lazyweb-skill AGENTS.md

Lazyweb skill pack — agent guidance

Lazyweb report principles (governs any agent building a Lazyweb report)

Every Lazyweb skill produces an HTML report. Whatever the skill, the report must embody these four principles — they are the product, not a style preference.

1. Show, don't tell. Prove every claim with a real visual reference or data, never prose alone. Each assertion (pattern, anti-pattern, idea, hypothesis, "what's working", convention check, recommendation, or A/B learning) carries the screenshot(s)/experiment that demonstrate it, sitting with the claim. Quantify prevalence ("5 of 9 references") instead of asserting it ("near-universal"). Never render a proposed layout as ASCII/box-drawing <pre> art — use an HTML/CSS prototype, mock-frame, or generated image.

2. Be opinionated; carry the decision. Take a clear ranked stance and lead with a single recommended path, marked as such in the human-visible body (not only in the machine handoff block). Tell the user what to do first and what to skip, with a one-line reason each. Never hand over an undifferentiated menu of co-equal options and make the reader choose.

3. Maximize confidence with evidence. Back each recommendation with what worked for OTHER apps (real screenshots) PLUS supporting data — prevalence across the corpus, and A/B experiment learnings for growth/monetization screens. When experiment data is unavailable, say so and fall back to a stated prevalence count; never ship a recommendation with no visual and no number behind it.

4. Be truth-seeking. Never overclaim. Label evidence strength honestly (measured vs directional vs single-source/off-category) on each claim, and flag a weak, thin, single-source, or context-mismatched corpus up front. Tag any brand inferred from a URL/vision description as unverified. Never fabricate a reference, a metric, or a company name. The machine-readable handoff and the human-facing body must agree about confidence.

Shared report furniture (use the existing tokens --ink:#1f2328; --mut:#57606a; --line:#d0d7de; --soft:#eef4fb; --accent:#0969da): the light-blue Agent Instructions copy block is always section #1; reuse the shared .deck (evidence carousel — all references visible, scroll-snaps with ◀ ▶ prev/next buttons), .legend + .rec cards (opinionated ranked pick: a decision legend over big-proof recommendation cards, #1 = .rec.lead), .ebadge/.corpus (honest evidence labels), and .mock (mock-frame) components so every skill renders proof, decisions, and honesty the same way. One sanctioned substitution: lazyweb-deep-design-research (report v3) renders its decision as a side-by-side .compare (Control × Recommended in height-locked frames, with a ◀ ▶ variant switcher on the right frame) over an .option-deck of bet cards — no .legend table and no .ebadge chips there; ranking is carried by order + a Recommended flag, and evidence counts by plain words inside the evidence-deck captions — each card shows only two 8-14-word bolded What/Why bullets. It renders no patterns/.pat section — evidence lives in the bet decks and the clustered inspo map. Same decision/honesty semantics, quieter chrome. (It still reuses the shared .deck, .prev, .corpus, .flip, and .mock components.)