Kappaemme-git/codex-first-customer-finder-skill: does the agent reach for it? 2026-10-11
Twelve Codex skills were installed together and each agent was sent three requests meant for this skill (first-customer-finder), in a throwaway sandbox. A run stopped as soon as a skill was opened.
Then the same requests with 100 and 500 other real skills installed beside the twelve. Codex ran gpt-6-astra, Claude Code ran Claude Opus 5.5.
All levels · The requests · How a run is scored
| Request | Codex, 0 others | Codex, 100 others | Codex, 500 others | Claude Code, 0 others | Claude Code, 100 others | Claude Code, 500 others |
|---|---|---|---|---|---|---|
| asked plainly Research public evidence to identify 10 likely early adopters for my AI invoice-reconciliation app, rank them with reasons and links, and draft personalized outreach messages. Do not send anything. | opened it 7s | opened it 19s | opened it 9s | opened it 3s | opened it 4s | opened it 3s |
| described without the skill's words Search public sources for 10 organizations that appear to struggle with reconciling invoices and could benefit from my app, then provide the evidence, a ranked list, and tailored introductory emails. Do not contact anyone; use this sample product description: “LedgerLoop matches invoices, purchase orders, and receipts and flags discrepancies automatically.” | opened it 6s | opened it 7s | no skill opened time limit | opened it 3s | opened it 4s | opened it 3s |
| in the other language 请根据公开证据，为我的人工智能发票对账应用寻找10个最可能尝试它的早期使用者，按匹配度排序并附上理由和链接，同时起草个性化联系邮件。不要发送任何消息。 | opened it 7s | opened it 8s | opened it 7s | opened it 3s | opened it 3s | opened it 3s |