cost-optimizer

name	cost-optimizer
version	1.0.0
description	Advises on the cheapest richapi tool path for a stated goal. Consulted before spending — turns "I want 500 CTOs with emails" into a cost-accurate plan, shows comparative costs for every viable path, and flags anti-patterns (loops, over-verification, redundant enrichment). Use when the user asks "what's the cheapest way", "how much will this cost", "I have N credits left — what can I do", or before any big-spend job.
benefits-from	["richapi-gtm"]
allowed-tools	["Bash","Read","AskUserQuestion"]
triggers	["cheapest way to","how much will this cost","I have N credits left","budget-conscious","save credits","cost estimate"]

cost-optimizer

The credit-math skill. Never guesses — always grounds in _lib/mcp-catalog.json.

Preamble

~/.claude/skills/richapi-gtm-skills/bin/richapi-skills-preflight

Read _lib/mcp-catalog.json in full before answering. Never cite a credit cost from your training data — it will be wrong.

Common goals → cost-ranked paths

Goal: "Find N prospects matching an ICP"

Path	Cost / result	Catch
A — people_search	0.1	Indexed cache; less precise on live LinkedIn state. Cheapest.
B — profile_search	0.1	Live LinkedIn, exact-string matching. No seniority/industry bands.
C — lead_search	0.5	Full 30+ filter surface, include+exclude, sessionId pagination. Use only when B can't express the ICP.

Rule of thumb: A < B < C. Always ask "can A or B satisfy this?" before picking C. If the ICP mentions seniority bands, company headcount bands, or exclusion criteria → it's C.

Goal: "Find emails for N people"

Path	Cost	When
email_finder (waterfall, sync)	≤2 per success, 0 on hard-fail	N ≤ 5, or you need the result in the current turn
find_emails (batch) + check_email_finding	2 per success, polling is free	N ≥ 10, or you have ≥100 to do

Rule: Never loop email_finder for >5. Always batch for >5.

Cost note: Both paths charge only on success (or soft-fail). A 100-person batch where 60% hit = 120 credits, not 200.

Goal: "Verify deliverability on N emails"

Path	Cost	When
email_verifier (waterfall, sync)	≤1 per verified, 0 on hard-fail	N ≤ 5
verify_emails (batch)	0.5 per email	N ≥ 10

Rule: Batch is cheaper per-unit and faster wall-clock for any list >10.

Goal: "Enrich N LinkedIn profiles"

Path	Cost	When
enrich_profile (singular)	1 per call	N = 1
enrich_profiles_bulk (up to 50)	1 per result	Always, for N ≥ 2

Rule: Never loop enrich_profile for N > 1. Chunk into 50s and use bulk.

Goal: "Understand a company"

Path	Cost	When
enrich_company	1	Firmographics only
web_tech_stack	2	Just tech
website_intelligence	5	Everything web — tech + pixels + emails + social + SSL
website_intelligence + enrich_company	6	Full picture

Rule: For 2+ angles on one site, website_intelligence always wins.

Budget-constrained mode

If the user says "I have X credits, what can I do":

1. Call check_usage (free) — confirm the actual balance.
2. Enumerate options fitting within X:
   • X // 0.1 = max people_search results
   • (X - 5) // 2 = max emails you can find after 1 enrich+site scan
   • etc.
3. Present 2–3 realistic scenarios, not a menu of every possibility.

Anti-pattern detector

Scan the user's plan for these before running:

Pattern	Fix
Loop of enrich_profile for many profiles	Switch to enrich_profiles_bulk
Loop of email_finder for 20+ people	Switch to find_emails batch
Verifying emails before classifying	Classify first (0.5), drop roles/disposables, then verify only work+personal
Running lead_search when people_search satisfies the ICP	Downgrade to people_search (5× cheaper)
web_tech_stack + web_pixels + web_social_links together	Collapse to one website_intelligence (5 credits total vs 5 combined — break-even on 3 tools; cleaner output with wi)
ad_details on every ad from ad_search	Cap at top 3–5 by engagement/recency
Polling check_email_finding at 500ms intervals	Use exponential backoff starting at 3s
Re-running search_reference_data every session	Cache locally — it's static catalog data

Output shape

Give the user a ranked cost table + a recommended plan:

Goal: "Find 50 VP Sales at SaaS companies 200–500 in Bay Area, with verified work emails"

Ranked paths:
  Path A (lead_search + enrich_profiles_bulk + find_emails + verify_emails)
    50 × 0.5  + 50 × 1 + 50 × 2 × 0.7_hitrate + 35 × 0.5 = 152 credits
  Path B (people_search instead of lead_search) — NOT VIABLE
    people_search can't express "Director+" seniority band; skip.
  Path C (skip verification — rely on find_emails confidence score)
    50 × 0.5 + 50 × 1 + 50 × 2 × 0.7 = 135 credits (−17, but ~5% bad addresses)

Recommended: Path A. 152 credits. Budget safety: <1% of a 25k-cap plan.
Fire? [confirm]

When to refuse a plan

If the user's plan is obviously wasteful and they haven't been informed:

1. Show the cheaper path + cost delta.
2. Ask confirmation before proceeding.
3. If they insist on the wasteful path, proceed — but note it in the final report.

Never silently let them burn credits.

No-call reporter

This skill does no paid MCP calls — only check_usage (free) if needed. It's pure planning.