scout-agent

Scout Agent — Domain Intelligence Protocol

> The intelligence layer of the Creator Engine. Researches a domain BEFORE any product is built. > Transforms "I want to build something for X" into a data-backed recommendation of WHAT to build.

---

Inputs

• `domain`: The domain query (e.g., "kubernetes security", "email marketing automation")
• `slug`: Sanitized slug derived from domain (e.g., "kubernetes-security")
• `creator`: Loaded creator.yaml profile (type, level, language, expertise)
• `language`: Output language (from creator.yaml or auto-detected)
• `intent`: personal | marketplace | explore — determines report depth and sections

Output

• `workspace/scout-{slug}.md` — Complete scout report

---

STEP 1: BASELINE TEST — What Claude Already Knows

**Purpose:** Establish the FREE alternative. What does any user get from Claude vanilla, without products?

**Method:** 1. Generate a comprehensive response AS IF you were Claude with no special products, answering: > "You are an expert in {domain}. Explain the key concepts, common approaches, best practices, > common pitfalls, and how a professional would approach this domain." 2. Write this response verbatim — this IS the baseline 3. Be honest and thorough — a weak baseline makes the gap analysis meaningless 4. Adapt depth to creator profile:

• **developer** → include code patterns, tooling, debugging approaches
• **prompt-engineer** → include cognitive patterns, prompt structures
• **domain-expert** → include what Claude ADDS beyond their existing knowledge
• **marketer** → include conversion patterns, audience psychology
• **operator** → include automation, monitoring, scaling patterns
• **agency** → include multi-client patterns, customization
• **hybrid** → balanced across all dimensions

**Output section:** `## 1. Baseline — What Claude Knows`

**Quality gates:**

• 300-600 words. Shorter → domain may be too narrow. Longer → consider decomposing.
• Must be substantive enough to stand alone as useful content (it IS the competition).
• Rate the baseline: Strong (80%+ coverage) / Moderate (50-80%) / Weak (<50%).

---

STEP 2: GAP ANALYSIS — Where Claude Falls Short

**Purpose:** Identify what's MISSING, SHALLOW, WRONG, or GENERIC in the baseline.

**Method:** Read the baseline critically through 4 lenses:

| Lens | Question | Example finding | |------|----------|----------------| | **Missing** | What topics weren't mentioned at all? | "No mention of supply chain attacks in k8s security" | | **Shallow** | What was mentioned but lacks depth? | "RBAC mentioned but no real-world role design patterns" | | **Wrong** | What common misconceptions appear? | "Implies network policies are enabled by default" | | **Generic** | What lacks real-world specificity? | "Best practices are textbook, no production war stories" |

**For each gap, rate severity:**

• **Critical** — Without this, Claude's advice could cause harm or significant waste
• **Significant** — Noticeable quality difference; expert would spot this immediately
• **Minor** — Nice to have; improves experience but baseline is functional

**Output section:** `## 2. Gap Analysis`

**Quality gates:**

• Minimum 5 gaps. Fewer than 3 → baseline was too generous or domain too narrow.
• More than 15 → ask creator to narrow scope before proceeding.
• Each gap must be specific and actionable (not "could be more detailed").
• At least 1 gap per lens. If a lens has 0, explicitly state "No {lens} gaps found."

**CRITICAL: After this step, RETURN CONTROL TO SKILL.MD for the mid-protocol gate.** The creator must see and approve the gaps before research begins.

---

STEP 3: MARKET SCAN — What Already Exists

**Purpose:** Know the competitive landscape before building.

**Method:** 1. Check if `myclaude` CLI is available:

   myclaude --version 2>/dev/null

2. If available, run market queries:

   myclaude search "{domain}" --json 2>/dev/null
   myclaude search "{related-term-1}" --json 2>/dev/null
   myclaude trending --json 2>/dev/null

3. If CLI unavailable, note it and continue — market scan is valuable but not blocking.

**For each competing product found, capture:**

• Name, type, author, downloads
• What it covers (from description)
• What it DOESN'T cover (gap relative to our analysis)
• Price point (if marketplace intent)

**If intent = "explore":** Lighter analysis — list products found, note obvious gaps, skip deep comparison.

**Output section:** `## 3. Market Landscape`

**Quality gate:** Even with no results, document the absence — "No existing products found for {domain}" is itself a valuable market signal (blue ocean or non-obvious niche).

---

STEP 4: RESEARCH — Fill the Gaps with Real Intelligence

**Purpose:** Transform identified gaps into actual knowledge using external research.

**Prerequisite:** WebSearch and WebFetch tools must be available. If not:

• Write to report: "Research skipped — no web tools available. Recommendations based on baseline + gap analysis only. Confidence: medium."
• Skip to Step 5.

**Method:** 1. For each **Critical** gap from Step 2, run targeted research:

   WebSearch: "{domain} {gap-topic} best practices 2025 2026"
   WebSearch: "{domain} {gap-topic} common mistakes production"

2. For each **Significant** gap, run focused research:

   WebSearch: "{domain} {gap-topic} expert guide"

3. For the most promising results (relevance + recency), fetch full content:

   WebFetch: {url} — extract key patterns, heuristics, examples

4. Synthesize findings per gap:

• **Key patterns** — things experts do that Claude didn't mention
• **Common mistakes** — things beginners do that experts warn against
• **Expert heuristics** — decision rules that only come from experience
• **Real examples** — case studies, war stories, production incidents
• **Source** — URL and date (unsourced claims stay in baseline, not research)

**Research budget:*