Seven stars. Zero gained in the last week. If you're scrolling SkillsMP looking for what's trending, blast-radius isn't it. But if you've ever shipped a change that looked safe on a diff and broke something three files away, this skill's methodology is worth your attention — even if you have to grit your teeth about the platform it ships on.
What this skill actually does
blast-radius is a review skill. You give it a diff or a change description, and instead of writing you a polite paragraph about what might go wrong, it forces the agent to find the single fact that makes a change safe and then prove it by running code. Not by citing a comment. Not by pointing at a line. By executing something that fails loudly if the assumption is wrong.
The skill's author is blunt about why: "A blast-radius writeup that sounds right is worthless. It reads as convincing whether or not it's true, and that is the trap you are walking into." That's not marketing copy. That's the entire philosophy compressed into one sentence, and it's the kind of thing that should make every developer who's ever written a PR review feel uncomfortable.
The gap it fills
Most code review tooling — and most AI-assisted review — stops at "here are the callers" or "here's my analysis." Grep finds callers in a second. The skill is explicit about this: "Listing the callers is not the job." The job is the breakage grep won't show you. The JSON an API returns. The DB column that's actually populated. The feature flag that defaults to false. The microtask timing that means your cleanup runs before your handler finishes.
This is a real gap. I've seen plenty of reviews that correctly identify every direct caller and still miss the fact that a function now runs in a different lifecycle hook, or that a dependency upgraded its pinned version and changed behavior on a corner case nobody thought to check. blast-radius formalizes the discipline of looking past the obvious.
What I actually found useful
Three things stand out from reading the SKILL.md:
First, the confidence ladder. For each safety fact, the skill forces you to rank how sure you are on a scale from "you said so" (worthless) to "you reproduced it in the running app." Any fact that can't reach step 4 — a script or test that runs the real code — gets explicitly marked as unproven. No sugarcoating. No hand-waving. That's a standard most human reviewers don't meet, let alone AI agents.
Second, the "one fact it's safe because of" heuristic. Most scary-looking changes are safe because of a single concrete reason. Find it, prove it, and most of the anxiety evaporates. This is genuinely good engineering thinking, and the skill formalizes it into a workflow step instead of leaving it as vague advice.
Third, the companion skill ecosystem. blast-radius is part of pstack-for-codex, which includes $how, $why, $interrogate, $tdd, and others. If you're already using the suite, blast-radius slots in naturally. If you're not, it still works standalone — but you'll miss the context of the workflow it was designed to live inside.
Who should install this
If you're a developer who regularly reviews diffs you don't fully trust, this skill's methodology is valuable regardless of whether you use it through Codex or transplant the reasoning into your own workflow. The confidence ladder alone is something I'd recommend any reviewer internalize.
If you're working in a codebase where "it compiles" isn't enough and where silent failures in distributed systems or async boundaries are real concerns, the "look where grep stops" instruction is directly applicable.
If you're already deep into the pstack-for-codex ecosystem, this is a natural fit alongside $why and $how.
Who should skip it
If you're looking for a Claude Code skill specifically — and you're not running Codex — the installation path is through the Codex CLI plugin marketplace, not .claude/skills/. The skill references ../poteto-mode/references/codex-agent-runtime.md and the entire pstack suite is Codex-native. You could extract the methodology, but the skill itself won't load in Claude Code without significant rework.
If you want a quick review summary or a formatted report, this isn't it. The skill explicitly says not to hand back a writeup. It wants proof. If your team culture rewards the well-written comment over the failing test, this skill will frustrate you.
If you're reviewing trivial changes — a typo fix, a config tweak — the overhead of writing a script to prove safety is absurd. The skill knows this too; it's designed for changes you "don't trust yet."
Honest concerns
The biggest issue is platform lock-in. This is a Codex plugin. The install instructions are codex plugin marketplace add Aqua-123/pstack-for-codex. There's no Claude Code equivalent. If you're a Claude Code user, you're either switching tools or manually porting the methodology, and neither is trivial.
The 7-star count with zero recent growth tells me this hasn't found its audience yet. That could mean the methodology is ahead of its time, or it could mean the skill has friction that prevents people from adopting it. I lean toward the latter — the requirement to run real code as proof is expensive, and not every change warrants it.
There's also the question of whether an AI agent can reliably execute step 4 (run a script that fails if you're wrong) without introducing its own errors in the test script. The skill acknowledges this by saying "say so out loud" if you can't get to step 4, but an agent writing a bad test script and calling it proof is a real failure mode.
The verdict
blast-radius is more a methodology than a tool, and as a methodology it's genuinely good. The insistence on running code over writing prose is the right instinct, and the confidence ladder gives you a framework to judge your own certainty instead of pretending you're sure about something you're not.
But it's a Codex skill in a Claude Code world, it has 7 stars, and the proof-running requirement means it's not always cheap to use. I'd install it if I'm in the pstack ecosystem. If I'm not, I'd steal the methodology and write my own version for whatever tool I'm actually using.
Either way, the core idea — that a convincing explanation is worthless without a failing test to back it up — is worth paying attention to.
Links: - SkillsMP: https://skillsmp.com/creators/aqua-123/pstack-for-codex/skills-blast-radius - GitHub: https://github.com/Aqua-123/pstack-for-codex/tree/main/skills/blast-radius