Vinicius Oike
@viniciusoike.bsky.social
Economist and Data Scientist | https://restateinsight.com/blog
São Paulo, Brazil
São Paulo, Brazil
One strategy I'm using is to ask ClaudeCode to write a small summary of what it's done + add context for future work when it's almost done with a task. I then use /clear, ask Claude to read the document, and give it instructions on the next task.
September 17, 2025 at 8:03 PM
One strategy I'm using is to ask ClaudeCode to write a small summary of what it's done + add context for future work when it's almost done with a task. I then use /clear, ask Claude to read the document, and give it instructions on the next task.
4. Avoid auto-compacting
The last point is directly tied to this one. ClaudeCode has a context limit (the UI warns when approaching this limit) and it auto-compacts by default. The quality of compacting can sometimes be unreliable and I found it best to just workaround it.
The last point is directly tied to this one. ClaudeCode has a context limit (the UI warns when approaching this limit) and it auto-compacts by default. The quality of compacting can sometimes be unreliable and I found it best to just workaround it.
September 17, 2025 at 7:58 PM
4. Avoid auto-compacting
The last point is directly tied to this one. ClaudeCode has a context limit (the UI warns when approaching this limit) and it auto-compacts by default. The quality of compacting can sometimes be unreliable and I found it best to just workaround it.
The last point is directly tied to this one. ClaudeCode has a context limit (the UI warns when approaching this limit) and it auto-compacts by default. The quality of compacting can sometimes be unreliable and I found it best to just workaround it.
You can ask for ClaudeCode to take notes of useful information and store it in separate md files to help build this knowledge.
September 17, 2025 at 7:57 PM
You can ask for ClaudeCode to take notes of useful information and store it in separate md files to help build this knowledge.
A great example of useful context to provide Claude with (from @sarahjohnson.io) gist.github.com/sj-io/3828d6...
Claude R Tidyverse Expert
Claude R Tidyverse Expert. GitHub Gist: instantly share code, notes, and snippets.
gist.github.com
September 17, 2025 at 7:55 PM
A great example of useful context to provide Claude with (from @sarahjohnson.io) gist.github.com/sj-io/3828d6...
3. Knowledge
Anthropic recommends that you create a CLAUDE.md file to store important context. Currently, I'm experimenting with an entire claude/ directory stashed with useful content. Before starting a new task I ask Claude to read the relevant content.
Anthropic recommends that you create a CLAUDE.md file to store important context. Currently, I'm experimenting with an entire claude/ directory stashed with useful content. Before starting a new task I ask Claude to read the relevant content.
September 17, 2025 at 7:53 PM
3. Knowledge
Anthropic recommends that you create a CLAUDE.md file to store important context. Currently, I'm experimenting with an entire claude/ directory stashed with useful content. Before starting a new task I ask Claude to read the relevant content.
Anthropic recommends that you create a CLAUDE.md file to store important context. Currently, I'm experimenting with an entire claude/ directory stashed with useful content. Before starting a new task I ask Claude to read the relevant content.
Ex: I asked ClaudeCode to download an Excel file, import it in R, and make it tidy. The download, however, failed due to network issues and Claude decided to invent a dataset with fake values instead of stopping and pointing this error.
September 17, 2025 at 7:51 PM
Ex: I asked ClaudeCode to download an Excel file, import it in R, and make it tidy. The download, however, failed due to network issues and Claude decided to invent a dataset with fake values instead of stopping and pointing this error.
2. Manually approve
While ClaudeCode seems magical when it "just works" it often doesn't. Auto-accept can lead to errors that hard to detect since Claude can sometimes omit them and say that everything worked according to plan.
While ClaudeCode seems magical when it "just works" it often doesn't. Auto-accept can lead to errors that hard to detect since Claude can sometimes omit them and say that everything worked according to plan.
September 17, 2025 at 7:47 PM
2. Manually approve
While ClaudeCode seems magical when it "just works" it often doesn't. Auto-accept can lead to errors that hard to detect since Claude can sometimes omit them and say that everything worked according to plan.
While ClaudeCode seems magical when it "just works" it often doesn't. Auto-accept can lead to errors that hard to detect since Claude can sometimes omit them and say that everything worked according to plan.
1. Make plans
Always start a new task in plan mode and invest some time making sure that Claude has a solid plan. A neat tip from Anthropic is to use "think more...", "think hard about potential issues...", "think step by step ... ".
"think" < "think hard" < "think harder" < "ultrathink."
Always start a new task in plan mode and invest some time making sure that Claude has a solid plan. A neat tip from Anthropic is to use "think more...", "think hard about potential issues...", "think step by step ... ".
"think" < "think hard" < "think harder" < "ultrathink."
September 17, 2025 at 7:45 PM
1. Make plans
Always start a new task in plan mode and invest some time making sure that Claude has a solid plan. A neat tip from Anthropic is to use "think more...", "think hard about potential issues...", "think step by step ... ".
"think" < "think hard" < "think harder" < "ultrathink."
Always start a new task in plan mode and invest some time making sure that Claude has a solid plan. A neat tip from Anthropic is to use "think more...", "think hard about potential issues...", "think step by step ... ".
"think" < "think hard" < "think harder" < "ultrathink."
Worst of all, since the numbers were generated with *data* they were heralded as an objective fact of the universe.
a computer monitor displays the message data does n't lie
ALT: a computer monitor displays the message data does n't lie
media.tenor.com
August 21, 2025 at 2:41 PM
Worst of all, since the numbers were generated with *data* they were heralded as an objective fact of the universe.
There's certainly some latin to know in economics: ex ante, ex post, ad hoc, per capita, ceteris paribus, de gustibus non est disputandum...
August 20, 2025 at 6:26 PM
There's certainly some latin to know in economics: ex ante, ex post, ad hoc, per capita, ceteris paribus, de gustibus non est disputandum...
I recently made this switch. Took me about 2-3 weeks to get adjusted but I've been enjoying Positron a lot. Highly recommend reading this positron.posit.co/migrate-rstu...
Overview – Positron
positron.posit.co
August 20, 2025 at 6:11 PM
I recently made this switch. Took me about 2-3 weeks to get adjusted but I've been enjoying Positron a lot. Highly recommend reading this positron.posit.co/migrate-rstu...