Quotes — Using skills to pay the bills

Verbatim quotes from transcript.md. Speech-to-text artifacts preserved; likely intended terms noted in brackets only when needed for searchability.

On skills sprawl

• [sprawl: customer voices] "We have a plethora of skills distributed across so many repositories that it's become incredibly difficult and almost impossible to get your head wrapped around how to unify, centralize and properly distribute skills across all teams." — quoted from a lead architect (Section: Skills sprawl — definition)
• [sprawl: wild west] "It's a free form right now for the most part. Everyone's been doing their own thing. It's the wild west, right?" (Section: Skills sprawl — definition)
• [sprawl: scale] "we have 2 million plus skills across GitHub across 44,000 repos. It's an insane amount." (Section: Skills sprawl — definition)

On the failure modes

• [drift] "newer skills are shipping. And teams aren't keeping up with those newer versions." (Section: The five failure modes)
• [rot] "having an outdated skill can often be just as bad as having no skill at all." (Section: The five failure modes)
• [overloading] "there is actually a limit to the amount of skills you can have with most coding agents, often expressed as a percentage of the context window… Most agents will truncate that description, which means you end up in a situation where your skills might not be activated because the description has been shortened." (Section: The five failure modes)

On the agentic equation

• [equation] "an agent is a function of the modern [model], the harness and the context" (Section: The agentic equation)
• [lever] "context is effectively the best lever that we can pull to affect the outcome of the agent here." (Section: The agentic equation)
• [why skills] "skills again feel like an important powerful piece because they can encode business or domain specific logic that the model won't have been trained on." (Section: The agentic equation)

On skills-as-software practices

• [decompose] "Don't write monolithic skills, right? Don't write huge epic skills. Think about decomposing breaking skills down into smaller pieces." (Section: Decompose)
• [extend, don't edit] "Very tempting to just edit them directly… Please don't do that. Your teammates won't like you." (Section: Extend, don't edit)
• [avoid global] "This is the works on my machine problem. Skills that are installed on a developer's machine live outside of the repo… Two different people run the same task against the same code base and get very different output because one has the skill and the other one doesn't." (Section: Avoid global skills)
• [Boris Cherny, quoted by Moss] "we ban any local setup. All agent workflow improvements, hooks, skills, etc. Must be checked into the repo for everyone." (Section: Avoid global skills)
• [registry as single source of truth] "you have this one sort of truth. Everyone's pulling the same notion of the scale, same conversion." (Section: Publish to a registry)
• [supply-chain / minimum release age] "the malware here was notable because… it had a level of persistence where it added a global clawed code hook to your machine to try and keep itself installed on the user's machine. Right. So another good reason not to use any kind of global state coding skills." (Section: Publish to a registry — re: the "mini sha loot tack" / likely Shai-Hulud incident)
• [agent-agnostic] "the agents we're using in six months might look nothing like today's… Skills are effectively your durable asset. The agent is just the room type [runtime]." (Section: Don't lock to one agent)
• [team asset] "making them safe to contribute right if people feel like a skill is highly dependent on [or relied upon] by many others in the company they might not feel safe contributors." (Section: Make context a team asset)

On evals

• [Adam Savage, quoted by Moss] "the difference between screwing around [and] science is [writing] it down" (Section: Measure with evals)
• [evals as experiment] "if you can't measure it… you can't improve it." (Section: Measure with evals)
• [do you even need the skill?] "Does the skills still work? Do you even need the skill? Is the agent[,] sorry[,] has the model just got better and you no longer need to worry about kind of documenting some of these processes…" (Section: Measure with evals)

On the CDLC and the 2005 analogy

• [CDLC] "Everything you do to keep code healthy already has a direct equivalent for context right you probably just haven't really been doing it." (Section: CDLC)
• [we know the ending] "the good news is we already know how the story ends right we don't have to spend another 10 years coming up these processes of rediscovering those we can apply what we've learned with the sdlc to the CDLC right we just have to make sure we're not making the same mistakes again." (Section: 2005 PHP story)

From the Q&A

• [non-technical users] "I don't think [there] is a good answer expecting folks to you know start running CLI commands in the tunnels [terminals] locally I think is going to be difficult." (Section: Q&A, Audience member 2)
• [dependencies between skills] "there isn't a standard way to understand it [at] the moment[;] dependencies [is] something that I think everyone's a bit scared to tackle… maybe just the agent being able to look at the skill and see hey I need to use this other skill by reading it" (Section: Q&A, Audience member 4)