ckelsoe/prompt-architect

Prompt Architect

Transform vague prompts into expert-level, structured prompts using 31 frameworks across 7 intent categories.

Works with Claude Code, ChatGPT, Gemini CLI, Cursor, GitHub Copilot, Windsurf, OpenAI Codex, and 30+ Agent Skills compatible tools.

Quick Start

npx @ckelsoe/prompt-architect

The interactive installer detects your AI agents (Claude Code, Gemini CLI, Cursor, Copilot, Codex, and more) and lets you choose where to install.

Tip: Use npx rather than npm install. Both work — npm install runs the same installer via the postinstall hook and installs to every agent it detects — but only npx gives you the interactive menu to choose specific targets.

Published on the public npm registry. No authentication required.

---

Table of Contents

• Overview
• Key Features
• Example Transformation
• Supported Frameworks
• Quick Start
• Installation
• Verifying Your Installation
• Updating
• Usage
• Framework Selection Guide
• Project Structure
• Documentation
• Contributing
• Support
• License

---

Overview

Prompt Architect is an Agent Skills compatible skill that elevates your prompting capabilities through:

• Intelligent Analysis - Evaluates prompts across 5 quality dimensions (clarity, specificity, context, completeness, structure)
• Framework Recommendation - Suggests the best framework(s) for your specific use case with clear reasoning
• Guided Dialogue - Asks targeted clarifying questions to gather missing information progressively
• Systematic Application - Applies selected framework to transform your prompt
• Iterative Refinement - Continues improving based on feedback until perfect

Target Audience:

• Developers using AI coding agents (Claude Code, Gemini CLI, Cursor, Copilot, etc.)
• Prompt engineers optimizing LLM interactions
• AI practitioners seeking systematic prompt improvement
• Teams wanting consistent, high-quality prompts

---

Key Features

31 Frameworks Across 7 Intent Categories

| Framework | Best For | Complexity | |-----------|----------|------------| | CO-STAR | Content creation, writing tasks | High | | RISEN | Multi-step processes, procedures | High | | CRISPE | Comprehensive prompts with multiple output variants | High | | BROKE | Business deliverables with OKR-style measurable outcomes | Medium | | RISE-IE | Data analysis, transformations (Input-Expectation) | Medium | | RISE-IX | Content creation with examples (Instructions-Examples) | Medium | | TIDD-EC | High-precision tasks with explicit dos/don'ts | Medium | | RACE | Expert tasks requiring role + context + outcome clarity | Medium | | CARE | Constraint-driven tasks with explicit rules and examples | Medium | | CTF | Simple tasks where situational context drives the prompt | Low | | RTF | Simple, focused tasks where expertise framing matters | Low | | APE | Ultra-minimal one-off prompts | Low | | BAB | Rewriting, refactoring, transforming existing content | Low | | Tree of Thought | Decisions requiring exploration of multiple approaches | Medium | | ReAct | Agentic / tool-use tasks with iterative reasoning | Medium | | Skeleton of Thought | Structured long-form content (outline-first) | Medium | | Step-Back | Principle-grounded reasoning (abstract first, then specific) | Medium | | Least-to-Most | Compositional multi-hop problems (simplest first) | Medium | | Plan-and-Solve (PS+) | Zero-shot numerical/calculation reasoning | Low | | Chain of Thought | Reasoning, problem-solving | Medium | | Self-Consistency | Robust answers: sample many reasoning paths, majority-vote | Medium | | Chain of Density | Summarization at fixed length, maximum information density | Medium | | Iterative Compression | Shortening text toward a target length | Low | | Self-Refine | Iterative output quality improvement (any task) | Medium | | CAI Critique-Revise | Principle-based critique and revision (Anthropic) | Medium | | Devil's Advocate | Strongest opposing argument against a position | Low | | Pre-Mortem | Assume failure, identify specific causes | Low | | RCoT | Verify reasoning by reconstructing the question | Medium | | Chain-of-Verification | Fact-check a draft by verifying each claim independently | Medium | | RPEF | Recover/reconstruct a prompt from an existing output | Low | | Reverse Role Prompting | AI interviews you before executing | Low |

Composable technique (layered onto any framework, not one of the 31): Few-shot / in-context examples — the discipline of adding well-chosen, well-ordered worked examples to a prompt.

Quality Scoring System

Every prompt is evaluated across:

• Clarity (1-10): Is the goal clear and unambiguous?
• Specificity (1-10): Are requirements detailed enough?
• Context (1-10): Is background information provided?
• Completeness (1-10): Are all necessary elements present?
• Structure (1-10): Is the prompt well-organized?

Progressive Disclosure

• Asks 3-5 targeted questions at a time (never overwhelming)
• Adapts questions based on framework selection
• Builds understanding iteratively
• Confirms assumptions before proceeding

Continuous Iteration

• Shows before/after comparisons with explanations
• Refines based on your feedback
• Can switch frameworks if initial choice doesn't fit
• Continues until you're satisfied

---

Example Transformation

Before

"Write about machine learning"

Analysis Scores:

• Clarity: 2/10 (vague goal)
• Specificity: 1/10 (no details)
• Context: 0/10 (no background)
• Completeness: 2/10 (missing most elements)
• Structure: 3/10 (single sentence)
• Overall: 1.6/10

---

After (CO-STAR Framework)

CONTEXT:
Creating content for a business blog aimed at C-level executives exploring
how AI/ML could benefit their organizations. Readers understand business
strategy but have limited technical ML knowledge. Part of an emerging
technologies series.
OBJECTIVE:
Create an engaging article helping executives understand practical machine
learning applications relevant to their companies. Focus on demonstrating
tangible business value and real-world implementation without overwhelming
technical details.
STYLE:
Professional blog style combining narrative with bullet points. Include 2-3
real-world case studies. Structure with clear subheadings every 150-200 words.
Balance storytelling with concrete information. Avoid jargon; when necessary,
provide plain-language explanations.
TONE:
Professional yet approachable and conversational. Confident and authoritative
without being condescending. Practical and business-focused rather than
theoretical.
AUDIENCE:
C-suite executives and senior managers at mid-to-large enterprises who:
Make strategic technology investment decisions
Understand business metrics and ROI
Have limited technical ML knowledge
Value practical examples over theory

RESPONSE FORMAT:
800-word article structured as:
Compelling headline (10 words max)
Brief hook (2-3 sentences)
3-4 main sections with descriptive subheadings
Mix of paragraphs and bullet points
Clear call-to-action conclusion

Result Scores:

• Clarity: 9/10
• Specificity: 9/10
• Context: 10/10
• Completeness: 9/10
• Structure: 9/10
• Overall: 8.8/10

---

Supported Frameworks

CO-STAR (Context, Objective, Style, Tone, Audience, Response)

Best for: Content creation, writing tasks, communications

Components:

• Context - Background information and situation
• Objective - Clear goal and purpose
• Style - Writing style and formatting approach
• Tone - Voice and emotional quality
• Audience - Target reader characteristics
• Response - Expected format and structure

Example Use Cases: Blog posts, emails, presentations, marketing copy, documentation

---

RISEN (Role, Instructions, Steps, End goal, Narrowing)

Best for: Multi-step processes, systematic procedures

Components:

• Role - Expertise and perspective needed
• Instructions - High-level guidance
• Steps - Detailed methodology
• End goal - Success criteria
• Narrowing - Constraints and boundaries

Example Use Cases: Code reviews, workflows, systematic analysis, project planning

---

RISE (Dual Variants)

RISE-IE (Input-Expectation)

Best for: Data analysis, transformations, processing tasks

Components:

• Role - Expertise needed
• Input - Data format and characteristics
• Steps - Processing methodology
• Expectation - Output requirements

Example Use Cases: CSV analysis, data processing, file transformations, report generation

RISE-IX (Instructions-Examples)

Best for: Content creation with reference examples

Components:

• Role - Expertise needed
• Instructions - Task guidance
• Steps - Workflow process
• Examples - Reference samples

Example Use Cases: Creative writing, template-based content, style matching

---

TIDD-EC (Task, Instructions, Do, Don't, Examples, Context)

Best for: High-precision tasks requiring explicit boundaries

Components:

• Task type - Nature of the work
• Instructions - What to accomplish
• Do - Explicit positive guidance
• Don't - Explicit negative guidance (what to avoid)
• Examples - Reference samples
• Context - Background information

Example Use Cases: Code generation with standards, compliance tasks, quality-critical work

---

CTF (Context, Task, Format)

Best for: Simple tasks where situational background matters more than expertise framing

Components:

• Context - Situation and background
• Task - What needs to be done
• Format - Output structure

Example Use Cases: Handoff documents, mid-project updates, situation-driven requests

---

RTF (Role, Task, Format)

Best for: Simple, well-defined tasks where expertise framing drives output quality

Components:

• Role - Expertise required
• Task - What needs to be done
• Format - Output structure

Example Use Cases: Quick conversions, simple formatting, straightforward requests

---

APE (Action, Purpose, Expectation)

Best for: Ultra-minimal prompts — the simplest structured framework

Components:

• Action - What to do (one clear verb-driven instruction)
• Purpose - Why it's needed (one sentence)
• Expectation - What a good result looks like

Example Use Cases: Quick summaries, single-function code, one-off requests, rapid iteration

---

BAB (Before, After, Bridge)

Best for: Transforming, rewriting, or refactoring existing content

Components:

• Before - Current state and what's wrong
• After - Desired end state
• Bridge - Transformation rules and constraints

Example Use Cases: Code refactoring, copy rewrites, tone changes, document restructuring, version migrations

---

RACE (Role, Action, Context, Expectation)

Best for: Medium-complexity tasks needing expertise + background + explicit success criteria

Components:

• Role - Expertise or persona
• Action - What needs to be done
• Context - Situational background and constraints
• Expectation - What a good result looks like

Example Use Cases: Technical reviews, expert analysis, contextual recommendations, documentation with standards

---

CRISPE (Capacity+Role, Insight, Instructions, Personality, Experiment)

Best for: Comprehensive prompts where you want multiple output variants to compare

Components:

• Capacity & Role - Expertise level and professional persona
• Insight - Background context and situational knowledge
• Instructions - The specific task
• Personality - Tone, voice, communication style
• Experiment - Request N variants to compare (the defining differentiator)

Example Use Cases: Marketing campaigns (A/B variants), content with tone options, strategic analysis needing multiple angles

---

BROKE (Background, Role, Objective, Key Results, Evolve)

Best for: Business deliverables with measurable outcomes and built-in self-improvement

Components:

• Background - Situation and context
• Role - Professional persona
• Objective - The task
• Key Results - Measurable business outcomes (OKR-style)
• Evolve - AI self-critiques and suggests 3 improvements

Example Use Cases: Sales process improvements, content strategy with KPIs, product decisions tied to metrics

---

CARE (Context, Ask, Rules, Examples)

Best for: Tasks with explicit constraints, compliance requirements, or quality standards

Components:

• Context - Situation and background
• Ask - The specific request
• Rules - Explicit constraints, dos/don'ts, standards
• Examples - Reference samples showing the quality bar

Example Use Cases: Healthcare/legal content, UI error messages, interview questions with bias constraints, brand-compliant copy

---

Skeleton of Thought (SoT)

Best for: Structured long-form content — generate outline first, then expand

Approach:

• Phase 1: Generate a concise skeleton/outline (key points only)
• Phase 2: Expand each skeleton point independently (parallelizable)
• SoT-R variant: Route non-suitable questions to standard generation

Example Use Cases: Technical documentation, structured reports, tutorials, any multi-section content

---

Step-Back Prompting

Best for: Principle-grounded reasoning — abstract to the underlying concept first

Approach:

• Generate a higher-level "step-back" question about underlying principles
• Answer the abstract question to retrieve principles
• Use those principles as context to answer the original specific question

Example Use Cases: STEM problems, architecture decisions, debugging, any task where first-principles reasoning matters

---

Least-to-Most (LtM)

Best for: Compositional multi-hop problems with ordered dependencies

Approach:

• Decompose the complex problem into ordered subproblems (simplest first)
• Solve each subproblem sequentially, feeding each answer into the next
• Use accumulated answers to solve the original problem

Example Use Cases: Multi-domain questions (legal + technical), complex calculations, architecture problems with prerequisites

---

Plan-and-Solve (PS+)

Best for: Zero-shot numerical and calculation reasoning

Approach:

• "Let's first understand the problem, extract relevant variables and their corresponding numerals, and devise a complete plan. Then, let's carry out the plan, calculate intermediate values, pay attention to computation, and solve the problem step by step."

Example Use Cases: Financial calculations (MRR, CAC, payback), math word problems, resource estimation, any zero-shot reasoning task

---

Self-Refine

Best for: Iterative quality improvement of any output

Approach: Generate → produce specific actionable feedback → refine → repeat until stopping criterion Research: Madaan et al. NeurIPS 2023 — +5-40% improvement across 7 task types

Example Use Cases: Code review and rewriting, writing improvement, plan refinement, pre-submission QA

---

CAI Critique-Revise

Best for: Aligning output to an explicit stated principle or standard

Approach: Initial output → critique against a specific principle → revision addressing every critique point Research: Anthropic Constitutional AI (arXiv 2212.08073, 2022) — principle-driven alignment

Example Use Cases: Plain language compliance, brand voice enforcement, epistemic quality (claims vs. assertions), legal/regulatory language standards

---

Devil's Advocate

Best for: Generating the strongest possible opposing argument against a position

Approach: Explicitly instructs the AI to attack a position as forcefully as possible — not balanced, not a straw man, but maximum-strength opposition Research: ACM IUI 2024 peer-reviewed study

Example Use Cases: Decision stress-testing, architecture reviews, countering groupthink, stakeholder preparation, debiasing

---

Pre-Mortem

Best for: Identifying specific failure causes before they happen

Approach: Assume the project has already failed → describe the failure → work backwards to specific causes with warning signs Research: Mitchell, Russo & Pennington (1989), popularized by Gary Klein (HBR 2007) — prospective hindsight improves identification of reasons for future outcomes by ~30%; LLM application is practitioner-level, with no dedicated AI paper

Example Use Cases: Project kickoffs, product launches, technical migrations, high-stakes strategic decisions

---

RCoT (Reverse Chain-of-Thought)

Best for: Verifying that reasoning addressed all conditions in a multi-constraint question

Approach: Generate answer → reconstruct the question from the answer → cross-check conditions → correct overlooked items Research: Academic backward reasoning literature (ACL 2025, NAACL 2025)

Example Use Cases: Multi-condition logic problems, complex requirements analysis, high-stakes reasoning verification

---

RPEF (Reverse Prompt Engineering)

Best for: Recovering a reusable prompt template from an existing output

Approach: Provide an output (and optionally the input) → AI analyzes it for tone, structure, constraints, persona → generates a reusable template with [PLACEHOLDER] variables Research: Li & Klabjan, EMNLP 2025 (arXiv 2411.06729)

Example Use Cases: Recovering lost prompts, codifying successful one-time outputs, building style templates, understanding system prompt behavior

---

Reverse Role Prompting (AI-Led Interview)

Best for: Complex tasks where you know the goal but struggle to specify all requirements

Approach: Provide a minimal intent statement → AI asks targeted clarifying questions → executes once context is complete Research: FATA framework (arXiv 2508.08308, 2025) — ~40% improvement over standard prompting

Example Use Cases: Complex strategy tasks, non-expert users, requirements gathering, generating complete prompts from an interview

---

Tree of Thought

Best for: Decisions where multiple approaches need systematic comparison

Approach:

• Defines the problem and constraints
• Explores 2-5 distinct solution branches in parallel
• Evaluates each branch against defined criteria
• Synthesizes into a reasoned recommendation

Example Use Cases: Architecture decisions, debugging with multiple hypotheses, technology selection, strategic trade-offs

---

ReAct (Reasoning + Acting)

Best for: Agentic tasks that interleave reasoning with tool use

Approach:

• Defines goal, available tools, and constraints
• Alternates Thought → Action → Observation cycles
• Each observation informs the next thought
• Continues until goal is reached

Example Use Cases: Agentic workflows, multi-step research, debugging with tools, data investigation

---

Chain of Thought

Best for: Complex reasoning and problem-solving

Approach:

• Breaks down reasoning into explicit steps
• Shows work and intermediate conclusions
• Verifies logic at each stage
• Builds to final answer

Example Use Cases: Math problems, debugging, decision analysis, logical reasoning

---

Chain of Density

Best for: Iterative refinement and compression

Approach:

• Starts with baseline version
• Progressively refines through iterations
• Increases information density
• Optimizes for specific goals

Example Use Cases: Summarization, content compression, explanation optimization

---

Installation

Choose the method that matches your AI tool:

Claude Code

/install-skill https://github.com/ckelsoe/prompt-architect/tree/main/skills/prompt-architect

Plugin system (enables updates):

/plugin marketplace add ckelsoe/prompt-architect
/plugin install prompt-architect@prompt-architect-marketplace

Codex CLI

$skill-installer install https://github.com/ckelsoe/prompt-architect/tree/main/skills/prompt-architect

Other Agents (Gemini, Cursor, Copilot, Cline, Zoo Code, etc.)

Copy skills/prompt-architect/ from this repo to ~/.agents/skills/prompt-architect/

npm (alternative)

npx @ckelsoe/prompt-architect

Installs to both .claude/skills/ and .agents/skills/ in one step.

ChatGPT

ChatGPT natively supports Agent Skills. Available on Business, Enterprise, Edu, Teachers, and Healthcare plans.

1. Download prompt-architect.zip (direct download, always latest version) 2. In ChatGPT, click your profile icon → Skills 3. Click New skill → Upload from your computer 4. Upload the .zip file — do not unzip it first

ChatGPT scans uploaded skills before making them available; this one is plain Markdown with no scripts, so the scan is typically immediate.

Use the .zip, not the .skill file also attached to the release. They are

the same archive, but ChatGPT's uploader expects a .zip containing a single

top-level folder.

Windsurf

Windsurf does not yet support Agent Skills natively:

cat adapters/for-windsurf.md >> .windsurfrules

Any LLM API or Chat Interface

Paste the contents of adapters/system-prompt.md into your system prompt field. Works with OpenAI API, Anthropic API, Google Gemini API, Custom GPTs, or any tool with a system prompt.

See adapters/README.md for detailed instructions for each platform.

---

Verifying Your Installation

After installing, test with:

"Help me improve this prompt: write a technical blog post"

The skill will automatically: 1. Analyze your prompt and score it 2. Recommend the best framework (likely CO-STAR) 3. Ask clarifying questions 4. Generate an improved, structured prompt 5. Explain the changes made

---

Updating

# npm
npm update -g @ckelsoe/prompt-architect
Check version
npm list -g @ckelsoe/prompt-architect
Claude Code plugin
/plugin update prompt-architect@prompt-architect-marketplace

Migrating from v2.x? The package was renamed from @ckelsoe/claude-skill-prompt-architect. See MIGRATION.md for details.

---

Usage

Basic Usage

Simply ask Claude Code to improve any prompt:

"Help me improve this prompt: [your prompt here]"

Example Interaction

``` User: "Help me improve this prompt: create a function to process user data"

Skill: I'll analyze this prompt and help improve it.

Analysis:

• Clarity: 4/10 (basic goal stated)
• Specificity: 2/10 (what kind of data? what processing?)
• Context: 1/10 (no background)
• Completeness: 2/10 (missing many details)
• Structure: 3/10 (single sentence)
• Overall: 2.4/10

Recommendation: RISEN framework (multi-step process requiring methodology)

**Questions*