Which Claude Code framework should your team use? Start with none

In September this blog looked at what happens when installing outruns choosing: A dev team in a folder, and the question it doesn't answer. This is the follow-up, with evidence.

The two most-starred add-ons for Claude Code, Superpowers and Everything Claude Code (ECC), hold roughly 300,000 and 270,000 to 280,000 GitHub stars [1][2]. Stars measure attention, not quality. A search run on 17 October 2026 looked for an independent test showing that either one, or Spec Kit, GSD or BMAD, makes Claude Code write better code than plain Claude Code. It found none [3].

What the controlled studies measured

The closest thing to proof tests the ingredients rather than the brands. SWE-Skills-Bench ran 49 coding skills against about 565 real-repository tasks, each task with and without the skill. 39 of the skills changed nothing. The average gain was 1.2 percentage points (89.8% to 91.0%). Seven skills helped, by up to 30 points, and three made results worse, by up to 10 [4].

Instruction files did no better. A study of AGENTS.md-style files found they did not generally improve success and raised cost by more than 20% on average [5].

There is an encouraging result, and it is about size. In SkillsBench, curated skills lifted pass rates across domains from 33.9% to 50.5%, and focused skills of three modules or fewer beat large bundles [6]. For software engineering alone the lift was smaller: 4.5 points in an early version [7].

None of these studies ran Superpowers or ECC, so applying them to those two is an inference. Several used smaller, older models such as Claude Haiku 4.5 [4], so newer models may behave differently.

What the head-to-head tests show

Only two public comparisons turned up that put the tools side by side, and each is a single task. In Felipe Fontoura's test on 13 October, BMAD took 217 seconds and $1.89, Spec Kit 473 seconds and $2.84, and Superpowers 40 seconds and $0.33. All three passed his 13 of 13 checks [8]. In Chase AI's test in April, plain Claude Code built a website in about 20 minutes and 200,000 tokens, Superpowers in about 60 minutes and 250,000, and GSD in about 105 minutes and 1.2 million. He called GSD's output not meaningfully better [9].

Look at Superpowers across the two: far quicker than BMAD and Spec Kit in one, three times slower than plain Claude Code in the other. Different tasks, models and dates, so these are anecdotes. One author sells a competing tool, the other judged quality by eye [8][9].

Superpowers' own authors say it uses a lot of tokens and builds more slowly than working without it [10]. Their claims of faster, cheaper builds compare version 6 with version 5, not with plain Claude Code, and come without a public results table [11].

What most comparisons measure instead

The vendor evals check whether the agent followed the framework's workflow, not whether the code improved [12]. A framework can score perfectly on that and still not move the outcome.

One study of repositories that adopt spec tools suggests where the value may come from: the developer having articulated goals before delegating [13]. If that is right, the habit matters more than the brand.

Pick by the problem you actually have

This part is my judgment, built on the evidence above. It is not a measured result.

Default to plain Claude Code and a short, hand-written CLAUDE.md (The instruction file getting shorter, not longer). Add one layer, and only for a failure you can name.

If it builds the wrong thing, or decisions drift between sessions, try a spec tool: OpenSpec is the light one, Spec Kit the heavier and more auditable. If it builds the right thing badly or skips tests, try Superpowers. If long sessions lose the plot, try GSD Core.

If you want specialists, take two to five specific skills from ECC or wshobson/agents instead of the bundle. ECC ships about 70 agents and 300 skills, and the count changes between releases [14]. For parallel work, start with Claude Code's own subagents; agent teams are experimental and off by default [15].

Do not stack two full workflow frameworks. Each brings its own planning, testing and review flow, and they will compete for the agent's attention. And read the hooks before installing anything globally: one audit counted 48 hook scripts and more than 500 auto-loadable instruction files on a full ECC install, and a malware look-alike clone has been reported [16]. Every always-loaded file also costs tokens every session (Your setup costs tokens every session).

Run your own test before you commit

The cheapest way to settle it is a small experiment. Take 10 to 15 real tickets. Write the acceptance tests before the run and hide them from the agent. Run each ticket at least three times on three or four setups: plain, plus Superpowers, plus one spec tool, and optionally ECC with a minimal install. Keep the model, the Claude Code version and a fresh clone constant.

Compare hidden-test pass rate, tokens and dollars, time, and defects found in review, using medians rather than single runs. My estimate is one to two days and $100 to $300 of API spend [17]. Adopt a framework only if it beats the run-to-run noise at a cost multiple you accept.

The summary is short. Process tools change cost and time a lot, and nobody has shown they change quality. So start plain, add one thing, and watch what it does to your own numbers.

Sources and confidence

Star counts and prices move, so re-check live numbers before relying on them. Labels: Certain means read on the primary source; Likely means a credible secondary source, or a figure that moves; Verify means sources conflict or it is an estimate.

1. Superpowers has about 300,000 GitHub stars (about 300.8k on 17 Oct 2026). Likely. github.com/obra/superpowers
2. ECC has 267,959 to 282.1k stars depending on the tracker. Verify. github.com/affaan-m/ECC, plus star trackers that disagree
3. No independent head-to-head benchmark of these frameworks against plain Claude Code was found. Likely: this is an absence in one search, not proof. Re-check arXiv and each repository
4. SWE-Skills-Bench: 49 skills, about 565 tasks, 39 with no change, average +1.2 points, 7 helped up to +30, 3 hurt up to -10 (the abstract states these as percentages; I read them as points); run with Claude Code and Claude Haiku 4.5. Certain. arxiv.org/abs/2603.15401
5. Context files do not generally improve success and raise cost by more than 20% on average. Certain. arxiv.org/abs/2602.11988
6. SkillsBench: curated skills 33.9% to 50.5% (+16.6 points); skills of three modules or fewer beat larger bundles. Certain. arxiv.org/abs/2602.12670
7. SkillsBench software-engineering gain of 4.5 points (early version, as reported by a survey). Likely. arxiv.org/abs/2606.11435
8. Fontoura test, 13 Oct 2026: BMAD 217 s and $1.89, Spec Kit 473 s and $2.84, Superpowers 40 s and $0.33, all 13 of 13; OpenSpec also ran in that test but its result was not captured; the author sells a competing tool. Likely. felipefontoura.com/articles/spec-driven-development-tools-compared
9. Chase AI test, 13 Apr 2026: plain about 20 min and 200K tokens, Superpowers about 60 min and 250K, GSD about 105 min and 1.2M; times and tokens are approximate, quality judged by the author. Certain as reported by the author. chaseai.io/blog/gsd2-vs-claude-code-which-ai-coding-tool-wins. The three-times-slower ratio (60 vs 20 minutes) is my arithmetic on those approximations
10. Superpowers' authors on its token use and speed. Certain. primeradiant.com/blog/2026/superpowers-6.html
11. Superpowers 6 claims "up to 50% faster and up to 60% cheaper" than version 5, with no public table; the post admits an evaluation bug. Certain that these are the vendor's claims; the numbers themselves are Verify. primeradiant.com/blog/2026/superpowers-6.html
12. The Superpowers eval suite grades workflow compliance and compares Superpowers versions, not plain Claude Code. Certain. github.com/prime-radiant-inc/superpowers-evals
13. A study of spec-tool adoption suggests the benefit may come from developers articulating goals before delegating. Likely. arxiv.org/html/2608.25202
14. ECC v2.2.3 lists 71 agents and 311 skills; the September post counted 68 and 286, so the numbers move. Verify. github.com/affaan-m/ECC/releases/tag/v2.2.3
15. Claude Code agent teams are experimental and disabled by default. Certain. code.claude.com/docs/en/agent-teams
16. An audit counted 48 global hook scripts and more than 500 auto-loadable instruction files on a full ECC install, and reported a malware look-alike clone. Likely. dev.to/joergmichno/we-audited-the-viral-213k-star-everything-claude-code-repo-and-found-a-malware-clone-in-the-wild-14hb
17. The test size, 1 to 2 days and $100 to $300, is my estimate. Verify. No external source; run one ticket first and measure