# bshp.io

> Personal site for Ben Bishop. Focus: multi-agent systems, frontier model evaluation, and the tools built around that work.

## Links

- [Developers Overview](/developers): All machine-readable surfaces
- [OpenAPI](/openapi.json): Public read API specification
- [API Index](/api/v1): Public read API root
- [Projects JSON](/api/v1/projects): Static file: /api/v1/projects.json
- [Articles JSON](/api/v1/articles): Static file: /api/v1/articles.json
- [Stack JSON](/api/v1/stack): Static file: /api/v1/stack.json
- [RSS Feed](/feed.xml)
- [Sitemap](/sitemap-index.xml)
- [GitHub](https://github.com/BenBish)

## Projects

- [astro-llms](/llms/projects/astro-llms.md): An Astro integration that generates curated llms.txt agent surfaces and per-entry Markdown mirrors from your Content Layer collections. No HTML scraping, no Starlight lock-in.
- [grok-plugin-cc](/llms/projects/grok-plugin-cc.md): A Claude Code plugin that delegates code review, adversarial review, and coding rescue tasks to Grok through xAI's official Grok Build CLI.
- [Mission Control](/llms/projects/mission-control.md): A React + Express + SQLite dashboard that ingests Claude Code, Codex CLI, Hermes, ComfyUI, and Lemonade telemetry into one place - sessions, activities, consumption, runtime health, failures, jobs, and generations. Run the hub on a central host; attach collectors on the machines where you do AI work.
- [local-model-plugin-cc](/llms/projects/local-model-plugin-cc.md): A Claude Code plugin that delegates code review, adversarial review, and coding rescue tasks to local models through the codex CLI.
- [Better Bet](/llms/projects/betting-picks.md): Sports betting picks management with an agent API, real-time SSE activity feed, and analytics dashboard.
- [Model Prompt Tests](/llms/projects/model-prompt-tests.md): A Markdown library of prompts for evaluating models across coding, planning, writing, and reasoning - each with a scoring rubric.
- [Archie Orchestrator](/llms/projects/archie-orchestrator.md): An OpenClaw plugin giving agents task lifecycle, queue tracking, and review gates. Includes a detailed account of where it broke down in real end-to-end testing.

## Articles

- [Four Flash Models on Practical Prompt Tests](/llms/articles/flash-tier-shootout-benchmark.md): GLM-5.3-Flash led this four-model benchmark at 4.83 and had the lowest completed-response latency, while Qwen3.8-Flash had the lowest recorded cost.
- [How much can you build on an 11-hour flight?](/llms/articles/how-much-can-you-build-on-an-11-hour-flight.md): I reconstructed one LHR to SFO flight from Mission Control, git, and Linear: 5 hours 46 minutes of active work, 27 merged PRs, and several limits the logs cannot explain.
- [I finally hit the SuperGrok limit](/llms/articles/i-finally-hit-the-supergrok-limit.md): What it took to use a full $30 SuperGrok weekly allowance: 31 coding-agent sessions, 9.66 million input tokens, and a lot of deliberate work.
- [DeepSeek V4 Flash vs Pro on Practical Prompt Tests](/llms/articles/deepseek-v4-flash-vs-pro-benchmark.md): DeepSeek V4 Pro scored 4.50 to Flash's 4.38 in one practical prompt-test batch, but cost about 21 times as much for candidate answers.
- [Grok Imagine: a small hands-on image review](/llms/articles/grok-imagine.md): Four create and edit calls through Grok Build image tools: what held up for decorative site assets, what this sample does not prove, and when I still reach for code or real screenshots.
- [July 2026 agent usage, measured](/llms/articles/july-2026-agent-usage.md): A July extract from the Fedora Mission Control hub: 402 sessions across hosted agents and local OpenCode with Qwen 3.6 plus Hermes, alongside $14.46 of separate Direct API Spend.
- [Local Gemma vs Grok 4.5 and Sonnet 5 on Practical Prompt Tests](/llms/articles/local-gemma-vs-frontier-benchmark.md): A model-prompt-tests run that put a local Gemma candidate next to Grok 4.5, Sonnet 5, and GPT-5.5: peer scores, latency reality, and what a high local score does not prove.
- [astro-llms: llms.txt from Astro content collections](/llms/articles/astro-llms.md): A content-collection-first Astro integration for curated llms.txt agent surfaces. Why HTML scraping is the wrong default, how setup works, and what the build emits.
- [Mission Control: provider spend, budgets, and burn rate](/llms/articles/mission-control-spend-budgets.md): Mission Control now pulls account-level billing from provider APIs, surfaces Direct API Spend on the homepage, and adds budgets, burn-rate forecasts, and spend alerts without mixing them into agent session costs.
- [Kimi sold out: capacity-constrained AI demand](/llms/articles/kimi-capacity-constrained.md): Moonshot AI's Kimi subscriptions sold out. That is a capacity signal, not a demand problem: when labs turn away paying users, compute cost is the binding constraint.
- [OpenAI model naming and independent release lines](/llms/articles/openai-model-naming.md): Why version-coupled model names break down as product lines multiply, what Anthropic's shift away from shared version labels fixed, and a guess at independent OpenAI lines with their own ladders.
- [Adding Grok support to Mission Control](/llms/articles/adding-grok-support-to-mission-control.md): How Grok Build CLI became a first-class source in Mission Control: desktop collector, source filter, sessions, activities, and honest status lights.
- [Grok Build for Code Review in Claude Code](/llms/articles/grok-plugin-cc.md): A look at grok-plugin-cc, a Claude Code plugin that routes review and rescue workflows through xAI's Grok Build CLI, with setup, safety gates, and test results.
- [Mission Control: unified AI usage after the OpenClaw pivot](/llms/articles/mission-control.md): What Mission Control is today - a multi-source usage and runtime dashboard - and how it was rebuilt from an OpenClaw-only activity feed into collectors for Claude Code, Codex, Hermes, ComfyUI, and more.
- [Grok 4.5 vs Sonnet 5 on Practical Prompt Tests](/llms/articles/grok-45-vs-sonnet-5-benchmark.md): A benchmark writeup comparing Grok 4.5 and Sonnet 5 across 13 practical prompt tests, with methodology, caveats, and raw artifacts preserved in model-prompt-tests.
- [Local Models for Code Review in Claude Code](/llms/articles/local-model-plugin-cc.md): A look at local-model-plugin-cc, a Claude Code plugin that routes review and rescue workflows through the Codex CLI and local OpenAI-compatible model servers.
- [Benchmarking Model Prompt Tests](/llms/articles/model-prompt-tests-benchmarking.md): Introducing the first benchmark runner for model-prompt-tests: a small harness for running prompt suites across model providers, scoring outputs against rubrics, and publishing comparable reports.