13 prompting techniques for better vibe coding.

small changes that yield massive performance improvements 🧵👇

1. if you're using anthropic models, you can use hard-coded "magic words" in your prompt to increase the token budget.

> think
> think hard
> think harder
> ultrathink

2. when role assigning, spell out responsibilities, tasks, and expected output.

don’t rely on titles the way most do with writing tasks.

ex:

❌ “you are a senior UI engineer that builds world-class, award-winning front-ends.”

✅ “you are a UI engineer responsible for designing, building and organizing reusable UI components. your job is to document all user stories and only build or refactor components to satisfy the user stories.”

3. add context dumps and information about the situation to the beginning of a prompt, and the task request at the end.

research shows LLMs attend most to tokens near the start (and end) of long prompts.

add details about the app, the end‑user, the problem your product is solving, your lack of technical skills, etc.

the model provider is more likely to “save” this information and reference it in follow‑up requests, and the agent is more likely to stick to the directive.

4. not every directive needs a thorough prompt.

when you get stuck (or lazy), sometimes “make it better” yields unexpected wins over a 1 000‑word context dump.

when in doubt, use the socratic method with reasoning models—start with a question (i.e. "how would you approach this?", "what do you think?").

or ask it to ask you questions (i.e. "what do you need from me to feel 100% confident you can build this?").

research shows iterative Q‑and‑A elicits more precise output and reduces hallucinations.

5. if you're cost-sensitive, you can turn a cheaper non-reasoning model (i.e. gpt 4o, sonnet 3.5) into a reasoning model by ending your prompt with "think step by step..."

6.  when something goes wrong, revert your changes and update the original prompt with additional context addressing the issue you observed.

it’s always cheaper (time and money) to restart than brute force prompting.

but if you insist on trying to fix an error, always ask it to determine the underlying root cause:

"trace the full user flow and identify where the issue originates. think deeply about the **root cause** and treat the symptom as a clue, not the target."

7. fill the gaps by asking the model to reflect on its work. after a task is completed, ask it to criticize itself.

here's a great prompt i stole from @shawmakesmagic:

"Please write a SCATHING review of my codebase -- everything that's wrong with it, in what ways it is badly implemented, larp code, stubs, demo and not production ready, etc. Catalog every weakness and explain why it's bad. Then write a detailed technical implementation plan on how to fix it."

and then ask it to implement its solution.

it's best to do this no more than once or twice before it starts inventing critiques.

8. know every model's strengths.

here's where i've personally found the most success:

> o3 = planning, task breakdowns and debugging
> sonnet 4 = specific, precise execution tasks
> gemini 2.5 pro = when you have wireframes or mockups, also for specific execution tasks
> opus 4 = large tasks or refactors

don't be afraid to switch models or ask different models attempt the same task.

9. know every platform's strengths.

every vibe coding platform has a highly researched, battle-tested system prompt that shapes the models behavior more than your own prompts.

here's where i've personally found the most success:

> lovable/v0 = reliable, consistent front-ends
> replit = python apps and backends fast (also the best vibe coding mobile app)
> cursor/windsurf = good for everything, but best for using MCPs and tool calling
> claude code = when i'm starting with or refactoring an existing codebase

you can also push a project to github and switch from one platform to another depending on what your current task is (i.e. create a working prototype in lovable and refine the back-end in claude code).

10. before asking the model to do anything, present your task and ask it to decompose it.

if you ask for a task outright, it's more likely to start working and lose track of the directive. instead, try this:

"before starting this task, break it down into it's smallest components and list out the sub-tasks."

the list it outputs will give it a to-do list that it can refer back to as it progresses and lowers the cost.

11. if you know what you want, you're guaranteed a performance lift by giving it examples or code snippets (or asking it to find examples before starting).

research shows examples boost performance more than fine-tuning.

whether its API documentation, component functionality, code structure and style, response formats, etc.

one example is good. many examples is better.

12. when giving context, references or examples, you should speak the language of LLMs.

most coding models are most comfortable using the formats they're trained on.

you can boost the model's understanding by providing documentation in markdown and presenting codebases in xml.

use tools like gitingest to convert github repositories into markdown or repomix to convert to xml. always look for llms.txt from api providers.

13. give the same prompt to every model to find the "majority vote" for a solution.

once you have multiple code outputs from each model, find where the most overlap is and commit to that direction.

this could be reviewing file diffs, a visual vibe check for UI output, asking different reasoning models how they would fix a bug, etc.

This app straight scam: they show you have some free votes but when you making votes they charging your USDC balance 🛑🚧🚨🚨🚨🚨

catsbeingweird

My cat every time I on the toilet. Is she guarding me?

enjoyart

world

the immortal blades from vermis by jeanette stals

Farcaster

🎁 I just got a reward from Monad Box! Try your luck and earn rewards too!

This crap is interesting but not working at All 😭

Bitcoin's positive price movements is meaningless nowadays

fc-updates

We’re planning to launch a new, optional “Pro” protocol subscription

1. At launch, subscribers will have a Pro badge on their profile

2. Subscription will be onchain and part of the protocol. 100% of the proceeds go to casters / developers in weekly rewards. This in addition to the existing $25K+ a week in rewards. 

3. Planning to add longer casts (10K characters) and up to 4 images per cast for pro subscribers (needs a protocol change). Additional features over time.

4. $120 a year / $10 a month 
Why are we doing this?

- The protocol needs to find long-term revenue source that can continue fund caster and developer rewards beyond any one company.  
- We originally thought storage fees could be that, but we’ve realized that storage fees should be essentialy free to help growth. 
- So we’re trying an opt-in pro subscription as another model

Why now?

- Figuring out a long-term revenue model that works will allow us to double down on investing in growth

Solana is low tier trash system that is about to save crypto. What a world to live in.

building farcaster infra @ /neynar 🪐 | casting @ /rish

writing about products & protocols on Farcaster | prev. Juicebox, Nouns, RAW DAO | casting about coffee, hi-fi, and other rabbit holes

product lead apps, social, and creators @base

Dude@ Warpcast is a shitpit at its best. It's so fragmented and userunfriendly. They impelenting things that doesn't make sense to vast audience making only 1% core users rich/alpha-valued. 

This wallet thing: I have the address but how I will know what are the use cases / how can I operate with it? No discord. No direct communication to the team members. It's shards of information devided by more sharding that you can dig by spending hourds to understand.

What is the point of this place?

@dwr.eth @jesse.base.eth @jacek @matthewb @aneri.base.eth @rish

gifART

/gifart seams to be restricted to some users, lemme know if you want me to invite you!

The simplicity and coolness are outrageous!

Warpcast

Why almost no one able to but warps because of the "update e-mail" error and how this can be fixed?

I am the stone the builder refused.
                    The Hat Stays On 😤 ~ Certified Degentlemen /sota Cofounder

many things:

• algo improvements (especially ones that help new users)
• pinned casts
• more users
• other clients or lite clients (including indices of frames/apps/creators)
• community efforts to surface quality accounts + content + frames
• spam filtering that doesn't silence new users by default

Read this and vote for real things when you can if you wanna see @farcaster improved

A sufficiently decentralized social network. farcaster.xyz

Design

Paragraph

I just subscribed to @antimofm.eth on /paragraph! Check it out:

Tonkeeper is such a trash. Like the worst ever experience for user.

coffee

https://zora.co/collect/zora:0xac0a7393d3414f60ce0a8c7384cfa146f0ecb1f9/1 👁️ 👁️ 👁️

I still don't understand how people trust their wallets to apps here. every interaction can be the last one before your funds get stolen.

is warpcast still even alive? let me know. Whole month passed since I posted anything.

moonhattan

@moonhattan