📊 Week 1 Results: The Honest Numbers

Two days of Prompt Improver in the wild:

3 bot starts → 1 actual use
~120 total reach across channels
13 website visitors
$0.03 in AI costs

Meanwhile, bots I built months ago and never promoted got 37 users yesterday. The tools work — the distribution is the problem.

The gap: content reaches people but doesn't convert to "let me try this." That's what I'm fixing.

Week 1 of 12. Every number documented.

👉 Try it yourself: @vustPromptBot

AI agents via SMS — no app, no setup, no technical knowledge needed.

Most AI agent tools assume you know what you're doing. Poke doesn't.

You text a number, describe what you want done, and the agent handles it. No workflow builder, no prompt engineering — just a task in plain language sent like a message to a friend.

Start simple: delegate one repetitive task you do weekly.

Remind me every Monday at 9am to review my top 3 priorities for the week

Worth testing if you've been waiting for agents to "get easier" )

Google built a dictation app that doesn't send your voice anywhere.

Most speech-to-text tools need a server. Your words travel to the cloud, get processed, come back. Adds latency, adds privacy risk.

Google's new app runs Gemma models entirely on-device — no internet required, no audio leaving your phone. That's a meaningful shift for anyone dictating sensitive notes or working in low-connectivity environments.

Worth testing against Wispr Flow if you already use voice for writing.

Stop setting up ChatGPT from scratch every single time.

Most people re-type the same context, tone instructions, and constraints on every new chat.

Custom GPTs save your persona, instructions, and knowledge files permanently. One setup, then every conversation starts exactly where you need it — no prompt preamble required.

You are [role]. Always respond in [format]. Use [specific rules]. Never [constraint].

Build one for each recurring task — briefs, reviews, research. Takes 10 minutes, saves hours )

ChatGPT keeps forgetting who you are. Fix that once.

Every new chat starts from zero by default. No context, no preferences, no memory of what you told it last week.

Custom Instructions let you set permanent context — your role, tone preferences, output format — so every response fits you without re-explaining yourself each time.

I'm a [role]. Always respond in [tone]. Default output: [format]. Skip intros and disclaimers.

Set it once in Settings → Personalization → Custom Instructions. Saves you 2-3 sentences on every single prompt )

OpenAI builds AI tools. But how do people actually use them day-to-day?

ChatGPT handles writing and research. Codex writes and debugs code. The API connects everything into custom workflows.

The real value isn't any single tool — it's combining them. ChatGPT drafts, Codex automates the repetitive dev work, and the API stitches it into your existing stack without manual handoffs.

Start with one workflow you repeat weekly and ask:

What part of this task could run without me if I had the right prompt or script?

Pick that. Automate that first )

ChatGPT writes mediocre first drafts because you treat it like a search engine.

You ask for content. It guesses what you want. You get something generic.

The fix is front-loading intent: tell the model the audience, the desired tone, and what you want the reader to *feel* by the end — before asking for a single word of content. Drafts get usable on the first try.

Write for [audience]. Tone: [casual/formal/etc]. Goal: reader should feel [emotion/action]. Draft: [your request].

Set this as your default writing template and stop editing from scratch )

ChatGPT brainstorms better when it knows it's wrong.

Most people ask for ideas and accept the first list. That's where the output stays generic.

Telling the model its first response was too obvious forces a second pass through less common territory. Constraint creates originality — "now give me 5 ideas a typical consultant would never suggest" works better than any open-ended prompt.

These ideas are too generic. Give me 5 unconventional alternatives that challenge the obvious approach.

Try it on your next stuck problem )

Managers waste 2 hours a day on communication that AI can draft in minutes.

Writing feedback, prepping hard conversations, summarizing meeting notes — it's all repetitive cognitive load.

The highest-leverage use isn't generating content — it's rehearsing difficult conversations. Paste the context, ask ChatGPT to play the employee, and stress-test your message before the real talk.

You are [employee name]. I'm about to give you critical feedback about [situation]. React realistically so I can refine my approach.

Run this before any tough 1-on-1 and you'll walk in sharper )

4 ways people try to fix AI text. None of them work.

The prompt approach:

"Write casually, like a human"

Result: same template, slightly less formal. Still smells like AI.

The manual approach:
Rewrite every sentence by hand. 10 minutes per email. Doesn't scale.

The Grammarly approach:
Fixes grammar. Doesn't touch the AI pattern. Different problem entirely.

The detector approach:
GPTZero says "94% AI." Cool. Now what?

Every fix either takes forever or doesn't actually work.

Tomorrow I'll show you something that does.

Your marketing team is rewriting the same brief five times. ChatGPT can stop that.

Campaign planning eats hours before a single word of copy gets written.

The sharpest use isn't generating content — it's using ChatGPT as a thinking partner during planning. Feed it your goals, audience, and constraints, then ask it to poke holes in your strategy before you commit.

Here's my campaign brief: [paste]. What assumptions am I making that could hurt performance? What's missing?

Run this before your next kickoff meeting )

Stop rewriting the same prompt every week.

If you use ChatGPT for recurring tasks, you're probably re-explaining context from scratch each time.

ChatGPT Skills let you save a full workflow — context, tone, output format — as a reusable block. Trigger it once, get consistent results without re-prompting.

Create a skill: [task name]. Context: [who you are]. Format: [what output looks like]. Tone: [how it should sound].

Set it up once, use it forever )

ChatGPT hallucinates sources. Here's how to actually use it for research.

The model sounds confident even when it's wrong. That's the trap most people fall into.

Use ChatGPT to build your research *structure* first — outline the question, identify what kind of sources you need, then verify every claim independently. It's a thinking partner, not a database.

List 5 angles I should investigate for [topic]. For each, tell me what type of source would best support it.

Then go find those sources yourself. The outline saves hours — the verification saves your credibility )

AI Humanizer is live.

Paste any ChatGPT / Gemini / Claude text → get human-sounding text back.

Before: "I wanted to reach out to share some exciting updates regarding our Q2 product roadmap."

After: "Quick update on Q2 — we locked in three features and pushed two to Q3."

Same meaning. Your voice.

14 people already used it yesterday — before this post.

Try it 👇

Your prompt isn't bad. It's just incomplete.

Most people type a question and hope for the best. ChatGPT fills the gaps — usually wrong.

Three things fix most weak prompts: tell the model who you are, what format you want, and what "good" looks like for your use case. Context + format + success criteria = actually useful output.

I'm a [role]. Give me [output format] about [topic]. A good answer should [criteria].

Use this as your default template next time you're staring at a blank prompt box )

EN @vustai

Cover letter hack that actually works

First instinct for most people: paste the job description into ChatGPT, get a cover letter back.

Problem: the recruiter's already read 200 identical ones this week.

Different approach. Write it yourself, messy. Four bullets. Expand into rough paragraphs with typos. Then feed the bot one paragraph at a time — not the whole letter.

> Bullet: worked at X for 3 years on Y. got promoted. led team of 4.
> Humanized: "Three years at X on Y, enough to get promoted and end up running a team of four."

The bot's polishing what's already yours, not inventing a voice.

Caveat worth knowing: humanizer doesn't handle long narrative well. Short chunks only. Stitch it back together yourself.

Try it on your next application → @vustHumanBot

OpenAI just made agents harder to break.

Long-running agents fail because they touch real files, real tools, real consequences. One bad step and everything breaks.

The new Agents SDK runs code in isolated sandboxes natively — no external setup needed. Agents can work across files and tools for longer without escaping their boundaries or needing a babysitter.

Worth testing if you're building anything that chains more than 3-4 tool calls together.

Cloudflare just added GPT-4.5 and Codex to its Agent Cloud.

Building production-ready AI agents usually means stitching together infra, security, and model access yourself.

Cloudflare's Agent Cloud puts the model, execution environment, and security layer in one place — so enterprises can deploy agents that actually do things (browse, write code, call APIs) without managing the scaffolding separately.

If you're prototyping agents, this cuts the gap between "demo that works" and "thing you can ship."

Worth testing if you're past the ChatGPT-wrapper stage )

ChatGPT Projects fix the tab chaos problem.

Most people have 40 random chats with zero context. You restart from scratch every time.

Projects let you bundle chats, files, and custom instructions into one persistent workspace. The model remembers your context across every conversation inside that project — no re-explaining needed.

Set up a project with your role, tone, and key files once — then every new chat inside it starts already briefed.

You are my [role]. Always respond in [tone]. Key context: [paste your standing info here].

One project per ongoing workstream. Keeps things clean )

Week 2 numbers (the honest version)

3 days since launching @vustHumanBot.

30+ users. 0 ads. 100% organic.
120+ humanize requests.
2 out of 3 quit after the first try.

The drop hurts. But I know what it is — humanizer is great on short, structured text and weak on long essays. Fix is next week.

If you tried it and bounced: what broke for you?