In Short

  • Problem: when a Codex process is already framed, does a tool like Stax really improve the result, or does it mostly add another layer of context?
  • What I tested: four variants on the same repo and the same SEO data pack: no Stax, with a skill, with Stax, and with Stax plus a skill.
  • Result: both blind scorings gave the same order: st+clear, clear, st+skill, skill.
  • Lesson: in this lab, Stax gave a small positive signal, especially on repository-side actionability. The skill helped formalize the work, but did not win by itself.
  • Limit: this is one lab, on one repo, one data pack, two Codex scorings, and no usable Claude score.

The Question I Wanted To Ask

I did not want to test Stax on a vague task. It is too easy to launch an agent tool and decide it “looks better” because it added files, instructions, or a more rigorous shape.

The better question was narrower:

if the process is already established, does Stax still improve anything?

To avoid judging by feel, I used a real but bounded task: produce an SEO quick-wins report from Google Search Console and Google Analytics exports for pezzoslabs.com, then make one very simple homepage change.

The report tested analysis quality. The file edit tested whether the agent stayed clean inside a real repository. The traces gave an imperfect but useful cost view: duration, tokens, touched files, validations, and retries.

The Four Runs

The lab compared four variants:

  • a clear Codex prompt without Stax;
  • a clear Codex prompt with a repo-local skill;
  • a Stax-assisted run;
  • a Stax-assisted run with the skill.

The same data pack and repo context were used to keep the comparison small. The point was not to prove a universal ranking. It was to see whether the extra surface changed the result in this particular workflow.

What The Scores Said

Both blind Codex scorings produced the same order:

  1. st+clear
  2. clear
  3. st+skill
  4. skill

That does not make the result universal, but it is more useful than a feeling. The best output was the Stax run with a clear prompt, not the run that merely had the most process around it.

The qualitative difference was also narrow. st+clear connected recommendations to the repository a little better, pointing more naturally to files such as src/pages/..., src/data/accompagnement-pages.ts, or src/data/structuredData.ts.

That is a signal, not a doctrine. On this already-clear process, Stax seemed to help the agent produce a slightly more actionable code-side output.

What I Keep From It

In this lab, Stax helped a little. The signal was strongest where the answer needed to be actionable in the repository: what to change, where to look, and how to connect data to next steps.

The skill did not win alone. That does not mean skills are useless. It means a skill is not a magic quality layer. It has to match the task and the available evidence.

What Not To Copy Blindly

I would not generalize from this single benchmark. It used one repo, one SEO pack, one task shape, and scoring by Codex rather than a diverse human panel.

I also would not say that Stax made the workflow cheaper. In the comparable pairs, total duration and token counts increased with Stax. I could not observe exactly what Stax did or did not send to its service either; the lab trace only says the orchestration did not pass GA/GSC exports to Stax commands, and network behavior or telemetry cannot be proven from stdout alone.

The useful thing to copy is the discipline: compare variants on the same input, reveal the mapping after scoring, record costs and limits, and avoid turning a small positive signal into a product claim.

Where The Evidence Lives

The public repo contains the reports, prompts, scorings, synthetic traces, diffs, revealed variant mapping, and RESULTATS.md. It does not contain the raw GA/GSC CSV exports, full Codex JSONL logs, or complete worktrees.

It is useful for rereading the protocol and comparing the published outputs. It is not a way to replay the lab exactly with private data, and it does not prove a general performance claim for Stax, Codex, or skills.

The Rule I Keep

When a tool adds context or another layer of method, I do not want to conclude too fast that it improves the work. I want to put it back on the same input, score the outputs before revealing variants, and then look at the cost. In this specific lab, Stax deserves a second test. It does not yet deserve a general rule.