Competitive benchmark

Competitive Website Benchmark Skill

A three-phase competitive teardown: Benchmark every site on a shared rubric → Synthesise Steal / Differentiate / Avoid → optionally Hand off to build.

This skill is an orchestrator. It owns the comparative logic (matrix + strategy) and pulls in specialist skills for craft depth rather than reinventing them.

---

Step 0 — Gate: Choose Audit Depth (ASK FIRST, DO NOTHING ELSE YET)

This is a hard gate. Before fetching, detecting, or scoring anything, present the two depth modes to the user and wait for their explicit choice. Do not begin analysis until they answer. Frame the choice descriptively so they pick with full information:

⚡ Fast Benchmark Uses this skill's own rubric across every site. Each competitor and your own site scored on all seven axes from a direct read of the pages. You get the full comparison matrix and the Steal / Differentiate / Avoid synthesis. You won't get the granular craft teardown — exact hex values, type-scale diagnosis, line-by-line copy critique. Quick, broad, strategic. Best for a first read or when there are many competitors.

🔬 Deep Teardown Everything in Fast, plus the skill orchestrates specialist audits on every site: visual-aesthetics-audit for design (typography, palette, spacing, with fix values), ux-writing analysis for copy, and the persuasion lens for conversion hooks. Far richer and more defensible, but heavier — for your site + 3 competitors that's ~12 deep passes before synthesis. Slower and more thorough. Best when you're committing to act on it.

Throttle note: If the user picks Deep Teardown AND there are 4+ competitors (5+ total sites), warn them about run length and offer to either (a) run Deep on their own site + the top 2-3 competitors and Fast on the rest, or (b) proceed full Deep anyway. Respect their call. For 3 or fewer competitors, honour Deep directly without friction.

Use ask_user_input_v0 to present the Fast / Deep choice as tappable options.

---

Step 1 — Gather & Validate Inputs

Once depth is chosen:

1. Confirm which URL is the user's own site and which are competitors. If unclear, ask — the whole synthesis hinges on knowing whose side you're on.
2. Fetch each site. Use Claude in Chrome if available (screenshots at 1440px desktop + 375px mobile, page text, accessibility tree). Fall back to web_fetch for HTML/text, and note that visual analysis is inference-based.
3. For each site capture: hero/above-the-fold, primary nav & IA, key CTAs, pricing or conversion path, social proof elements, footer. These are the comparison surfaces.

If any site is unreachable or behind auth, flag it and proceed with what's available.

---

Step 2 — Detect Business Model (Auto, Ask Only If Unclear)

Infer the model from the sites themselves — it sets the evaluation lens. Read → references/business-models.md for detection signals and per-model "what good looks like."

Model	Tell-tale signals	What "hooking the user" means
B2B	Demo/contact-sales CTAs, ROI/case studies, integrations, "trusted by [logos]"	Credibility, proof, low-risk next step (demo, trial)
B2C	Direct signup, broad appeal copy, app-store badges, lifestyle imagery	Speed to value, emotional resonance, social proof
D2C	Product catalogue, add-to-cart, shipping/returns, brand story	Aspiration, frictionless purchase, brand identity
Marketplace	Two-sided messaging (buyers + sellers), listings, trust/safety	Liquidity signals, trust between strangers

Detect per-site (competitors may differ from the user — note it if so). If the model is genuinely ambiguous after reading the sites, ask the user once. Otherwise state your detected model and proceed.

---

Step 3 — Benchmark Every Site on the Seven Axes

Score the user's site and every competitor on the same seven axes, through the lens of the detected business model. Read → references/benchmark-rubric.md for the full scoring criteria per axis.

#	Axis	Evaluates
1	Visual Design & Brand Expression	Aesthetic quality, cohesion, distinctiveness, craft
2	UX Writing & Messaging Clarity	Copy clarity, value articulation, microcopy, CTA quality
3	Brand Voice & Positioning	Distinct voice, sharp positioning, "why us" clarity
4	Conversion & Persuasion Hooks	How effectively the page drives the target action (model-specific)
5	Information Architecture & Journey	Navigation logic, findability, flow to conversion
6	Social Proof & Trust Signals	Reviews, logos, testimonials, security, credibility cues
7	Differentiation Clarity	Is it obviously different from the others, or category wallpaper?

Scoring: Each axis gets a 🔴/🟡/🟢 RAG + score /10, per site.

Deep Teardown orchestration (only if Deep was chosen)

For each site, invoke the specialist skills and fold their findings into the relevant axis:

• Axis 1 (Visual) → run visual-aesthetics-audit logic; pull its six-dimension diagnosis and fix values into this axis.
• Axis 2 (UX Writing) → apply ux-writing evaluation to hero copy, CTAs, microcopy, error/empty states.
• Axis 4 (Hooks) → apply the Trust & Persuasion deep lens from the ux-audit skill (its references/deep-lenses.md — Fogg + Cialdini). Flag dark patterns explicitly — a competitor using fake scarcity is not a strength to steal; it's a trap to avoid.

In Fast mode, score all seven axes from this skill's own rubric without invoking the sub-skills.

---

Step 4 — Build the Comparison Matrix

Output a side-by-side matrix: axes as rows, sites as columns, RAG + score in each cell. The user's own site is always the first/anchor column, visually distinguished.

| Axis | YOUR SITE | Competitor A | Competitor B | Competitor C |
|---|---|---|---|---|
| Visual Design | 🟡 6 | 🟢 8 | 🟡 5 | 🟢 9 |
| UX Writing | 🔴 4 | 🟢 8 | 🟡 6 | 🟡 6 |
| ... | | | | |
| **Overall** | **5.1** | **7.8** | **5.6** | **7.2** |

Immediately below the matrix, a one-line positional read: where the user ranks, and their single biggest gap and single biggest existing strength relative to the field.

---

Step 5 — Synthesise: Steal / Differentiate / Avoid

This is the payload. Read → references/synthesis-method.md for the full method.

For each axis (or grouped where it reads better), produce three lenses:

• 🟢 STEAL — A competitor does this measurably better; it's adoptable without becoming derivative. Name the competitor, the specific thing, and how to adapt it (not copy it).
• 🔵 DIFFERENTIATE — Where every site looks the same. Sameness in a crowded category is a liability; this is the gap the user can own. Bias the synthesis toward this — differentiation beats imitation. Be concrete about the opening.
• 🔴 AVOID — A mistake one or more competitors make (clutter, dark patterns, weak positioning, broken journeys). Name it so the user doesn't inherit it.

The cardinal rule: Do not produce a "become more like your competitors" report. The most valuable output finds the unoccupied position — where the user can be deliberately different and better, not a cleaner clone. If the whole category shares a blind spot, that blind spot is the strategy.

End the synthesis with The One Move: if the user changed only one thing about their site to compete, what is it and why.

---

Step 6 — Structure the Report

# Competitive Website Benchmark — [Your Brand] vs [N] Competitors
**Date:** [today] · **Business model detected:** [model] · **Depth:** [Fast / Deep]
**Sites analysed:** [your site] (anchor) + [competitors]

## Executive Summary
- Where you rank: [X of N]
- Biggest gap: [axis + one line]
- Biggest existing strength: [axis + one line]
- The One Move: [single highest-leverage change]

## Comparison Matrix
[the seven-axis × all-sites table with RAG + scores + overall row]
**Positional read:** [one line]

## Per-Axis Findings
For each of the 7 axes:
- The matrix row restated
- What the leader on this axis is doing (named)
- Where the user sits and why
- [Deep mode] specialist findings (visual fix values / copy critique / persuasion read)

## Strategic Synthesis
For each axis or grouped theme:
- 🟢 STEAL — [competitor] does [X]; adapt as [Y]
- 🔵 DIFFERENTIATE — everyone does [X]; you could own [Y]
- 🔴 AVOID — [competitor] / the category does [X]; don't inherit it

**The One Move:** [the single highest-leverage play]

## Dark Pattern Watch (if any found)
[Competitor persuasion tactics that are manipulative, not admirable — flagged so the user
doesn't copy them]

## Caveats
[Unreachable sites, auth walls, inference-based visual notes, dynamic states not audited]

## Next Steps (handoff offer)
[See Step 7]

---

Step 7 — Deliver & Hand Off

• Save as competitive-benchmark-[your-brand]-[date].md.
• Use present_files to surface it.
• Then offer the build handoff (this skill does NOT build — it briefs):
  • Want to act on this? → Offer to package the findings into a brief for the product-requirements skill to produce a build/redesign PRD.
  • Want deeper market/positioning/GTM strategy? → Hand off to the market-intelligence skill (business model, pricing, channel strategy — the layers this skill deliberately doesn't cover).
  • Want the visual direction spelled out? → Offer a full visual-aesthetics-audit design brief for the user's own site.
• Also offer: re-benchmark after changes; deep-dive a single competitor; convert findings into a prioritised action backlog.

---

Orchestration Principles

• Ask depth first, always. Never start analysis before the Fast/Deep gate is answered.
• Anchor on the user. Every score, every synthesis line exists to answer "so what should I do?" — competitor analysis is a means, not the product.
• Differentiation over imitation. Push the user toward an unoccupied position. A report that just says "copy the leader" has failed.
• Same lens for everyone. Score every site on the identical rubric through the same business-model lens, or the comparison is meaningless.
• Flag manipulation, don't admire it. Dark patterns are AVOID findings, never STEAL.
• Don't reinvent — orchestrate. In Deep mode, route to the specialist skills; in Fast mode, use the built-in rubric. Don't duplicate their logic inline.
• Hand off, don't sprawl. Building the thing is product-requirements' job; deep market strategy is market-intelligence' job. This skill benchmarks and directs — then passes the baton.
• Be honest about inference. Visual/motion findings from web_fetch alone are estimates. Say so.