Alex Prompter Profile picture
Jun 27 18 tweets 6 min read Read on X
Most people try to get better code by tweaking prompts.

But one overlooked trick quietly boosts coding accuracy by 2x.

And barely anyone’s using it.

Here's the method 🧵: Image
1/ Coding with AI isn't like writing essays.

LLMs often fail on real code problems because:
– The spec is long
– Small details matter
– One wrong line breaks the whole thing

Single prompts just don’t cut it anymore. Image
2/ That’s where AlphaCodium comes in.

A new approach designed by CodiumAI researchers:
Instead of relying on prompts alone, it breaks the problem down step-by-step before generating any code.

A key step? Bullet Point Analysis. Image
3/ So what is Bullet Point Analysis?

Before writing a single line of code, the model is asked to explain the problem in structured bullet points:

- Goal
- Inputs
- Outputs
- Rules
- Edge cases

Why? To force the model to reason semantically. Image
4/ Think of it like this:

Humans use outlines to clarify their thoughts.
Bullet points help AI do the same.

Each bullet acts like a mini anchor - locking in a semantic chunk of meaning. Image
5/ And it works.

When they added bullet point analysis to their coding flow, GPT-4 accuracy jumped from 19% → 44% on a competitive coding benchmark (CodeContests).

That's more than 2x improvement - with no extra training. Image
6/ Why it matters for business too:

LLMs are great at language, but struggle with business complexity.

Bullet Point Analysis fixes this by:
– Forcing systematic thinking
– Reducing overlooked factors
– Creating accountability checkpoints
– Improving decision traceability
7/ Here are 5 prompts that use Bullet Point Analysis:

BUSINESS STRATEGY ANALYSIS:
Adopt the role of a strategy consultant for [COMPANY] in [INDUSTRY].

STEP 1: Before providing any recommendations, analyze this situation by creating bullet points for each area:
- Business context and current position
- Market dynamics and competitive landscape
- Internal capabilities and resources
- External opportunities and threats
- Success metrics and constraints
- Stakeholder interests and concerns

STEP 2: After completing your bullet point analysis above, now provide your strategic recommendations based on each bullet point you identified.

Context: We need to [GOAL] within [TIMEFRAME] with [BUDGET] budget targeting [MARKET].
8/ PRODUCT DEVELOPMENT DECISION:
Adopt the role of a product manager for [PRODUCT] at [COMPANY].

STEP 1: Before deciding on features, break down this product decision into bullet points:
- User needs and pain points we're solving
- Technical feasibility and resource requirements
- Business impact and revenue potential
- Competitive positioning and differentiation
- Implementation risks and mitigation strategies
- Success metrics and validation methods

STEP 2: Based on your bullet point analysis, recommend specific product features and development priorities.

Context: We're building [PRODUCT] for [MARKET] to achieve [GOAL] within [TIMEFRAME].
9/ MARKET ENTRY STRATEGY:
Adopt the role of a business development director for [COMPANY].

STEP 1: Before recommending market entry approach, structure your analysis in bullet points:
- Market size, growth, and opportunity assessment
- Customer segments and buying behavior analysis
- Competitive landscape and positioning gaps
- Regulatory requirements and market barriers
- Resource requirements and investment needs
- Risk factors and mitigation strategies

STEP 2: Using your bullet point analysis, recommend the optimal market entry strategy with specific tactics.

Context: Entering [MARKET] with [PRODUCT] to achieve [GOAL] within [TIMEFRAME].
10/ OPERATIONAL PROBLEM SOLVING:
Adopt the role of an operations consultant for [COMPANY] in [INDUSTRY].

STEP 1: Before proposing solutions, analyze this operational challenge through bullet points:
- Current process flow and identified bottlenecks
- Root causes and contributing factors
- Impact on customers, employees, and business metrics
- Available resources and capability constraints
- Implementation complexity and change management needs
- Success metrics and measurement methods

STEP 2: Based on your bullet point analysis, provide specific operational improvements with implementation roadmap.

Context: We need to solve [GOAL] within [TIMEFRAME] with [BUDGET] resources.
11/ CODING ARCHITECTURE DECISIONS:
Adopt the role of a senior software architect for [COMPANY].

STEP 1: Before recommending any code architecture, analyze this development challenge in bullet points:
- Business requirements and technical specifications
- Scalability needs and performance constraints
- Integration requirements with existing systems
- Security requirements and compliance needs
- Development team capabilities and timeline constraints
- Maintenance and long-term evolution considerations

STEP 2: Using your bullet point analysis, recommend specific architectural patterns, technologies, and implementation approach.

Context: Building [PRODUCT] for [MARKET] to handle [GOAL] within [TIMEFRAME].
12/ SYSTEM DESIGN PLANNING:
Adopt the role of a system design engineer for [COMPANY].

STEP 1: Before designing any system, break down the requirements into bullet points:
- Functional requirements and user interactions
- Non-functional requirements (performance, security, reliability)
- Data flow and storage requirements
- Scalability and availability constraints
- Integration points and external dependencies
- Monitoring and maintenance considerations

STEP 2: Using your bullet point analysis, design the system architecture with specific components and technologies.

Context: Designing [PRODUCT] system for [MARKET] to achieve [GOAL] within [TIMEFRAME].
13/ The pattern is simple:

Step 1: "Before [solving], analyze through bullet points: [key factors]" Step 2: "Based on your analysis, recommend [solutions]"

This forces semantic reasoning before generating answers.

Always include the phrase "Before [action], analyze through bullet points" to trigger the structured thinking mode.

The AI literally cannot skip the analysis step.
14/ And it’s not just bullet points.

AlphaCodium’s full flow includes:
– Test-based iteration
– AI-generated test cases
– YAML structured outputs
– Modular code generation
– Double validation for decisions

📰 Dive deeper into full paper:
arxiv.org/pdf/2401.08500
The AI prompt library your competitors don't want you to find

→ Unlimited prompts: $150 lifetime or $15/month
→ Starter pack: $3.99/month
→ Pro bundle: $9.99/month

Grab it before it's gone 👇
godofprompt.ai/pricing
That's a wrap! If you found this useful:
1/ Follow me @alex_prompter for more AI tips.
2/ Like & RT this post:
@UnrollHelper

• • •

Missing some Tweet in this thread? You can try to force a refresh
 

Keep Current with Alex Prompter

Alex Prompter Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!

PDF

Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @alex_prompter

Jun 24
ChatGPT-4o is a genius tool!

You can convert any logo into 3D terrain.

Prompt 👇 Image
1/ Step 1: Attach your logo or image.

Step 2: Run the JSON prompt below.

Full prompt 👇:
Recreate [BRAND NAME] logo following my JSON aesthetic below:
{
"role": "You are a geospatial visualization generator trained in realistic topographic modeling and satellite cartography. Your task is to transform any uploaded logo, subject, or image into a 3D terrain model — where the shape of the logo is used to generate elevation features like ridges, cliffs, peaks, and valleys. The terrain must preserve the precise shape of the uploaded subject, and be rendered in high-resolution as if captured from a top-down satellite or terrain mapping camera. All labeling and layout must respect the original structure.",
"instructions": {
"use_uploaded_subject": true,
"preserve_shape_and_text": true,
"embed_into_terrain": true,
"integration_mode": "elevation model — logo becomes terrain heightmap",
"terrain_details": {
"height_map_behavior": "edges of letters or logo outline are ridgelines; inner zones become valleys or plateaus",
"contour_lines": true,
"elevation_style": "realistic — varying steepness, visible slope curvature",
"textures": [
"satellite-style shaded relief",
"hillshade lighting simulation",
"contour lines at regular intervals"
]
},
"visual_options": {
"color_mode": [
"earth-tones",
"satellite topography colors (greens, browns, shadows)"
],
"extra_features": [
"tiny hikers or tents walking on logo contours",
"survey tripods, geologist tools placed on ridges",
"small labels or altitude numbers subtly embedded"
],
"optional_environment": [
"low fog in valleys",
"sunrise light angle from top right",
"thin cloud layer from above"
]
}
},
"image_constraints": {
"aspect_ratio": "1:1",
"resolution": "minimum 2400x2400",
"camera_angle": "orthographic top-down or 3/4 aerial angle",
"background": "flat terrain blend or gradient sky",
"logo_legibility": "must remain readable as shape, even if terrainified"
},
"visual_style": {
"map_style": "hybrid between shaded relief and GIS terrain overlay",
"shading": "realistic terrain shadows and light falloff",
"labeling_behavior": "optional — no country names or real labels, only fictional map-style markers if needed"
},
"notes": "Do not abstract or reshape the uploaded subject. It must retain its form while being interpreted as an elevation-based geographic structure. Avoid fantasy or flat map styles — this must look like a photoreal topographic data map with logo features as terrain."
}Image
2/ Amazon Image
Read 7 tweets
Jun 23
Bigger AI ≠ better AI.

A 9-billion-parameter model just outscored GPT-4 on tough reasoning tests.

The trick is Adaptive Prompting.

Quick breakdown + prompts 🧵:
1/ In 2024, researchers tested Gemma-9B on eight reasoning benchmarks.

With Adaptive Prompting, it beat GPT-4 on arithmetic and matched it on commonsense tasks. Image
2/ Adaptive Prompting = dynamic, multi-stage prompts that guide the model to understand, validate, and refine each step before answering. Image
Read 15 tweets
Jun 20
ChatGPT-4o is an incredible tool.

You can turn any logo into an escherian stairwell!

Prompt 👇 Image
1/ Step 1: Attach your logo or image.

Step 2: Run the JSON prompt below.

Full prompt 👇:
Recreate [BRAND NAME] logo following my JSON aesthetic below:
{
"role": "You are a surreal architectural image generator trained in the visual logic of M.C. Escher. Your task is to integrate the uploaded logo, image, or subject into a recursive, mind-bending landscape — including infinite staircases, impossible arches, and multi-perspective geometry. You must preserve the exact shape, color, and text of the uploaded logo, embedding it structurally or symbolically into the architecture. Use Escher-style surrealism: logical yet impossible.",
"instructions": {
"use_uploaded_subject": true,
"preserve_shape_and_text": true,
"embed_as_architectural_form": true,
"visual_structure": {
"primary_theme": "infinite staircases, shifting planes, mirrored portals, recursive arches",
"embedding_mode": [
"logo forms the stair base or arch frame",
"logo appears repeatedly across surfaces as structural support",
"logo is distorted by forced perspective but still readable"
],
"dimension_behavior": "multi-orientation logic — objects face different gravities, some stairs go upside down"
},
"visual_treatment": {
"style": "M.C. Escher engraving style or clean 3D sketch",
"color_mode": [
"grayscale tones",
"or use only brand colors from uploaded logo — no other hues"
],
"shading": "linework, stippling, soft engraving-style shadowing"
},
"surreal_elements": [
"upside-down figures walking on stairs",
"open voids and abstract windows",
"nested doorways leading to copies of the same space",
"non-Euclidean geometry with repeating logo-infused elements"
],
"optional_effects": {
"reflection_in_water": true,
"skybox": "white void or soft cloudy abstraction",
"shadows cast by impossible shapes"
}
},
"image_constraints": {
"aspect_ratio": "1:1",
"resolution": "minimum 2000x2000",
"scene_type": "surreal architectural environment",
"logo_placement": "structurally fused into impossible space — must remain fully legible"
},
"visual_style": {
"render_type": "engraving-style 3D illustration",
"depth": "extreme perspective depth with recursive logic",
"textures": "clean, geometric, stone or concrete finish"
},
"notes": "Do not reinterpret the uploaded logo. It must remain fully intact — only embedded into the architectural recursion. Avoid fantasy colors, neon lighting, or cartoonish elements. The output must evoke mathematical surrealism, impossible space, and intelligent abstraction in the visual spirit of M.C. Escher."
}Image
2/ Amazon Image
Read 8 tweets
Jun 18
ChatGPT-4o is insane.

You can convert any logo into MS paint drawing!

Prompt 👇 Image
1/ Step 1: Attach your logo or image.

Step 2: Run the JSON prompt below.

Full prompt 👇:
Recreate [BRAND NAME] logo following my JSON aesthetic below:
{
"role": "You are a retro computer graphics renderer simulating a real Windows 95/97 desktop environment. Your task is to take any uploaded logo or image and redraw it in a crude MS Paint style — as if it were recreated manually in Microsoft Paint — while fully preserving its original shape and exact colors. The result must be displayed inside a complete MS Paint window that sits within a classic Windows 95/97 desktop environment.",
"instructions": {
"subject_behavior": {
"use_uploaded_image_only": true,
"preserve_shape_and_layout_of_subject": true,
"preserve_original_colors_exactly": true,
"no_random_color_replacement": true,
"recreate_subject_as_sloppy_MS_Paint_drawing": true,
"simulate_mouse_drawn_style": true,
"fill_color_bleed_and_jagged_edges": true
},
"interface_layer": {
"include_full_MS_Paint_window": true,
"show_toolbar_on_left": true,
"show_color_bar_at_bottom": true,
"top_bar_should_read": "untitled - Paint",
"simulate_pixel-alignment_of_window_elements": true,
"align_subject_centered_on_canvas": true
},
"desktop_context": {
"render_desktop_background": "Windows 95 or 97 style blue gradient wallpaper",
"optional_elements": [
"My Computer icon",
"Recycle Bin icon",
"taskbar at bottom with Start button"
],
"position_paint_window_centered_with_padding": true,
"allow_window_to_occupy_70-85% of frame": true,
"keep_full_window_visible": true,
"no cropping of toolbar or edges"
},
"visual_style": {
"aesthetic": "low-res MS Paint, Windows 95 desktop",
"rendering_style": "screenshot with humor and nostalgia",
"text_style": "blocky system fonts or jagged hand-drawn text if part of original image"
},
"image_constraints": {
"aspect_ratio": "16:10",
"resolution": "1600x1000 minimum",
"output_type": "simulated desktop screenshot",
"no_transparency": true,
"full_window_frame_required": true
},
"notes": "Do not crop the MS Paint window. Center the entire Paint program in the image. Use the real Windows desktop as background so that aspect ratio doesn't require cropping. The uploaded subject must remain visually recognizable and match original colors exactly. Style should feel nostalgic, humorous, and like it was drawn by a 10-year-old."
}Image
2/ OpenAI: Image
Read 8 tweets
Jun 13
Why write prompts from scratch…

When AI can build them for you?

Try this easy method 👇🧵:
1/ Run this prompt with ChatGPT o3 or Claude Opus:

> Adopt the role of an expert prompt engineer.
Ask me 5 fast questions (goal, audience, must-have context, tone, format) before you create a detailed prompt for me.

After I answer these questions, perform a comprehensive research on best-fit prompt engineering technique based on my goal and context.

Think step-by-step and perform this task thoroughly.Image
2/ Result: Image
Read 9 tweets
Jun 11
ChatGPT just leveled up again.

A new method called “10-Shot + 1 AutoDiCoT” smashes through reasoning tasks.

It combines 10 human-crafted examples with 1 self-generated, smart one.

Here's why this works so well + prompts 👇🧵:
1/ The setup:

You want the model to reason step-by-step.

So you feed it:
• Full task context
• 10 typical examples (CoT format)
• 1 smart example it created itself (AutoDiCoT)

This combo boosts accuracy. Image
2/ What’s AutoDiCoT?

“Automatic Diverse CoT.”

It’s a method where the model generates multiple CoT answers…

Then selects the most consistent one.

It’s like having the model proofread itself. Image
Read 17 tweets

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Don't want to be a Premium member but still want to support us?

Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal

Or Donate anonymously using crypto!

Ethereum

0xfe58350B80634f60Fa6Dc149a72b4DFbc17D341E copy

Bitcoin

3ATGMxNzCUFzxpMCHL5sWSt4DVtS8UqXpi copy

Thank you for your support!

Follow Us!

:(