PARTLY

Partly true: Claude Code really does sometimes skip instructions and skills, but the claim it is 'deliberately dumbed down' to burn tokens has no evidence behind it and is contradicted by how most users pay; the reel ends in a coaching funnel

Facebook · Oct 2026

The claim'Claude Code sucks and all these influencers are full of it.' After hundreds of hours and dozens of projects, it fails simple instructions, skips skills or pretends it used them, and constantly hallucinates. 'I believe Claude Code is deliberately being dumbed down' to run people in circles and drive token usage up. 'Big tech doesn't give us the most powerful AI they have.' The fix: run open-source CLI agents on fine-tuned local open-weight models. Comment 'AI' or 'LOCAL' for a guide, consult call, group or coaching.

Two different claims are mixed together here. The experience part is partly true. Claude Code does not follow every instruction every time, and that is documented: Anthropic's own skills docs say Claude decides when to load a skill from its description, so automatic use is a judgment call, not guaranteed. Public bug reports describe Opus models ignoring CLAUDE.md rules or announcing work they never do. Anthropic itself admitted in an April 23, 2026 postmortem that three product changes degraded Claude Code quality for about six weeks. The accusation part is not supported. That same postmortem states 'We never intentionally degrade our models', and no leak, filing or independent test shows deliberate dumbing down to burn tokens. The incentive story also runs backwards for most users: Claude Code is mainly used on flat monthly plans ($20 to $100 and up) with usage caps, so wasted loops cost Anthropic compute rather than earning it more. One fair point: Anthropic does hold back its strongest model, Claude Mythos Preview, limited to vetted partners since April 2026 for cybersecurity safety reasons, which is publicly disclosed, not a token scheme. Running local open-weight models is a legitimate option for privacy and control, and the best open models are now close on some coding benchmarks, but they need serious hardware and still fail instructions too. The reel's real product is the 'comment AI / LOCAL' coaching funnel.

What holds up

Transcript and caption read in full; on-screen chapters: 'Claude Code fails constantly', 'Big tech dumbs down AI', 'Run local open source'; CTA to comment for guide, call, group, coaching.
Claude Code skills docs: Claude uses a skill 'when relevant' based on its description; explicit /skill-name invocation and disable-model-invocation give deterministic control.
Anthropic engineering postmortem, Apr 23, 2026: three product-layer causes degraded Claude Code Mar 4 to Apr 20; 'We never intentionally degrade our models.'
Claude pricing page: Pro and Max are flat monthly plans with 5-hour and weekly limits that include Claude Code. Claude Mythos Preview withheld to partners via Project Glasswing (Apr 7, 2026).

What doesn’t

'Deliberately dumbed down to burn tokens' is asserted with no evidence.
Flat-rate plans make wasted loops a cost to the vendor, not revenue.
Coaching and consult funnel behind 'comment AI / LOCAL'.
Local models trade one set of limits for another (hardware, setup, their own errors).

The catch

The frustration is real and partly documented. The conspiracy is not: there is no evidence of deliberate throttling, and the billing model most people use points the other way.

How to actually do it

Invoke must-run skills explicitly with /skill-name, and move anything that must always happen into hooks or scripts rather than prompt instructions.
Keep sessions short, restate key rules near the task, and check outputs against tests instead of trusting 'done'.
If you want local models for privacy, try an open-weight coding model with an open CLI agent on a side project first and compare results honestly.

Real reliability complaints are welded to an unproven 'deliberately limited' theory to sell coaching. Deterministic hooks and explicit skill calls fix more than switching vendors does.

Confidence: Medium
Posted by: an AI automation coach (about 8,800 plays) selling a free guide, consult calls, a community and 1-on-1 coaching; reel sent to Buddy Wed Oct 28, 2026, late night

See the original claim →

We test hype for free. We build the real thing for a living.

Thirty minutes, no pitch — and you'll leave with something useful either way.

Book a call with Todd or start with the free Business Checkup →

The Verdict Weekly

Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.