Human-approved checkpoints for Superpowers' planning flow

---

An add-on plugin for Superpowers that inserts Dex Horthy's (HumanLayer) 4-gate model — Product → Architecture → Program Design → Vertical Slices — as explicit, human-approved checkpoints into the Superpowers planning flow.

flowchart LR
    subgraph SP["Superpowers"]
        B["brainstorming"]
        W["writing-plans"]
        X["subagent-driven-development /<br/>executing-plans"]
    end
    subgraph FG["factory-gates (this plugin)"]
        A["architecture-gate"]
        P["program-design-gate"]
        V["vertical-slices-gate"]
    end
    B --> A --> P --> W --> V --> X

Loading

This is a non-invasive companion plugin. It does not fork or edit any Superpowers file. It adds three new skills that slot into the gaps between Superpowers' existing skills, plus a /factory-gates command for starting a feature with deterministic gate coverage.

🧩 Why

Superpowers already gates two of Horthy's four steps well:

Horthy's Gate	Superpowers today
1. Product	brainstorming — covers this well
2. Architecture	Implicit — mixed into writing-plans' "File Structure" section, not separately approved
3. Program Design	Missing — no explicit signature/call-stack contract before implementation
4. Vertical Slices	Mostly covered by writing-plans' task breakdown + execution handoff, but no explicit sign-off on slice order or multi-repo coordination

This plugin adds:

• architecture-gate — Gate 2. Component boundaries, data models, cross-cutting constraints. Runs after brainstorming, before writing-plans.
• program-design-gate — Gate 3. Types, signatures, call stacks, file layout — a "header file" for the feature, no bodies. Runs after architecture-gate, before writing-plans.
• vertical-slices-gate — Gate 4. A short, explicit sign-off on slice order and multi-repo/service sequencing. Runs after writing-plans has saved a plan, before subagent-driven-development / executing-plans starts.
• /factory-gates <feature description> — an explicit, user-invoked entry point. Starts the workflow with the gate sequence stated up front, rather than relying on brainstorming's own routing to pick it up. See "Known limitation" below for why this exists.

The result: brainstorming → architecture-gate → program-design-gate → writing-plans → vertical-slices-gate → subagent-driven-development / executing-plans.

📦 Install

Requires Superpowers already installed. From your project (or globally):

/plugin marketplace add /path/to/factory-gates
/plugin install factory-gates@factory-gates

Or point at a git remote once you've pushed this directory to one:

/plugin marketplace add chrschy/factory-gates
/plugin install factory-gates@factory-gates

⚠️ Known limitation — read this

Warning

Superpowers' own brainstorming skill ends with an explicit instruction: "Invoke the writing-plans skill... Do NOT invoke any other skill." Because this plugin doesn't edit Superpowers' files, that instruction still exists verbatim. architecture-gate's description is written to be highly specific ("use immediately after brainstorming, before writing-plans") so Claude picks it up as the more specific match — and Superpowers' own using-superpowers router skill instructs Claude to invoke any applicable skill, not just the one another skill points to. This is a soft override, not a hard one, and this repo measures it empirically (tests/gate-routing/) rather than just asserting it works. A bare request can also skip brainstorming entirely and jump straight into an unrelated skill (e.g. test-driven-development) instead — a more severe pattern than losing only the architecture-gate handoff, and one the test harness now detects and classifies as a failure rather than an inconclusive result.

Three ways to make it more reliable, all measured, none of them a hard guarantee — each works by adding an explicit, early statement that brainstorming's hard "invoke writing-plans only" instruction has to compete against, not by editing Superpowers' own files or routing rules:

Mechanism	How	Measured result
(bare, no phrasing)	Plain feature request, no workaround	3 trials (small sample — see note below): 1 fail (skipped brainstorming entirely, jumped straight to another skill instead), 2 inconclusive (brainstorming fired correctly, then got detoured mid-conversation into an unrelated skill before running out of turn budget) — this is the baseline the mechanisms below improve on
Say so explicitly	Start the feature with "Use the factory-gates workflow for this."	22 formal trials, 0 fails; explicit phrasing also correlates with fewer non-completions than a bare request
/factory-gates <description>	Use the command instead of a plain message	3 trials, 0 fails, 0 inconclusive
CLAUDE.md project instruction	Add the snippet below to your project's CLAUDE.md once	3 trials, 0 fails, 0 inconclusive

Note

The bare-scenario sample above is intentionally small: verifying the test harness's own classification logic (see tests/gate-routing/) surfaced and fixed several real bugs in it, and confirming the fix consumed the available trial quota for this round. Expanding this sample is tracked as a follow-up.

Tip

Add this to your project's CLAUDE.md once, and you won't need to remember any of the above per feature:

## Software Factory Workflow
For any new feature or creative work, use the factory-gates workflow — architecture-gate and program-design-gate run before writing-plans, vertical-slices-gate runs before execution. Do not skip these gates even if a skill's own instructions say to invoke writing-plans directly.

🔗 Superpowers compatibility

Verified against Superpowers 6.2.0 – 6.3.0. There's no dependency-version mechanism in the Claude Code plugin format to declare this formally — plugin.json has no dependencies field, checked directly against every plugin manifest in this environment's local marketplace cache, including Superpowers' own. This is a documented, empirically-tracked stance, not a mechanically enforced one.

The coupling is narrow but real: factory-gates' routing strategy (see "Known limitation" above) depends on specific wording in Superpowers' own skills, not any formal API —

• brainstorming's hard "invoke writing-plans... do NOT invoke any other skill" instruction
• using-superpowers's "if a skill applies, you MUST use it" routing rule

If a future Superpowers release changes that wording materially, factory-gates' soft-override strategy could degrade silently — there is no automated check outside of actually running the test suites.

tests/gate-routing/lib/common.sh prints a warning to stderr if the installed Superpowers version falls outside the verified range (older: a real warning; newer: a softer note, since untested doesn't mean broken) every time a test suite resolves the Superpowers plugin directory. If you see one, or if you've just updated Superpowers and something in the factory-gates workflow feels off, re-run tests/gate-routing/run-all.sh — that's exactly the empirical check this exists for.

🙏 Credits

• 4-gate model: Dex Horthy, HumanLayer — as described in his "Harness Engineering is not Enough: Why Software Factories Fail" talk and various podcast appearances (2026).
• Base framework this plugs into: obra/superpowers (Jesse Vincent).

This is an unofficial community glue layer, not affiliated with HumanLayer or the Superpowers project.