The most effective skill I've been using

I have been using the LLMs from copy pasting chatgpt's output into the file then cursor, claude code and codex. They all shared traits of overcomplicating things, just augmenting fields and logic. As new models released, things improved, but not solved yet(though opus 5.5 is leap improvement).

It got really unbearable, when i had to review the PRs, the amount of diff didnt justify for the task clearly. This forced me to reverse engineer PRs diff to understand what is being tried to ship, because that huge diff is scary.

btw the changes are tested, it works, it does what author wanted, SURE, it does. Isn't it exactly how LLMs are trained?

Fine i won't rant, let me be bit professional(cough). So in one of the reverse engineering session, PR diff was too big to justify the task, so i thought let me throw this diff at the claude(Fable 5) and ask it to simplify, btw i didnt read the diff yet.

Prompts:

And impressively.

Pretty good result actually, there is one simplification that stuck with me.

So the PR had this pipeline where a worker polls a shared inbox and turns each customer email into a support message.

It decided to use inbox API to label emails, that removed the major plumbing, pretty cool. Duplicate email table was quite obvious cleanup.

If i can get this kinda result then i want to repeat this for every PR i get and obvious next step is to create a skill out of prompts from that session, i named it simplify-by-design-review.

I started to use it for every PR, though i didnt get any dramatic simplifications like i did for the first time.

Later, i added two more guiding instructions to address the issues observed with use.

Solve the right problem — the biggest simplification.

Understand the objective, the motivation, and the success criteria. What are we even solving here? Is this ACTUALLY the thing we intend to solve, or are we doing something that's not right? Half understanding the problem is more dangerous than not understanding it. If things are not clear, ask clarifying questions. If you cannot ask, state the assumption up front. Don't overfit the problem.

Once the requirements are crystal clear, question the product design thoroughly. Collect data, be data-based wherever possible.

("dangerous", lol)

This pushes the agent to think from scratch, instead of anchoring on the current plan/diff/context and question product decisions based on the data.(muaah)

How to think

Don't be lazy, don't work with assumptions or guesses — be thorough, be hard working.

dont always try to produce next possible answer, dont settle on the first thought, ponder, wonder, think through different ways, connect, mix and try non-standard, creative, mix and connect different dimension world domain ideas and concepts. Think ponder challenge your own ideas and seemingly possible answers and refine it

Second is supposed to push LLMs to produce THE answer instead of next possible answer.

I've been using this skill a lot. Now this has become quite generic skill and not particularly tied to PR diff. For instance, I use it to cut BS from drafted whatsapp message replies now.

The best result

I was working on a POC to validate an idea and to production if it works out. I can't share details but it was about data extraction. You had to understand patterns present in the raw data and write a logic to extract useful data.

Prototyped with GPT-Astra, after days of iterations, pipeline(model + prompt + architecture) reached some stability and performed well on unseen datasets. Harness had two tools, python_execute to run python code and submit_review to submit for final review.

I was preparing the code to be ready for the deployment, i wanted to clean up and simplify the built harness and flow, since GPT's like to overcomplicate things. By then i had Opus 5.5.

Sent

After 10 mins

I didn't expect this, we use Nodejs, in worst case i had to use Agents API because of the python_execute tool. Not only it got rid of python tools, it improved on the cost and time!! But in this case, i think its more of model than the skill, because right after this, i tried to get the similar result with GPT-6-Astra and tried multiple times couldnt get it. Opus just needed the right framework to think.

Though this is not a fair measure of anything, these results are not produced in an isolated setup with the same environment, though opus 5.5 is the best model at the time of writing.

Observation

I have been using simplify-by-design-review a lot, on PRs, plans, docs pretty much on everything that needs fresh view top-bottom, it helps to cut down the obvious bloat and sometimes cuts features before they're built.

I think it works best when you have built a system that works because it has rich details(code itself) and plumbing that you have tested and iterated, its stable, works. Inefficiencies become quite easy to identify once you have working system, getting the first inefficient system is like collecting data phase.