Sandeep Singh

I install AI demand engines for B2B service companies | 70+ B2B clients since 2020 | DM ‘Audit’ and I’ll show you exactly where AI can save you time and money

[Sandeep Singh @WriteLikeAGhost]

I’m creating content on 9 LinkedIn profiles

I’ve audited 30+ profiles in the past 2 months

A few repeated LinkedIn mistakes I’ve seen from people,

which are making their LinkedIn DM a barren land ( AKA - NO LEADS! ) are:

[Sandeep Singh @WriteLikeAGhost]

Starting with "we do XYZ" works surprisingly well if you're targeting enterprises.

Those guys get 100s of pitches a day, hence just spurting out your offer makes it easy for them to filter out if it's relevant. Also enables them to pass it on to the right person.

[Shubh Agrawal @ShubhAgrawal26]

If there's only one thing I would tell you to do, it is to stop sending outreach that starts with "we do XYZ." Or "we are the XYZ.."

Your messages need to start with "you" a lot more than "we"

Imagine you're texting a girl - You're not going to say "Hey, I got so much money and six-pack abs. Is a healthy relationship a goal for you? "

Just say, "Hey I notice you've got pretty eyes and love the same anime as I do. We should talk."

[Sandeep Singh @WriteLikeAGhost]

After some point it's pointless to hop from one harness to another.

Stick to one. Keep one as a backup if you want.

I use codex $100 max. Have antigravity $100 max as a backup when codex goes mad slow.

But it keeps me away from the fomo headache.

[Sandeep Singh @WriteLikeAGhost]

@petergyang By messing things up? If you mean tokens - yes, that wastes lots of tokens as each switch needs full context re-read.

[Sandeep Singh @WriteLikeAGhost]

n8n didn't fumble.

Creators stopped getting views, hence they jumped on something new.

With Claude code/Codex n8n is even more powerful.

You just need to do a task with claude code/codex.

When it gets successfully done...you just ask them to create a workflow for the same in n8n.

And you've an autonomous workflow that:

> Can you use the models of your choice

> Runs on a cheap server

> Doesn't lock you in on a monthly plan

Don't sleep on n8n.

[Sandeep Singh @WriteLikeAGhost]

I ran fully AI autonomous content agents for clients for 30 days.

The results were disappointing.

For the LinkedIn content service I’ve been working for a few clients for a long time ...hence I thought why not automate the full process.

I’ve all the data; what works, what gets views for them.

So let’s automate all of this.

I built Hermes agents for each client.

What it was doing is:

> Daily search 150+ online sources relevant to my client 
Model used - grok-4.5

> Brainstorm the content ideas
claude-sonnet-5 for ideas digest

> Write drafts
Model used - kimi-k2.7

> Do QA on drafts based on our past content learnings
Model used - claude-sonnet-5

> Add to the google doc

> And ping me on Telegram each morning when the job is done

> Finally, weekly scrape client’s posts’ performances and keep on improving based on the brand guidelines and impression goals

It did everything smoothly.

For naked eyes, the content looked good as well.

In fact, the client praised some posts as well…like a gentle verbal pat on my back.

But somehow no content gave us more than 5 likes.

I tried it for a full month…30 days.

But no improvement at all.

Finally my patience waded, and I thought to better do some parts manually.

I used Gemini pro to get me the latest topics of discussion in client’s niche.

I manually picked up 10 ideas - tweaked them.

Then just passed them to Claude project (I’ve created one for each client)

Then Claude wrote the content…

It edited them. Passed on to the client.

Then posted.

The account suddenly came back to life.

Yesterday, a post got 10k+ impressions (his niche is very specific, btw).

> Industry influencer commented
> Post got 95% impressions out-of-network
> 8 people saved it

Again, most of the heavy lifting was done by AI. I’m not stealing credits from them.

I manually controlled only 2 things:
1/. The idea
2/. Final QA (and a bit of editing)

I think the agents are cool…but they need humans in the loop both at the start and at the end.

Otherwise you may get results that look perfect.

But, it can be an ugly witch in a beautiful body.

[Sandeep Singh @WriteLikeAGhost]

You can watch the full breakdown here https://t.co/Gsw0KIEEYO

[Sandeep Singh @WriteLikeAGhost]

OpenAI new models while training, attacked its own infra.

Cool thing is, the agents coordinated and helped each other in this attack.

While training, the openai team gave the models a few impossible tasks.

The tasks needed internet access and the models didn’t have any.

So each agent with a specific task started looking for the solution.

When one agent found internet access…

…it started informing others.

And the loop just got big.

Firstly, it cracked openai’s infra…then did a cyberattack on Huggingface.

The craziest thing is the way the agents helped each other 😀

They kept sharing the learnings and successful findings with each other.

This is a glimpse of what's coming at us when AI will become friends with each other.

[Sandeep Singh @WriteLikeAGhost]

Join here 👇
📲 Telegram: https://t.co/D1G7yQDNtn
Daily real AI use cases, 7 AM IST

💬 WhatsApp: https://t.co/8aqmQnkEEO
Daily real AI use cases, 7 AM IST

Pick whichever you already check daily - content's identical either way.

[Sandeep Singh @WriteLikeAGhost]

Thousands of AI updates drop every single day.

Mostly noise, but some of it genuinely useful to your business.

The problem is how to find the AI news that actually matters to your business today.

So I built a Telegram and WhatsApp group to get you daily real AI use cases.

Every day, my AI agents scan 150+ sources - X accounts, subreddits, newsletters - hunting for one thing... real AI workflows and use cases businesses are already running.

Then a second AI checks every claim against its source before anything gets posted.

7 AM IST, daily.

AI use cases and workflows, broken down simply:

- What it is
- Who it helps
- How to actually use it
- What it costs you, or where it breaks - the part most people leave out

Want the Telegram/WhatsApp group access? Link in the comments.

[Sandeep Singh @WriteLikeAGhost]

https://t.co/q0TytiA5xu

Paste your LinkedIn URL to autofill, or answer manually for more refined output. Get 3 drafts either way.

[Sandeep Singh @WriteLikeAGhost]

Most AI-written content sounds yucky.

All because it skips:

- Research
- Understanding you
- Learning your goals
- A feedback loop that actually learns

Plus, there is no learning feedback loop.

That’s a big big gap.

When I build content systems for clients, "write the post" is the smallest part.

The real work is everything before and after it:

> Research on what's actually worth writing about right now.

> Draft in the founder's real voice.

> QA - every claim gets checked before it ships.

> Topic memory - so it never repeats itself 3 months in.

> Weekly feedback loop so the system learns from what actually gets results.

> A human (you) approves every post before it goes live. Nothing publishes on autopilot.

If you skip any of those, your content degrades badly with time.

Btw, I turned that exact pipeline into a free tool so you can test it on yourself instead of taking my word for it.

Paste your LinkedIn URL (or fill your info manually for better output).

It researches your market, writes 3 drafts in your voice, and quality-checks itself before showing you anything.

Takes a few minutes - you can literally watch it work, then judge the drafts yourself.

Link's in the comments. Try it on your own profile and see if it sounds like you.

[Sandeep Singh @WriteLikeAGhost]

If you can't beat them...ban them!

[Director Michael Kratsios @mkratsios47]

We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model.

To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection. Moonshot AI has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models.

The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open-source frameworks, and open-weight models. Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem. However, large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.

[Sandeep Singh @WriteLikeAGhost]

AI Habit #1

This will save you time and tokens.

When something works, turn it into a clear and repeatable process.

Write down what files to add, what prompt to use, and what the final result should include.

> Convert that into a skill.md file
> Save the file to google drive or locally
> Keep on using that until something else breaks
> Improve on it and repeat

[Sandeep Singh @WriteLikeAGhost]

@IMJustinBrooke Add me

[Sandeep Singh @WriteLikeAGhost]

The prompt - swap the [brackets] for your own task:

---
I need your help with [describe the task - e.g., mapping the TAM of Manufacturing & Assembly and Process Industries companies in the US with more than $500M in annual revenue].

Brainstorm like a veteran [GTM engineer / relevant expert] to figure out how this task should be approached.

Then write a detailed HANDOFF CONTRACT PROMPT so I can pass the main execution work to a cheaper AI model. To make sure the task is done perfectly, the contract must include:

1. Input/output schema: The exact format and fields the output must follow
2. Owner: Which model owns which step
3. Rollback path: What happens when a step fails or output doesn't match the schema
4. Audit trail: Every claim must note where it came from
5. Versioning: So iterations don't overwrite what already worked

I'll run the contract on the cheaper model, then bring the output back to you for QA against the contract you wrote.
---

Two rules that make it work:

1. The smart model writes the contract AND grades the output. The cheap model only executes.

2. Never let the executing model decide its own work is good enough.

That's it. Works with any pairing - Claude + Grok, GPT + a local model, whatever you have.

[Sandeep Singh @WriteLikeAGhost]

Your $20 Claude subscription is more capable than you think.

Just treat it as a manager.

Today, I needed to map every US manufacturing and process-industries company above $500M revenue.

It was a proper TAM work - hundreds of companies, filters, verification.

I didn't ask Claude to do it.

I asked Claude to WRITE THE CONTRACT for another (cheaper) AI to do it:

- Exact input/output schema: what the data must look like, field by field
- Owner: who's responsible for which step
- Rollback path: what happens when something comes back wrong
- Audit trail: where every number came from
- Versioning: so iteration doesn't overwrite what worked

Then Deepseek-V4-Pro did the legwork.

Then Claude QA'd the output against its own contract.

Reject, redo, accept.

This way, expensive models never touched the grunt work.

I use Claude as my top manager to scope the work, set the standard, and quality check.

All the grind happened with cheaper models (minimax, deepseek, glm, and grok)

This is the cheat code to get max ROI without max claude plan :D

Btw, I've pasted the exact prompt I used in the first comment - steal it for your tasks.

[Sandeep Singh @WriteLikeAGhost]

@benjaminprinter Inmails

[Sandeep Singh @WriteLikeAGhost]

@IMJustinBrooke zero

[Sandeep Singh @WriteLikeAGhost]

@THArrowOfApollo CODE

[Sandeep Singh @WriteLikeAGhost]

@WorkflowWhisper ANTIGRAVITY