Tech • AI • Robotics • Game
Do Not Use These Two Tools (Here’s What to Use Instead)
6/10TL;DR
SuperPower and BMAD are drawing heavy criticism from advanced Claude Code users who argue that the frameworks add massive prompt overhead, higher token costs and slower workflows without delivering meaningfully better code.
KEY POINTS
Why the backlash is growing
SuperPower and BMAD are marketed as productivity boosters for AI coding, with claims of faster delivery and better structure. Yet critics argue the opposite: the tools often force rigid multi-step workflows onto tasks that modern models can already complete directly. The core complaint is that they create the appearance of professionalism while making routine work slower and more expensive.
Large adoption numbers, mixed reputation
SuperPower is associated with roughly 300,000 GitHub stars and about 1 million installations in the Claude Code ecosystem, while BMAD shows about 54,000 stars. Those figures reflect strong hype and visibility, but they do not settle the question of whether the methods improve outcomes. Detractors say popularity has been driven more by social proof and branding than by measured gains in developer efficiency.
What these tools actually are
The criticism centers on the fact that both products are essentially large collections of Markdown instructions, rules and workflow files rather than new model capabilities. BMAD reportedly includes more than 31 skills, 274 Markdown files and over 12,000 lines of text, totaling about 1.7 million characters, or roughly 350 pages. SuperPower is described similarly, with more than 15 skills and extensive mandatory process files.
Heavy workflows for simple changes
Opponents say BMAD can require a brief, product requirements document, UX file, architecture document, epics, stories, builds and reviews before coding begins. SuperPower is criticized for comparable steps such as brainstorming, workflow setup, testing and code review even for ordinary feature work. The result, they argue, is a framework that treats small tasks as if they were enterprise-scale delivery programs.
Prompt injection and context bloat
A major technical complaint is that these frameworks inject large amounts of instruction text at session start through hooks and skill loading. Critics say this means the model reads pages of rules before addressing the actual request, and may be forced to invoke specific skills whenever there is even a slight chance they apply. In practice, that can consume context windows rapidly and leave less room for the real coding problem.
Research cited against long prompts
Critics point to broader findings on LLM behavior: reliability tends to fall as input length increases, and models often struggle with information buried in the middle of long contexts. This aligns with the widely discussed lost in the middle effect, where beginning and ending tokens receive more attention than dense instruction blocks in between. The argument is that bloated workflows do not merely cost more tokens; they can also make the model less accurate.
More agents, more cost
Another objection is the use of multiple sub-agents, fresh chats and repeated review loops for work that a single model could often complete in one pass. BMAD is said to start a fresh chat for each workflow stage to avoid context overflow, but critics counter that this simply multiplies token usage across steps. The overall effect is described as paying several times over to accomplish the same task.
The case for simpler skills
In contrast, lean prompting approaches are being held up as better alternatives. Small skills built around clarification questions and lightweight planning are favored over full frameworks. A frequently cited example is Matt Pocock’s compact prompting style, with some skills reportedly as short as 6, 25 or 40 lines, compared with framework files stretching into hundreds or thousands of lines.
Most software changes are small
Critics also argue that the average coding task simply does not justify heavyweight process. They cite internal engineering patterns showing that many code changes are tiny: the median change at Google is said to be about 24 lines, with roughly 10% affecting only one line. If most real work is small, they argue, forcing every change through a long specification-and-review machine is inefficient by design.
A three-step alternative
The proposed replacement is simple: define the need clearly, align the prompt with the intended outcome, then let the model execute. Instead of inventing teams of fictional agents or generating multiple planning documents, the model is asked direct questions such as what the feature is, who it is for, what it should look like and how completion will be verified. The premise is that stronger models now handle most coding tasks naturally, making many older prompt frameworks obsolete.
CONCLUSION
The dispute over SuperPower and BMAD reflects a broader shift in AI coding: as models improve, developers are increasingly questioning whether elaborate prompt frameworks add value or just overhead. For many advanced users, the emerging rule is straightforward: fewer instructions, less ceremony and more direct prompting.
Ask a question