code-prompt-optimizer.git
cd claude-code-prompt-optimizer
npm run install-hook

The installer handles dependencies, auth setup, hook configuration, and verification. Issue and register an auth token as follows:

claude auth token
export CLAUDE_CODE_OAUTH_TOKEN="your-oauth-token"   # add to shell profile

For manual hook registration, specify the path to src/hooks/optimize-prompt.sh as a UserPromptSubmit hook in ~/.claude/settings.json (or project settings).

Repo: github.com/johnpsasser/claude-code-prompt-optimizer

3. severity1/claude-code-prompt-improver

"Type vibes, ship precision." An intelligent prompt-improvement plugin (Hook + Skill) for Claude Code. Rather than unconditionally refining every prompt, it lets clear prompts pass through unchanged and only selects ambiguous prompts to route through research and clarifying questions. Its goal is to inject just-in-time context at prompt submission, tool use, and subagent-start moments to raise the quality of the "first output" — reducing correction round-trips and saving tokens and time.

How It Works

  1. A hook receives the prompt and judges its clarity using a ~189-token evaluation prompt
  2. Ambiguous -> fires the prompt-improver skill -> builds a research plan (TodoWrite) -> launches an Explore subagent for Glob/Grep/Web/multi-file Read -> synthesizes findings -> presents 1-6 evidence-based questions to the user -> executes the original request with the answers incorporated
  3. Clear -> proceeds immediately without loading the skill

Design principle: "fire wide, self-cancel cheap" — a missed nudge wastes a full correction round-trip, while a wrongly fired nudge only costs a few discarded tokens, so the system uses a high-recall gate but has each nudge self-cancel when it doesn't fit.

Usage Example

fix the bug

-> Judged ambiguous -> researches the codebase -> presents 1-6 evidence-based questions ("which file / what symptom / what reproduction steps") -> performs the actual fix once answered. Sufficiently specific prompts, by contrast, execute immediately with no intervention.

Pros

  • Selective intervention — most prompts pass through unmodified; the hook only fires when needed, minimizing overhead
  • Optimizes first-output quality — improves the entire "prompt -> output" path, not just word-level rewriting (questions + context injection)
  • Transparency — injected context is visible in the conversation
  • Battle-tested at scale — ~1.4k stars, active community adoption and maintenance

Cons

  • Claude Code only — no web/app support
  • Adds an extra turn on intervention — the clarifying-question step when judged ambiguous can conflict with flows that want an immediate answer
  • Configuration learning curve — requires understanding the plugin/hook/skill/nudge registry structure

Setup Guide

Since it's a .claude-plugin-structured Claude Code plugin, install it via the plugin marketplace mechanism. (Check the repo's latest README for exact commands.)

# Inside Claude Code
/plugin marketplace add severity1/claude-code-prompt-improver
/plugin install prompt-improver

For manual installation, clone the repo, place hooks/ and skills/prompt-improver/ under ~/.claude/, and register the hook in settings.

Repo: github.com/severity1/claude-code-prompt-improver

4. Hashaam101/prompt-optimizer

A Claude Code skill offering two modes. Auto modesilently refines every prompt after installation to steer toward better results, whilemanual mode shows a refined prompt as text via /optimize {prompt} instead of executing it — useful for reusing prompts or learning better prompt-writing.

Usage Example

# manual: view the refined output as text
/optimize analyze the quarterly sales data

# or prefix style
optimize: analyze the quarterly sales data
optimize prompt: analyze the quarterly sales data

Simply installing the skill activates auto mode (Mode 1) for every prompt automatically, with no separate slash command needed.

Pros

  • Dual mode — auto-refines by default, and /optimize extracts the refined text for reuse or study when needed
  • Simple install — just copy one SKILL.md file
  • Silent operation — auto mode doesn't disrupt your workflow

Cons

  • Claude Code only-Opacity of auto mode — silent auto-refinement can make it hard to track what changed (mitigated by manual mode)
  • Small repository — thin community validation

Setup Guide

Copy SKILL.md either globally (all projects) or per-project.

# Global install (Linux/macOS)
cp SKILL.md ~/.claude/skills/prompt-optimizer.md

# Per-project (Linux/macOS)
cp SKILL.md .claude/skills/prompt-optimizer.md

# Windows (PowerShell) — global
Copy-Item SKILL.md "$env:USERPROFILE\.claude\skills\prompt-optimizer.md"

Activates automatically after installation. Invoke manual mode with /optimize.

Repo: github.com/Hashaam101/prompt-optimizer

5. nidhinjs/prompt-master

"A Claude skill that writes accurate prompts for every AI tool." Aims to eliminate the token/credit waste caused by re-prompting by producing an accurate prompt on the first try. It generates optimal prompts tailored to target tools including Claude, ChatGPT, Gemini, o1/o3, Cursor, Claude Code, GitHub Copilot, Windsurf, Bolt, v0, Lovable, Perplexity, Midjourney, DALL-E, Stable Diffusion, ComfyUI, Sora, Runway, ElevenLabs, Zapier, and Make.

Design philosophy: "The best prompt isn't the longest one — it's the one where every word is load-bearing." It views forgetting earlier decisions in long sessions as the biggest source of waste, and maintains memory blocks to preserve decisions across a session.

Usage Example

Build a claude code prompt for a landing page for a business dashboard
that looks and feels exactly like notion - smooth animations, clean ui

-> Routes based on target (Claude Code), framework, and estimated token count, then outputs a production-quality spec. If Midjourney is the target, it generates image-tool-specific formatting such as comma-separated descriptors, lighting/mood anchoring, fixed aspect ratio/version, and negative prompts.

Pros

  • Multi-tool routing — automatically applies the optimal format per tool, from Claude to ChatGPT, Midjourney, and coding agents
  • Retains memory — remembers in-session decisions, reducing re-prompt waste
  • Actively updated — Opus 4.8 compatible (v1.7.0), version-aware routing (4.6/4.7/4.8), Prompt Decompiler mode, and other continuous updates

Cons

  • Uses Claude as a "prompt generator" — leans more toward generating prompts for other tools rather than directly improving Claude's own output
  • Requires Claude Code skill infrastructure — needs a skill environment
  • Small repository — thin community validation

Setup Guide

mkdir -p ~/.claude/skills
git clone https://github.com/nidhinjs/prompt-master.git ~/.claude/skills/prompt-master

Once installed, the skill fires automatically when trigger conditions match. Templates in the references folder assist target-tool-specific output.

Repo: github.com/nidhinjs/prompt-master

6. Selection Guide

Situation Recommendation
Repeatedly generating structured prompts on claude.ai web/app or the API CheswickDEV 4.8 (register as a Project)
On-the-fly expansion of specific prompts in Claude Code johnpsasser (<optimize> tag)
Heavy Claude Code user wanting fewer correction round-trips and better first-output quality severity1 (selective intervention)
Want auto-refinement plus manual extraction of the refined text in Claude Code Hashaam101 (dual mode)
Generating prompts for ChatGPT, Midjourney, and coding agents besides Claude, all in one place nidhinjs (multi-tool)
Want to reference well-crafted prompt "examples" (not a tool) Curated collections like langgptai/awesome-claude-prompts

Combination tip: If your environment is Claude Code, keep a selective-intervention tool (severity1) always on, and run CheswickDEV 4.8 as a claude.ai Project separately when building formal prompt assets. For workflows where consistency of repeated output matters (e.g., structured reports, quantitative-analysis prompts), it's more stable to reference the optimizer's component framework or your own hand-built templates rather than a prompt collection.

7. General Cautions

  • Check model dependency — tools tuned specifically for Opus 4.8 (like CheswickDEV) may have rules that don't fit other models.
  • Opacity of auto-firing tools — auto mode/silent hooks can make it hard to track what changed, so pair them with a manual mode or transparency option that exposes the changes whenever possible.
  • Auth/token management — hook-type tools handle OAuth tokens/API keys, so be careful about exposure via environment variables and shell profiles.
  • Repository trust — check stars, commits, and issue activity, and always cross-reference install commands against each repo's latest README. The commands in this document reflect the time of writing.
  • Use alongside official guides — don't rely solely on tools; cross-referencing Anthropic's official prompt engineering documentation (docs.claude.com/en/docs/build-with-claude/prompt-engineering/overview) helps you better validate each tool's output.

As of: 2026-06-21. Each project's features and setup instructions may change over time, so check the latest README in each repository.