SOURCE NOTES / SEPTEMBER 2026
The experience
behind the advice.
This guide combines my supplied prompt-history analysis, my account of two years of building, and current official documentation. They answer different questions.
What the personal counts establish
They show how often I used a habit or the scale of my work. They do not prove the habit caused better results, and they have not been independently audited for this page. Quotes are reproduced from the two supplied working-note documents. Examples and copyable prompts are editorial templates, not transcripts or tested guarantees.
| Count | Meaning and source |
|---|---|
| 95K prompts · 100B+ tokens | Rounded two-year usage figures supplied by Paul. Tokens indicate processing volume, not unique words or original content. |
| 100+ agents | Total deployments over two years across tools including OpenClaw, Hermes, Paperclip, and GrokBot. Not the number active now, and not helper dispatches. |
| 1,407 unattended sessions | Summer sessions reported as having no human prompt. A session is not an automation, a completed task, or a saved work hour. |
| 68% fewer prompts | Reported June-to-August decline. The chart shows only those endpoints, indexed to June = 100 and August = 32. High commit activity is a separate qualitative observation; no commit chart is invented. |
| 186K search appearances | 102K Bing AI citations + 84K Google generative AI impressions across Paul’s sites. An editorial aggregate of different platform counters and potentially different periods, not unique reach or 186K citations. |
| Habit counts | 9,000 short prompts, 476 plan-first prompts, 139 used skills, 617 verification prompts, 832 compactions, and 1,444 screenshot-led prompts are quoted from the supplied analyses. Their categories may overlap. Counts document use, not effectiveness. |
| Hundreds of automations | Paul’s qualitative account. No exact total is derived from the 1,407-session count. |
| “One app” | An instructional starting point on the connector guide, not an observed statistic. |
Why some stronger-sounding numbers are absent
The notes report review hit rates, total findings, and rejected findings. Those findings are not presented here as independently proven bugs. Likewise, declining prompt counts are not translated into productivity multipliers or hours saved. The practical value is in the workflow readers can try and inspect.
Tool guidance changes
The official sources below were consulted in September 2026. Product documentation does not prove a reader’s installed version, permissions, subscription, or connected account supports the same behavior. The prompt workbench explicitly asks for a fresh documentation check before model-specific prompts are written.
- Anthropic: prompting best practices ↗
- OpenAI: prompt engineering ↗
- Claude Code: project memory ↗
- Claude Code: skills ↗
- Claude Code: best practices ↗
- Composio: sessions via MCP ↗
- Pipedream: MCP for end users ↗
- Hermes: profiles ↗
- Hermes: scheduled tasks ↗
- Hermes: security ↗
- OpenClaw: multi-agent routing ↗
- Google: AI features and your website ↗
- Paul: published AI citation research ↗
Public tools
The Second Brain, Voice Profile Builder, and Get Cited by AI repositories were checked as public GitHub projects. Inspect their instructions before using them. The Bing citation research has its own frozen study datasets and limits; a newer headline total does not automatically update every research claim.
Saved guides use this browser’s local storage. Prompt-workbench text is not saved or sent. This guide has no account system and does not send workbench inputs to any service.
Choose a guide ↗