Uploaded October 2026 | Updated November 2026, 3 hours ago
Running every prompt on the biggest frontier model gets expensive fast. Here are five ways I cut token usage in my AI coding harness: pick a cheaper model for simple work, put "don't create extra files or tests" rules in your steering file, split spec-driven development so a cheaper model handles implementation, let ESLint, Prettier and hooks do the deterministic work, and keep an eye on your context window and MCP servers. I'm showing it in Kiro, but every tip works in Claude Code, Cursor, Copilot and the rest.
Also, I recorded this one outside. Let me know if you like the new background.
Did I miss a tip? Leave it in the comments.
Check out my last video on why frontend developers are cooked
youtu.be/nWJa28MS-gI?si=lHchZnQYpJyrzVR0
Follow me on socials!
X: https://x.com/ErikCH
LinkedIn: / erikhanchett
Bluesky : https://bsky.app/profile/erikch.bsky....
Website: programwitherik.com
https://www.kiro.dev - my favorite harness!
Join my newsletter and get updated when I release cool stuff -
programwitherik.com/newsletter
## Links
- Pushing Kiro's Free Tier to Its Limits (Laura Salinas): dev.to/aws/pushing-kiros-free-tier-to-its-limits-5e34
- Kiro models and credit multipliers: https://kiro.dev/docs/models/
- Kiro steering: https://kiro.dev/docs/steering/
- Kiro specs: https://kiro.dev/docs/specs/
- Kiro hooks: https://kiro.dev/docs/hooks/
## Timestamps
0:00 Token maxing is dead
0:14 Tip 1: Use cheaper models for simple tasks
0:37 Tip 2: Add token rules to your steering file
1:06 Tip 3: Spec-driven development with two models
1:36 Tip 4: Let linters and hooks do the work
2:09 Tip 5: Watch your context window and MCP servers
3:11 Wrap up
Running every prompt on the biggest frontier model gets expensive fast. Here are five ways I cut token usage in my AI coding harness: pick a cheaper model for simple work, put "don't create extra files or tests" rules in your steering file, split spec-driven development so a cheaper model handles implementation, let ESLint, Prettier and hooks do the deterministic work, and keep an eye on your context window and MCP servers. I'm showing it in Kiro, but every tip works in Claude Code, Cursor, Copilot and the rest.
Also, I recorded this one outside. Let me know if you like the new background.
Did I miss a tip? Leave it in the comments.
Check out my last video on why frontend developers are cooked
youtu.be/nWJa28MS-gI?si=lHchZnQYpJyrzVR0
Follow me on socials!
X: https://x.com/ErikCH
LinkedIn: / erikhanchett
Bluesky : https://bsky.app/profile/erikch.bsky....
Website: programwitherik.com
https://www.kiro.dev - my favorite harness!
Join my newsletter and get updated when I release cool stuff -
programwitherik.com/newsletter
## Links
- Pushing Kiro's Free Tier to Its Limits (Laura Salinas): dev.to/aws/pushing-kiros-free-tier-to-its-limits-5e34
- Kiro models and credit multipliers: https://kiro.dev/docs/models/
- Kiro steering: https://kiro.dev/docs/steering/
- Kiro specs: https://kiro.dev/docs/specs/
- Kiro hooks: https://kiro.dev/docs/hooks/
## Timestamps
0:00 Token maxing is dead
0:14 Tip 1: Use cheaper models for simple tasks
0:37 Tip 2: Add token rules to your steering file
1:06 Tip 3: Spec-driven development with two models
1:36 Tip 4: Let linters and hooks do the work
2:09 Tip 5: Watch your context window and MCP servers
3:11 Wrap up










