namesmith

Expert Mindset

The goal is not to generate impressive variety — it is to surface 5 names the user will actually want to register. Keep this in mind throughout.

Key calibrations that non-experts get wrong:

• The challenger rejects ~50% of names by design. A 60-70% rejection rate is normal and healthy. If the challenger rejects 90%, the names were too generic. If it rejects only 20%, the criteria may not have been applied strictly enough.
• .com is almost always taken for good coined words — bias toward .io, .ai, .app as primary targets from the start, not as fallbacks. Short coined names (≤6 chars with unusual consonant clusters) are the exception — those are often unclaimed.
• "Another round" is not random regeneration. Each round should incorporate rejection patterns as steering signals. Read the most common failure dimension from the challenger output and adjust the archetype mix accordingly — if distinctiveness kept failing, shift away from metaphorical/evocative toward coined.
• Name quality is multi-dimensional. A name scoring 7/10 across all five dimensions beats a name scoring 10/10 on memorability and 3/10 on context fit.
• B2B names and B2C names should feel different. B2B buyers want precision and credibility — short coined words and descriptive names perform better. B2C buyers want warmth and story — metaphorical, evocative, and compound names work better. If the user hasn't specified, infer from the product description.
• Include one "safe anchor" name per round. Every list should have one descriptive-but-memorable name that clearly communicates what the product does. Users who reject all the creative names often accept the anchor — and it sets the tone against which they evaluate the others. Without an anchor, users feel unmoored by too much novelty.

---

NEVER

• NEVER show the internal name list to the user before Phase 4. The challenger must run first. Showing unfiltered names undermines the entire workflow and exposes weak candidates.
• NEVER generate more than 4 names from one archetype. If 12 out of 18 names are coined words, the distribution is broken and the challenger will lack diversity to filter from.
• NEVER retry the MCP domain probe more than once. A second retry on a failed probe adds delay without new information. Go straight to fallback.
• NEVER loop "another round" more than 3 times without asking the user to reconsider their brief. After 3 rounds, the problem is usually an unclear or over-constrained brief, not insufficient generation.
• NEVER present challenger scores to the user unprompted. Users want name options, not a branding exam transcript. Share scores only if the user explicitly asks why a name was filtered out.
• NEVER include a name you have already identified as culturally problematic in the list passed to the challenger. Filter these out in Phase 1 — the challenger does not recheck cultural safety.
• NEVER skip deduplication before Phase 2. Near-identical names waste challenger capacity and confuse domain results.
• NEVER use "another round" to fix a threshold problem. If the user wants more names because the threshold is too strict (they liked a rejected name), that is a filtering discussion — adjust the threshold in the current results rather than generating 15 new names.

---

Phase 0 — Understand the Business

Before asking questions, extract as much as possible from $ARGUMENTS. Most users provide more signal than they realize — read between the lines:

• "pet care startup" → B2C, warmth tone, compound/evocative archetypes will resonate
• "fintech SaaS for CFOs" → B2B, precision tone, short coined/descriptive archetypes will resonate
• "AI-powered recipe app" → consumer, playful tone, coined/short words work

If $ARGUMENTS contains a clear description of what the product/service does, who it serves, and what tone is desired (≥ 20 words covering these points), proceed to Phase 1 immediately.

Otherwise, use AskUserQuestion to gather:

1. What does your product or service do? (one sentence)
2. Who is the target customer?
3. Preferred tone: professional / playful / technical / human / other?
4. Any words, concepts, or languages to include? Anything to avoid?

Accept short answers — do not prompt for more detail than needed.

---

Phase 1 — Generate Name Candidates

MANDATORY READ before generating: Load ${CLAUDE_PLUGIN_ROOT}/references/naming-criteria.md — use Section 1 only (Generation Archetypes).

Do NOT load mcp-fallback.md or Section 2 of naming-criteria.md at this phase — they are for later phases.

Generate 15–20 name candidates that cover all 6 archetype types. This phase is high freedom — the archetype constraints prevent clustering, not artistry. Apply full creative judgment within them.

The 5-second test: For each name you generate, ask: "If someone heard this name at a conference badge, would they still remember it by end of day?" If not, the name needs more distinctiveness before you include it.

Archetype	Example	Max names
Invented / coined word	Kodak, Xerox, Etsy	4
Compound word	Dropbox, GitHub, Snapchat	4
Metaphorical / evocative	Amazon, Oracle, Stripe	3
Descriptive-but-memorable	Basecamp, Mailchimp	3
Short coined (≤6 chars)	Uber, Lyft, Fiverr	3
Domain-hack friendly (root ≤8 chars)	del.icio.us style	3

Rules:

• No two names may be phonetically or visually near-identical (e.g. Veltora / Veltara)
• Each name must be 3–14 characters total
• Bias toward names where .io or .ai is plausibly unclaimed (.com is assumed taken for most good coined words)
• Apply cultural/linguistic safety rules from naming-criteria.md Section 1 — exclude any name that fails these checks before passing to the challenger

Store internally as a list: [name, archetype, 1-sentence rationale]

Do NOT display this list to the user yet.

---

Phase 2 — Sub-Agent Challenge

Launch an Agent call to challenge and score all names from Phase 1.

Pass the following in the prompt:

• The business description (from Phase 0)
• The full name list: each entry as Name | Archetype | Rationale
• The full text of Section 2 from ${CLAUDE_PLUGIN_ROOT}/references/naming-criteria.md

The agent is defined at ${CLAUDE_PLUGIN_ROOT}/agents/name-challenger.md. Instruct the agent to follow its scoring rubric and produce the required structured output.

Parse the agent's output:

• Extract all names with Verdict: KEEP
• If fewer than 5 names are KEEPed, take the top 5 by score from the full list (relaxed threshold)
• Note in Phase 4 output if the threshold was relaxed

survivors = the final filtered list (5–15 names)

---

Phase 3 — Domain Availability Check

MANDATORY READ before probing: Load ${CLAUDE_PLUGIN_ROOT}/references/mcp-fallback.md.

Do NOT re-read naming-criteria.md at this phase.

This phase is low freedom — follow the probe-then-fallback pattern exactly. Do not improvise alternatives to the probe step.

Probe: Attempt one search_domains call for the first name in survivors, with tlds: [".com"].

If the probe succeeds (MCP healthy):

For each name in survivors, call check_domain_availability for .com, .io, .co, and .app.

• If .com is taken for a name, also call generate_domain_variations to surface creative alternatives
• Process up to 5 names concurrently

If the probe fails (tool not found, HTTP error, or timeout):

• Do not retry
• Display the UNAVAILABLE or DEGRADED notice from mcp-fallback.md
• Proceed to Phase 4 with domain data marked as — (manual check)
• Include the manual check URL per the fallback template

---

Phase 4 — Present Results

Display a results table:

## Business Name Candidates

| Name      | Type               | .com | .io | .app | Notes                          |
|-----------|--------------------|------|-----|------|--------------------------------|
| Veltora   | Coined             | ✓    | ✓   | ✓    |                                |
| NestRun   | Compound           | ✗    | ✓   | ✓    | .com taken; nestrun.io free    |
| ...       | ...                | ...  | ... | ...  | ...                            |

If domain check was unavailable, replace availability columns with a single Domain column containing the manual check URL from mcp-fallback.md.

If the challenger threshold was relaxed, add a note: * Threshold relaxed to top 5 by score — consider running another round for stronger candidates.

Then use AskUserQuestion to offer three options:

1. Explore one name further — deep-dive with domain variations
2. Run another round — generate fresh candidates (informed by what was rejected and why)
3. Done — end the session

If option 1 (Explore): Ask which name. Call generate_domain_variations for that name across multiple TLDs and suffix patterns. Present the variations with availability status. If generate_domain_variations is unavailable, list common TLD alternatives manually (.ai, .co, .app, .io, .dev, -hq.com, get[name].com).

If option 2 (Another round): Summarize the rejection reasons from the challenger output. Identify the most common failure dimension (e.g., "5 names failed distinctiveness, 3 failed context fit"). Adjust generation accordingly — if distinctiveness was the top failure, reduce metaphorical/evocative names and increase coined. Return to Phase 1.

If option 3 (Done): End cleanly. Do not summarize unless the user asks.

---

Edge Cases

Scenario	Action
MCP tool not found	Follow mcp-fallback.md UNAVAILABLE template
MCP returns HTTP error or timeout	Follow mcp-fallback.md DEGRADED template; do not retry
All names rejected by challenger	Take top 5 by score; note threshold was relaxed
Challenger output is malformed / unparseable	Treat all names as KEEP with score 5; add note about parsing failure
User provides < 20 words of context	Phase 0 clarifying questions handle this
Generation produces near-identical names	Deduplicate before Phase 2
generate_domain_variations tool unavailable	Use manual TLD alternatives listed in Phase 4 Explore path
Agent tool unavailable	Perform challenger scoring inline using the rubric from naming-criteria.md Section 2
User hits "another round" 3+ times	Surface the pattern: ask user to revisit the brief before generating again