Claude Token Optimization

The need

I use Claude every day for long pieces of work, and I noticed that the same result can cost a lot or very little depending on how you work: which model you pick for each step, how much material gets read, when it pays to start over in a new chat. There is plenty of advice on these topics, almost never backed by a test.

I wrote down the rules I was using, checking them against my daily work, and collected them in a skill: instructions Claude reads and applies on its own. The next step is opening them up to people who use Claude in other ways, because a rule born from one person's use needs to be tested by others.

Constraints

The rules only count if answer quality stays the same: savings that make the result worse don't count. Each proposal enters the next version only after it has been tested, credited to the person who suggested it. No personal, company or account data, and language that makes sense to people who don't write code. The material is public under the MIT licence. The project is independent and is not affiliated with or endorsed by Anthropic.

Forks in the road

1. Choice

   The rules live in a shared skill and in a version to paste into custom instructions, so people who don't install skills can use them too.

   Discarded alternatives

   A list of tips in an article; a single prompt to copy.

   Reason

   A skill applies itself in every chat; an article gets read once and forgotten. The paste-in version keeps people who use Claude without skills on board.

2. Choice

   A proposal becomes a rule only after a test in real work. Measurements help but aren't required: a clear description is enough to start.

   Discarded alternatives

   Accepting proposals by vote or by conviction; asking everyone for full measurements.

   Reason

   Without a test, a rule is just an opinion. Demanding numbers from everyone would shut out people with a good idea but no time to measure it.

3. Choice

   Anyone who wants to contribute is asked for ideas and experience, never prompts, files or conversations.

   Discarded alternatives

   Collecting real chat examples.

   Reason

   Conversations contain data about people and companies, and ideas are enough to improve the rules.