Five habits to avoid Claude and Codex usage limits plus 3 advanced tips

·Jul 15, 2026 11:58 PM·

Here's the video I promised. Five habits that keep me from hitting Claude/Codex limits (plus 3 bonuses for advanced builders), kept it quick and unpolished on purpose so I could actually ship it. Would love to know: which of these are you already doing, and which one are you going to try first? Kim W., Jared R., Akshayata S., Marisa S. thank you for saying usage limits are a real problem, this one's for you.

❤️3

5 comments

· Sorted by Oldest

Ozan D.
·
·
the usage efficiency angle is underrated. we see the same pattern in our AI skills data at aisa.to — the highest-scoring users aren't the ones who use AI the most, they're the ones who've figured out when NOT to use it and how to get max value from each interaction. constraint-driven habits beat unlimited access every time
1
Sana C.
·
·
Here you go Sussy C.
❤️1
🙌1
Artur M.
·
(edited)·
Sana C. thanks here are the cliff notes • Choose API Use Over MCPs: Connect only to essential Model Context Protocol (MCP) servers and execute everything else through APIs using tools like the GTM docs registry for updated documentation. • Avoid Waiting for Auto-Compacting: Initiate a dated-and-timed context handoff before reaching ~80% token usage to prevent context drift and hitting limits. • Create a Context Handoff Skill: Build a dedicated skill that instructs the AI to automatically include critical decisions and learnings during handoffs. • Reduce File Bloat in Context Index Files: Keep root files (like claude.md) lightweight by using them strictly as indexes to link out to specific protocol files. • Automate and Manually Review Context File Size: Implement an automatic agent to trigger bloat reduction when files hit size limits, and manually audit index files every couple of weeks. • Select Model Combinations for Sub-Agents: Explicitly instruct the AI to use the cheapest and most efficient model combinations (e.g., Sonnet or Haiku instead of Opus) when spawning sub-agents. • Use the Ponytail Skill for Lean Coding: Install specialized prompts/skills that force the AI to write the absolute minimum number of code lines required for a task. • Review Business Canonical Layers: Regularly audit and compress business canonical files in your second brain to prevent them from maxing out character limits.
Artur M.
·
·
fyi a lot of these issues are out-of-the-box handled by Hermes 🤓
Sana C.
·
·
Artur M. thanks. From my outbound perspective--I've seen Hermes being interesting after workflows have been setup for orchestration on improving agents and testing campaigns, reply rates etc. But the actual ideation, setup & orchestration, testing itself which is much of a WIP until breakout campaigns how could that work?

Ozan D.
·
·
the usage efficiency angle is underrated. we see the same pattern in our AI skills data at aisa.to — the highest-scoring users aren't the ones who use AI the most, they're the ones who've figured out when NOT to use it and how to get max value from each interaction. constraint-driven habits beat unlimited access every time
1
Sana C.
·
·
Here you go Sussy C.
❤️1
🙌1
Artur M.
·
(edited)·
Sana C. thanks here are the cliff notes • Choose API Use Over MCPs: Connect only to essential Model Context Protocol (MCP) servers and execute everything else through APIs using tools like the GTM docs registry for updated documentation. • Avoid Waiting for Auto-Compacting: Initiate a dated-and-timed context handoff before reaching ~80% token usage to prevent context drift and hitting limits. • Create a Context Handoff Skill: Build a dedicated skill that instructs the AI to automatically include critical decisions and learnings during handoffs. • Reduce File Bloat in Context Index Files: Keep root files (like claude.md) lightweight by using them strictly as indexes to link out to specific protocol files. • Automate and Manually Review Context File Size: Implement an automatic agent to trigger bloat reduction when files hit size limits, and manually audit index files every couple of weeks. • Select Model Combinations for Sub-Agents: Explicitly instruct the AI to use the cheapest and most efficient model combinations (e.g., Sonnet or Haiku instead of Opus) when spawning sub-agents. • Use the Ponytail Skill for Lean Coding: Install specialized prompts/skills that force the AI to write the absolute minimum number of code lines required for a task. • Review Business Canonical Layers: Regularly audit and compress business canonical files in your second brain to prevent them from maxing out character limits.
Artur M.
·
·
fyi a lot of these issues are out-of-the-box handled by Hermes 🤓
Sana C.
·
·
Artur M. thanks. From my outbound perspective--I've seen Hermes being interesting after workflows have been setup for orchestration on improving agents and testing campaigns, reply rates etc. But the actual ideation, setup & orchestration, testing itself which is much of a WIP until breakout campaigns how could that work?