Thirty skills in, three kept: a crypto investor's personal analyst setup

Who built it

0xJeff, a crypto/DeFi investor who writes a Substack. This is part two of his series on using Hermes as a personal analyst; his stated aim is to speed up his learning and free time for higher-value work such as talking to founders.

How it works

The first kept skill fetches and analyzes X posts. Cron jobs run each morning, follow a set of accounts on macro, geopolitics, tech and AI, and produce a daily report. The author started with a cookie-based command-line X client but removed it after finding it could post to his timeline even when he wanted read-only behavior, and moved to the official X API v2 with clearer permissions.

The second skill works through his X bookmarks. Hermes pulls 10-20 bookmarks a day through the API, uses a browser tool to open X long-form articles the API cannot return, ranks about 15 against his interests, summarizes them and files the insights in a wiki, turning a pile of saved posts into a searchable record.

The third skill, Reflect, depends on Hindsight, an external memory provider installed as a Hermes plugin. Hindsight keeps full conversation turns and offers a reflect tool that synthesizes across stored memories. The author uses it to produce a daily 'top 5 insights' list with reasons each matters to him, drawn from patterns in his past sessions and preferences.

Around these three, the setup also uses market-data sources (CoinGecko, DeFiLlama), a social-activity skill, and Hermes's built-in delegation for side tasks. The author notes that skills and cron jobs multiplied quickly to around 30, which made configuration changes and debugging harder.

Cost was the recurring theme. He reports about $0.50 a day in X API credits, and a single day where the Hindsight memory setup, routed through OpenRouter to Claude Sonnet 4.6, burned more than $50. In his words, "Self-learning + self-remembering make Hermes very useful", but he treats the bill as something to watch closely.

Lessons you can use

• Expect to try many skills and keep few. Measure each against a real daily need and retire the rest, because skill and cron sprawl makes debugging harder.
• Prefer official APIs with explicit, limited permissions over cookie-based tools that can act as you; a read-only intent in the prompt is not a permission boundary.
• Memory providers call a model in the background. Check which model your memory provider uses and cap it, or a cheap-looking feature can become the biggest line item.
• Use browser automation only where the API cannot reach, in a fresh isolated browser profile without sensitive logins.
• Send results somewhere durable (a wiki or notes folder) so the daily digest builds into a record rather than disappearing in chat.

Try a smaller version

1. 1

   Write one skill that summarizes a fixed list of sources you already read, then attach it to a morning cron job and check a week of output before adding anything else.

2. 2

   Use the web search and extraction tools first; add browser tools only for the pages they cannot read, and use a separate browser profile.

3. 3

   If you want cross-session reflection, try the built-in memory before an external provider; if you add one (for example via hermes memory setup), point it at a low-cost model and check spend after the first day.

4. 4

   Have the digest write to a dated file in a notes folder so you can compare days and judge whether the skill is worth keeping.

5. 5

   Review your skill and cron list every couple of weeks and remove jobs you no longer read.

Caveats

• All costs and usefulness claims are one user's anecdotal reports and were not verified.
• The author says Hermes sometimes hallucinated during X fetches, including producing random content, so outputs needed checking.
• The article does not list the 30 skills, state where Hermes ran, or which chat channel delivered the reports; nothing here is financial advice.

Spotted something wrong or out of date? Tell us on X at @HermesWatcher↗ (opens in a new tab) and we will fix it.