Breaking

Community Codex stack of DualMode, Caveman, and ADHD skills claims up to 5x usage

September 1, 2026 at 04:41 EDT

Software Dev & Coding
AI Agents
Open Source

A community Codex setup stacking DualMode, Caveman, and ADHD-style skills is being promoted as a way to stretch usage by as much as 5x on ChatGPT plans. The DualMode piece is a small MIT skill that assigns execution to GPT-5.6 Luna and keeps GPT-5.6 Sol off the repository.

Sep 1, 2026 · Codex

Claimed 5x Codex usage from a skill stack

DualMode sends GPT-5.6 Luna to execute and GPT-5.6 Sol a 1–3K token packet. Caveman and ADHD skills then cut filler in the replies. The 5x figure is a community claim, not a published test of the stack.

Claimed usage stretch

1–3K

Tokens in a DualMode packet

65%

Caveman advertised output cut

8.5%

Measured cut on agentic tasks

Claimed stretch, drawn to scale

Five stacked blocks versus one. This is the community 5x claim, not a DualMode lab result.

Default

Claimed

Caveman: advertised versus measured

Column height tracks the percentage. Chat-style 65% versus 8.5% on forced agentic tasks.

65%

Advertised
chat output

8.5%

Measured
agentic output

Most agent tokens are code and tool output, which Caveman leaves untouched.

DualMode loop

1. Luna boot
Read repo, gather facts

→

2. Compress
1–3K token packet

→

3. Sol plans
No repo access

→

4. Luna executes
Edit, test, verify

What holds up

Official Luna credits are far cheaper than Sol. DualMode keeps Sol on a tiny packet. Caveman did not hurt task quality in a paired test.

What does not

The stacked 5x figure is unpublished. Caveman’s 65% cut is chat, not agent coding. Some developers call the terse mode branding for “be brief.”

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.