Every AI gadget is some mix of three things:
| Earns its place by | You know it when | |
|---|---|---|
| 🟢 Tool | removing a step from someone's day | it breaks and the workflow breaks with it |
| 🟡 Toy | delight: materials, surprise, room to explore and mod | people keep it on the desk with nothing to do |
| 🔴 Trash | nothing: it adds friction, leaks privacy, or duplicates a free phone app | it is in a drawer within a month |
That question is worth asking before building hardware, not after.
T3 turns it into an audit an AI agent can run on any product page, as the
t3-hardware-scoring agent skill.
npx skills add toyworks/tool-toy-trashor, in Claude Code:
/plugin marketplace add ToyWorks/tool-toy-trash
/plugin install t3-hardware-scoring@tool-toy-trashThen hand your agent a product link:
Run a T3 audit on https://example.com/some-ai-pendant
Any agent that supports the Agent Skills standard can load it.
product page + reviews
└─ brand-blind strip names, so the hype can't vote
├─ tool auditor ┐
├─ toy auditor ├─ fixed rubrics, verbatim evidence only, run in parallel
└─ trash auditor ┘
└─ eagle eye re-check the fatal flags: "on-device" claims that can't be, apps it duplicates
└─ judge normalize, apply the gates and the veto → label
- Evidence first. No score above 0 without a quote from the source text, and a test checks that every quote really is in it.
- Math in code. Normalization, the Gray Zone and the veto live in
synthesize_results.py, not in the model's head. - One veto. A single Eagle Eye flag (privacy tension, a core flaw, app redundancy) forces Trash into the label, however good the rest looks.
A run on the fictional AI pendant in the test fixtures (the "second memory" that promises nothing leaves the device, plus a $19/month plan):
tool 33.3 toy 27.3 trash 71.4 composite -38.1
→ Trash (Eagle Eye) confidence: Review Required
skills/t3-hardware-scoring/
├── SKILL.md the procedure the agent follows, step by step
├── references/ auditor rubrics, brand blinding, Eagle Eye, the judge's rules
├── scripts/ validate + merge the auditor reports, synthesize the verdict
└── tests/ unit tests, a golden worked audit, link and consistency checks
The rules in full: t3-classification.md.
skills/t3-hardware-scoring/tests/run_tests.shStdlib Python only, no install step. CI runs the suite and a link check on every push and pull request.
Change a rule and its test in the same commit; each test names the spec section
it pins. Pull requests are welcome, especially new Eagle Eye red flags: add them
to trash-red-flags.md
first, and the trigger-consistency test will tell you where else they go.
MIT