Summary

A fifteen-minute tools round-up in which Jack Roberts walks through seven GitHub repositories and services he says pair well with GPT-6 Astra and Claude Fable 5.1, demonstrating each on screen. The argument is entirely by demonstration rather than evidence: each tool is shown working once, in the presenter's own environment, and recommended on that basis. Two of the seven are his own or promoted products, and the video carries a mid-roll pitch for his paid course, so the selection should be read as a creator's showcase rather than a survey. Its value to this archive is as a snapshot of what the agent-tooling layer looked like in mid-September 2026 — specifically the appearance of security scanning for agent skills, and of model-versus-model cross-checking as a named practice.

• NVIDIA SkillSpector is presented as a repository that scans agent skills before installation for hidden instructions, unsafe access requests and data-leak paths, returning a risk rating. Roberts reports it had roughly 17,000 GitHub stars at the time of recording — a figure he states on screen and does not source further. He demonstrates running it across his own installed skill inventory.
• The existence of SkillSpector is the more durable observation here than the tool itself: prompt-injection and over-broad permissions in third-party agent skills are being treated as a distinct security problem with its own scanning tooling, rather than as a general AI-safety topic.
• Roberts describes a 'Claude + Codex loop' as a working practice: one model drafts a plan, the second reviews it for weak spots before anything is built. He claims each model catches mistakes the other misses because every model has blind spots regardless of quality, and says he uses the loop for perhaps 10% of his AI work — the complex builds where the decision matters.
• He names Claude Fable 5.1 and GPT-6 Astra as 'the two best models in the world right now'. This is his unsupported opinion, offered without benchmarks, and should be read as one creator's ranking rather than an established position.
• DeepSeek is reported to have released its own agent harness — the environment around a model, as distinct from the model itself. Roberts uses a car-and-driver analogy: the harness is the car, the model the driver, and any model can in principle run in any harness. He says the DeepSeek harness can reach files, a terminal, tools and a browser, and is fully customisable.
• His stated reason for using it is cost — he describes the DeepSeek V4.1 model as 'so cheap' — and hitting token limits elsewhere. No pricing figures are given.
• The harness-versus-model distinction is the reusable idea in this entry. It separates two things usually discussed as one, and it explains why the same model can perform differently in different products.
• 'God's Eye View' is shown as an interactive 3D globe aggregating public data feeds — flights including military flights, ships, satellites, earthquakes and public cameras — with a conversational interface. Roberts demonstrates asking it to fly to the Pentagon, switch to a cockpit view on a nearby aircraft, and play a local radio station. It aggregates independent public sources rather than providing any private data.
• 'Slop Monster' is Roberts's own repository, declared as such. It scores text on vocabulary, AI sentence constructions, punctuation cadence and rule-of-three rhythm, then rewrites vague copy into specific copy — his worked example turns 'trusted, reliable and built to last' into 'six nails per shingle, every shingle'. He says he built it by combining existing anti-slop repos and adding a sales-and-marketing dimension he found missing from all of them, citing Steve Krug, Daniel Priestley and Alex Hormozi as sources.
• 'Floot' is a commercial service, not a repository, and Roberts flags that himself. It takes an interface designed inside a ChatGPT or Claude conversation and publishes it as a working app, supplying the authentication, database, forms and notifications. His framing is that the back end is the part people avoid, and that chat-native site publishing does not cover it. A free tier is mentioned; no pricing is given.
• 'Archifi' converts a plain-English description into an interactive architecture or process diagram, with several display modes and export. Roberts demonstrates it mapping a sales system and argues its real use is onboarding, handoff and tracing a request through an unfamiliar codebase.

Why it matters

Three things in this entry outlast the tool list. First, security scanning of agent skills now exists as its own category with a major vendor's name on it, which says the threat of a malicious or over-permissioned skill has moved from theoretical to tooled. Second, the harness-versus-model distinction gives the Handbook a cleaner way to talk about why the same model behaves differently in different products — it is a vocabulary gain, not a news item. Third, deliberate model-versus-model cross-checking is being described by practitioners as routine practice rather than as diligence, which reinforces the existing operating principle about cross-checking claims against a second source or model. Everything else here is perishable and promotional: the tool names, the star counts and the model ranking are one creator's snapshot of a fortnight, and two of the seven items are things he sells or built.

9a6oCFbs0O8-transcript.txt