Mohabbat

AI × blockchain × security × finance.
Software engineer → Agentic engineer.
I make AI write clean code I am proud to ship.

[Mohabbat @HaMohX]

#reset

[Mohabbat @HaMohX]

This is what you can achieve without loop engineering!

[Mohabbat @HaMohX]

Nothing worse than burning your whole $200 Claude limit by Sunday and waiting until Friday to get it back.

Happened to me 3 times, five days locked out. of course I was not happy. so I had to do something about it.

Two fixes, and it feels so much better with no drop in quality I can notice:

set Opus 4.8 to medium effort
turned off all memory (mem-claude and Claude's own)

[Mohabbat @HaMohX]

https://t.co/NtmwHsX3dn

[Mohabbat @HaMohX]

@marclou wrote "32 Principles of a Viral Product."
I turned all 32 into a Claude skill that audits your SaaS against them.
The link in the first comment
https://t.co/bo4ZbTwNM0

[Marc Lou @marclou]

https://t.co/Yju8RrWuOV

[Mohabbat @HaMohX]

DRY? LLMs do not care about it by default.

“Don’t repeat yourself” is a simple principle from The Pragmatic Programmer

by Andy Hunt and Dave Thomas.

The idea is clear:

Do not write the same logic twice.

Do not copy code.

Reuse what already exists.

But AI often fails here.

You can add “respect DRY” in your claude.md, rules file, or coding skill.

That helps.

But it does not solve the problem.

AI is not lazy by nature. But it often takes the path of least effort. In code, that often means:

It creates a new function.

It writes similar logic again.

It misses existing helpers.

It adds more code instead of using what is already there.

Example:

You ask Claude to build a feature.

Most of the needed functions already exist in the codebase.

But the LLM does not always check all files deeply. It may find one or two useful functions, then rewrite the rest.

That is how clean code slowly becomes spaghetti code.

So what is the solution?

Human control.

AI output needs review.

A rule in a markdown file is not enough because:

The tool may not load the rule.

The rule may get lost lost in long context.

The model may ignore it under task pressure.

The model may not search the codebase enough.

So you need to ask direct questions:

How did you implement this?

Did you apply the DRY principle?

Which existing functions did you reuse?

Did you search the codebase before writing new logic?

Which new code is really needed?

The best AI coding workflow is not:

“Write the code and trust it.”

It is:

“Write the code, explain your choices, prove reuse, then let me review.”

AI can write fast.

But humans still need to control quality.

[Mohabbat @HaMohX]

Stop Claude from eating your subscription.

Use these projects to cut token usage by 50% or more.

Headroom -> https://t.co/eicp155cim
Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 60-95% fewer tokens, same answers.

caveman -> https://t.co/CWbekfR3Qn

Forces shorter, direct replies. Useful when output tokens are the leak. Around 75% fewer output tokens without losing technical accuracy.

context-mode -> https://t.co/pzoymj4pdr

Keeps raw tool data out of the context window. Its examples show 315 KB becoming 5.4 KB, about a 98% reduction.

rtk -> https://t.co/JLTBIyyfIK

A CLI proxy that filters command output before it reaches the model. Useful for cutting noisy terminal output. 60% to 90% token savings.

semble -> https://t.co/WeL7Dh0X30

Code search for agents. Instead of grepping and reading everything, it returns only the useful code snippets. Around 98% fewer tokens than grep-based workflows.

code-review-graph -> https://t.co/fOVo9Qnesp

Builds a knowledge graph for code review so agents can see how files, functions, and changes connect before pulling in extra context.

codeburn -> https://t.co/SnxXRRCN5n

Bonus: tracks token usage, cost, model, tool, project, and provider. Runs locally and reads session data from disk.

Tokens are budget. Context is infrastructure.

Stop feeding Claude everything.

Feed it what matters.

[Mohabbat @HaMohX]

Gem article 💎

[jessy @13yearoldvc]

https://t.co/DGdLu79krs

[Mohabbat @HaMohX]

Do not ask: "How much can AI automate?"

Ask: "Which parts should AI do, and which parts must humans check?"

Simple rule:

Low-risk task + easy to check = more AI automation

High-risk task + hard to check = more human control

Low-risk coding tasks: AI can do more

Examples:

-> Write a small helper function

-> Rename variables

-> Generate test data

-> Format code

-> Add simple comments

-> Create boilerplate code

Human check is easy, so AI can take a higher role.

High-risk coding tasks: human must stay in control

Examples:

-> Write authentication logic

-> Change smart contract withdrawal rules

-> Edit payment code

-> Handle private keys

-> Change validator signing logic

-> Modify access control

-> Design database permissions

Here, one small AI mistake can create a real security issue.

Practical AI coding workflow

-> Human defines task and rules

-> AI writes code draft

-> AI explains why it made the change

-> Human reviews risk areas

-> Tests and security checks run

-> Human approves merge