Specter👁 @0xSpecter_

Claude is like an employee who needs to be introduced to the company anew every morning. Skills are simply his personal business. Write down who you are and how you work once, and he stops asking. 20 such files later, and the person has an AI that remembers him better than he remembers himself on Monday morning.

Specter👁 @0xSpecter_

a month ago

x.com/i/article/2067…

1 7 37 7K 35

If you explain the same thing to Claude more than once, it's a signal to save it as a file (Skill) instead of repeating it again. One file = Claude automatically picks up the context in any new chat, without reminders. Works for writing style, code reviews, weekly reports-any recurring task.

Specter👁 @0xSpecter_

a month ago

x.com/i/article/2067…

1 7 37 7K 35

Every new chat with Claude began with the same style explanation-until it occurred to me to just write it down once The file is called a Skill. Claude automatically finds and applies it, without reminders 20 such files later, routine tasks like code reviews and weekly reports now take significantly less time

Specter👁 @0xSpecter_

a month ago

x.com/i/article/2067…

1 7 37 7K 35

he got tired of introducing himself to Claude same voice rules. same preferences. same don’t suggest this again. same weekly review format. every new chat started like the last one never happened. so he stopped treating memory like something the model should magically have. he wrote it down. one file for the rules. one file for the decisions. one file for the repeated workflows. that is the whole trick. Claude does not need to remember you in some mystical way. it needs a place to check before answering. the useful setup looks boring: - CLAUDE.md is how you want it to work. - MEMORY.md is what already happened. - SKILL.md is what it should do the same way every time. once those files exist, the conversation starts later in the relationship. Claude can see the old decisions. it can avoid the failed approach. it can reuse the weekly review format. it can stop asking you to explain your taste like you just met. the model did not become personal. the workspace did.

Specter👁 @0xSpecter_

a month ago

x.com/i/article/2067…

1 7 37 7K 35

POV: You spend three weeks explaining who you are to Claude, and then you find out you could have done it just once This file is called a Skill. Write down the style, tone, and rules once-and Claude never asks again. One guy already has 20 of them. Code reviews, weekly reports, analyzing client feedback-it all works automatically, without a single reminder. The funniest thing about this story isn't that AI can do this. It's that 90% of people still type the same thing over and over again and don't even realize there's a "stop" button

Specter👁 @0xSpecter_

a month ago

x.com/i/article/2067…

1 7 37 7K 35

One guy started every new chat for three weeks straight with the same line: "I blog, no corporate jargon, that's my style." Over and over again. Every time. On the fourth try, he stopped and thought-what if I just wrote it down once? Saved it to a file. Closed the chat. Opened a new one, without a single word of context. Claude found the file himself and applied the style, as if he'd always remembered it. Then he did the same with twenty recurring tasks-code review, weekly report, feedback analysis. The most interesting thing isn't about the AI's memory. It's about being the first to notice a pattern that everyone else simply tolerated in silence.

Specter👁 @0xSpecter_

a month ago

x.com/i/article/2067…

1 7 37 7K 35

Interesting case study in how a small habit saves real time: instead of re-explaining his style and context to Claude every session, the author wrote it down once as a file (a Skill) - and the model now picks it up automatically whenever the task matches. The most useful part isn't the trick itself, it's the diagnosis: when Claude gives inconsistent results on the same request, the problem usually isn't the model - it's a vague spec. Tighten the description and instructions, and the behavior stabilizes. And the one piece of advice that cuts against the hype: don't build twenty Skills in a weekend. One that actually works, for a task you've genuinely repeated, beats twenty half-finished ones.

Specter👁 @0xSpecter_

a month ago

x.com/i/article/2067…

1 7 37 7K 35

Three weeks. Same paragraph. Fourth time typing it from scratch into a new chat. Most people would just keep doing that forever and call it "the cost of using AI." One file fixed it permanently - and it didn't stop there. Twenty Skills later, the model stopped forgetting who he is, how he reviews code, and what Friday is supposed to look like. The pattern here isn't really about AI. It's about noticing what you repeat and refusing to repeat it a third time.

Specter👁 @0xSpecter_

a month ago

x.com/i/article/2067…

1 7 37 7K 35

@ScottyBeamIO For that kind of money, I'd like the agent to fly for me

I kept catching myself re-explaining the same things to Claude in every new chat - style, context, rules. Skills fix exactly that: write the instruction once into a SKILL.md file, and it loads itself from then on, no reminders needed. What's interesting here isn't the idea itself, but the breakdown of what goes wrong. A vague skill description ("review code like a senior engineer") produced different behavior every time on the same prompt. Once the trigger was narrowed to an actual phrase, the steps were spelled out, and the output format was pinned down, the behavior became consistent. So inconsistency isn't the model glitching - it's the instruction being underspecified. And the closing takeaway lands harder than the list of twenty ready-made skills: don't build a dozen in advance, just get one working properly - for whatever task you've had to explain more than twice this month.

Specter👁 @0xSpecter_

a month ago

x.com/i/article/2067…

1 7 37 7K 35

x.com/i/article/2067…

@N01ennn I'm glad you liked it!