agentsclimarketplace

13yo eval

Skill kronael/tools/skills/13yo-eval

UX walkthrough of a web app from a 13-year-old novice's eyes (with manual access). NOT for functional/e2e testing (use software), design-craft critique (use design-eval), or visual fixes (use visual).From its SKILL.md

Install
npx -y skills add kronael/tools --skill 13yo-eval

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 6 stars6 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

9.0 KB, as published. Nobody here has run it

13yo-eval - fresh-eyes UX walkthrough

Solo UX evaluation through one persona: a bright 13-year-old novice who can read the manual. Goal: find what's confusing, what only the manual explains, what stays unclear even after the manual. Methodology source: <cwd>/docs/ux-13yo/research-uxtest.md. Produced by scavenge.

When to invoke / NOT

INVOKE: UX walkthrough, fresh-eyes pass, novice walkthrough, "test as a 13yo", post-feature usability check, periodic hygiene pass.

NOT for: functional/e2e tests (use software), design-craft critique (use design-eval), visual fixes (use visual), a11y conformance audit, performance/load testing, apps with no users beyond the team.

The persona

Bright, literate, will read the manual. Catches: undefined acronyms, unexpanded jargon, screens that assume domain priors, intimidation by density, moments where they can't predict what a click will do, buttons whose verbs have specialised meaning.

Blind to: financial judgement, expert pain, performance, correctness.

Walkthrough lens for every task probe (cognitive walkthrough, Wharton & Polson): (1) does the persona know what to try, (2) do they notice the control, (3) do they connect control to effect, (4) after acting, do they see progress? Record WHICH question failed — it names the fix (discoverability vs labeling vs feedback).

Consequence overlay — automatic, not optional

Run the six-prompt overlay on any screen showing:

  • numeric data the user is asked to evaluate or compare (anything beyond decoration)
  • time-sensitive state — data that ages, where "as of when?" matters
  • actions with external / persistent consequence — sign, submit, commit, send, transfer, publish, deploy, schedule, delete
  • the same entity surfaced in multiple places — cross-screen consistency only matters when entities recur
  • a distinction between preview / draft / staging and live / committed / real

Finance is one specialisation; the same six prompts apply anywhere state has weight outside the browser tab (admin panels, deploy tools, schedulers, e-signing, healthcare records — anywhere the user acts on quantified state with real-world consequence).

Six prompts, fixed order, each pass / fail / NA + one sentence:

  1. Units & denominators — every number labelled?
  2. Freshness — what timestamp / epoch?
  3. Comparability — same metric, same name across sibling screens?
  4. Reconciliation — number traceable to a source?
  5. Irreversibility — real-money clearly distinguished from preview?
  6. Cross-screen consistency — same formatting, same colour intent?

Setup (do BEFORE running the protocol)

  1. Start dev server. Note URL.

  2. Identify every route and every inner panel (slide-overs, modals).

  3. Create <cwd>/docs/ux-13yo/ if missing. Drop methodology (research-uxtest.md), protocol (protocol.md), template (report-template.md). If they don't exist, scavenge them via scavenge.

  4. Scaffold every per-screen report file NOW, before walking anything. Each file has these headings empty, ready to fill:

    • ## Cold findings (pre-manual) — appended during steps 1–11
    • ## Post-manual findings — appended ONLY in step 12
    • ## Glossary, ## Consequence overlay, ## Off-happy states, ## Viewport, ## Keyboard, ## Delights, ## Verdict

    Appending as you go survives interruption. End-of-run writes do not.

Per-screen protocol

ALWAYS in order. Append to the scaffolded file after each step.

  1. Land cold (Perfetti's 5-second test). Fresh tab, no scroll, 5 sec. One sentence: what is this for, who? Skip for inner-only panels — 5-second tests only mean something on landing/overview.
  2. Trunk test (Krug). Site ID? Page name? Main sections? Where am I in them ("you are here")? — all without scrolling.
  3. Primary job(s). Plural allowed for dashboards.
  4. Concrete user question. Pick one. Answer using only the screen.
  5. First-click probe. Two tasks. BEFORE each click, write the expected outcome (think-aloud discipline); click; record match or mismatch — a violated expectation IS the finding. Flag wrong first clicks: a right first click predicts ~87% eventual task success vs ~46% wrong (Bailey & Wolfson) — that ratio motivates the flag, but n=1 carries no statistic; label "first-click, single evaluator".
  6. Hover & inspect. Timebox 5–8 min. Tooltips: present / missing / carrying essential content (essential content does not belong in a tooltip).
  7. Microcopy scan. Glossary table: term → presentation → guess → real meaning.
  8. Consequence overlay (mandatory if triggers above — six fixed lines).
  9. Off-happy states — ACTUALLY trigger. Throttle, block API, empty filter. Imagined states are not evidence. If unreachable: "not observed, attempted: ___".
  10. Viewport sweep. 1440 / 1024 / 768. Flag 375 only if mobile is claimed.
  11. Keyboard pass. Tab order, focus ring, every interactive reachable.
  12. NOW read the manual. Open /docs (or equivalent). Append every finding to ## Post-manual findings — NEVER edit or merge earlier observations. Retroactive reclassification destroys the discoverability evidence the protocol exists to preserve.
  13. Delights. What worked. End-of-screen.

Classification

Each observation gets one:

  • CLEAR — got it without help.
  • MANUAL — manual closed the gap → discoverability bug (info exists, not where eye lands).
  • STILL UNCLEAR — manual didn't close it. Sub: docs-gap / prior-domain-knowledge / UI-ambiguous.

Finding format

**[cite|inference] [location]** — what happened. Persona narration
(inference). Classification: CLEAR / MANUAL /
STILL UNCLEAR (sub). Evidence: <screenshot path or observed text>.

Tags are evidence types, not severity:

  • cite — observed this run: UI text, screenshots, click outcomes
  • inference — persona reasoning; persona narration is ALWAYS inference, NEVER cite — there is no real 13-year-old to quote

NEVER use severity adjectives: small, minor, major, critical, huge, tiny, serious, important. Triage adds severity later — that's their job.

Subagent fan-out

For ≥5 screens, fan out via subagents using the browse skill (drives the agent-browser CLI via Bash — spell this out in the subagent prompt).

  • Default 1–2 subagents. 3+ only with explicit budget for contention (usage caps hit fast in parallel).
  • Bucket related screens (Basic+Expert pair, detail panel with parent).
  • Each subagent OWNS its bucket exclusively.
  • Subagents scaffold report files FIRST, append per step. NEVER let a subagent write only at the end of its run — interruption loses everything.
  • Main thread aggregates into summary.md.

Salvage rule

If a subagent captured evidence (screenshots) but the report write failed, spawn a writer-only follow-up subagent that uses the existing evidence + a minimal re-walk. Do NOT re-run the full protocol.

Rules

  • ALWAYS hard-separate ## Cold findings (pre-manual) and ## Post-manual findings. NEVER merge them.
  • ALWAYS run the consequence overlay on any trigger-listed screen; render it as exactly six lines in the fixed order.
  • ALWAYS cite evidence captured this run; findings without it get cut.
  • ALWAYS use the browse skill for browser automation (spell out in subagent prompts).
  • NEVER let evaluator knowledge leak into persona narration — if you (the model) know why the UI is that way, the persona doesn't; record the confusion, not the excuse (a facilitator never explains the UI).
  • NEVER use severity adjectives in findings. Tags are cite / inference — evidence type, not severity.
  • NEVER use imagined error states as evidence — trigger them or mark not-observed.
  • NEVER claim multi-persona coverage from one run.
  • ALWAYS record methodology limits in summary.md: one evaluator, one persona, one moment. One evaluator finds ~35% of problems (Nielsen); overlap runs 5–65% (Hertzum & Jacobsen) — a sample, never a census.

Anti-patterns

  • Importing Bailey & Wolfson 87/46 % as if it applied to solo screenshot tests — it doesn't; the ratio is the heuristic.
  • Copy-editing instead of microcopy scanning — the question is "would a novice understand", not "is this the best wording".
  • Hardcoding 375 px viewport flags on desktop-first tools.

Reference

Full methodology: <cwd>/docs/ux-13yo/research-uxtest.md. Oracle critiques (v1 + v2): <cwd>/docs/ux-13yo/oracle-critique.md, <cwd>/docs/ux-13yo/oracle-skills-critique.md. Skill that produced this one: scavenge.

Keep looking

Skills are one crate of 610,220. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.