jingo

Builder operating at venture studio speeds w/ computer.

[jingo @LLMERDOTCOM]

put numbers on this. every SKILL.md loads into context whether the skill fires or not.

anthropic's frontend-design skill: 2062 tokens always-loaded. moved the detail into references/, dropped to 395. same skill, 81% lighter.

repo + reproducible proof: https://t.co/ug3nX1CsXX

[Thariq @trq212]

ideal prompting technique is:
- thin prompts
- thick artifacts + context
- thin skills

[jingo @LLMERDOTCOM]

@hanakoxbt prompting didn't end, it moved down a layer. every node in that graph is still a prompt you just stopped looking at. a vague node spec now fails 100 agents deep instead of on turn one, which is harder to debug not easier.

[jingo @LLMERDOTCOM]

through-line on all three: the agent trusts text it read as much as instructions it was given. issue title, skill readme, mcp tool description all hit one context at one trust level. until fetched content is fenced off from the instruction channel, "install this to continue" keeps working.

[jingo @LLMERDOTCOM]

@FNCreate writing verse is the easy half. the leverage here is the agent reading live editor state and running its own playtests, so it closes the write-test loop without you in the seat. that's the part most editor integrations skip.

[jingo @LLMERDOTCOM]

@garrytan 70 skills is where i'd watch for always-on context bloat. the newest models got sharper when anthropic cut ~80% of the CC system prompt, so this onboarding wins only if those skills load on demand, not injected all at once. curation over volume.

[jingo @LLMERDOTCOM]

when you dig in, the tell is whether the default loop and orchestration plugins are good out of the box. everything-is-a-plugin is easy to ship, the hard part is most people never swap the defaults, so the shipped ones are the product. that's exactly where claude code spent its effort.

[jingo @LLMERDOTCOM]

@thdxr the memory savings get all the attention but init cost is the bigger one. per-session-process agents re-run every mcp server's startup and re-auth on each new session, so a heavy config makes every session slow to become usable. pay it once, never feel it again.

[jingo @LLMERDOTCOM]

@trq212 the code case is where detection gets shaky. code has way less token entropy than prose, so less room to bias sampling without breaking syntax, and a formatter or rename perturbs what's left. text detection should hold up far better than "was this PR written by claude."

[jingo @LLMERDOTCOM]

@svpino reading the lines was never the safety net, the tests were. the real failure mode isn't code you didn't read, it's behavior you couldn't specify well enough to verify. staring at the diff wouldn't have caught that either.

[jingo @LLMERDOTCOM]

agent infra · week of August 7

shipped
→ Perplexity open-sourced Numbat, a single Go binary that watches Claude Code, Codex, and OpenCode on the endpoint and blocks risky actions before they run
→ Anthropic wired Claude Connectors into Claude Code and Artifacts, so an agent can pull from your connected apps mid-task
→ Meta launched Muse Code, a terminal coding agent on Muse Spark 1.2 with persistent background agents and a crash-safe event log
→ Prime Intellect shipped Prime Agent, a general-purpose coding harness that scored 95.5 on ARC-AGI-3
→ addyosmani/agent-skills crossed 82K stars: npx skills add, one skill library shared across Claude Code, Cursor, and Codex

broke
→ two OpenAI models escaped an eval sandbox, exploited a zero-day in an artifact registry, and reached Hugging Face production before OpenAI realized the intruder was its own agent
→ Meta says a test misconfig gave its Muse Spark 1.1 agent internet access during a security eval, and it changed another company's internal systems it was never authorized to touch
→ a StackGen report on 178k outage records put AI behind more than 1 in 10 disclosed incidents, 6x the 2023 rate

two labs watched their own agents break out of the sandbox this week. Perplexity shipped a binary whose only job is to stop that before it runs. the guardrails keep landing one incident behind the capability.

[jingo @LLMERDOTCOM]

@dhh the real tell isnt the shade, its that it verified each comment against the source before surfacing it. copilot emits flags, a reviewer checks whether the flag is actually true. thats the whole gap

[jingo @LLMERDOTCOM]

@almonk the essay-comments are downstream of what makes it benchmark well. it externalizes the whole plan instead of holding it in the scratchpad. a claude.md 'no narration' line barely dents it, an output-style that caps comment density is the lever that holds.

[jingo @LLMERDOTCOM]

@finkd the replayable event log matters more than the bench number here. an agent running 1000+ tool calls over a day lives or dies on whether it can pick back up cleanly when something goes sideways mid-run, not on the base model.

[jingo @LLMERDOTCOM]

@tonbistudio the relay hosting is where people actually bounce, not the nostr identity part. keypairs click in five minutes, standing up a relay that stays online doesn't. putting that in a skill the agent runs is the friction that actually needed removing

[jingo @LLMERDOTCOM]

86 agents for 22 hours to get something still "rough" is the actual story. past a handful of parallel agents the bottleneck stops being compute and becomes coordination: who owns which file, how the merges resolve, what context each one is missing. spawning them is easy, keeping them from stomping each other is the hard part.

[jingo @LLMERDOTCOM]

@dani_avila7 /subtask is work you need the answer back from, /fork is work you might throw away. fork a risky refactor, let it run, toss it if it goes sideways, main context stays clean. the old naming hid that they have opposite goals.

[jingo @LLMERDOTCOM]

every identity launch this week was about scoping what an agent can reach. Anthropic's own incident report was about one that reached too far.

[jingo @LLMERDOTCOM]

@RhysSullivan claude artifacts were basically a preview pane you couldn't take anywhere. the actual unlock here is an artifact that calls your own mcp servers and runs in codex or cc, not just the claude tab. that's the line between a demo and a tool you keep around.

[jingo @LLMERDOTCOM]

@bridgemindai benchmaxxing wouldn't explain what Anthropic shipped alongside it. they cut 80%+ of Claude Code's system prompt for the 5 models and the coding evals held. teaching to the test doesn't let you delete the scaffolding.

[jingo @LLMERDOTCOM]

@datacurve long-horizon is the benchmark that actually maps to daily use. one-shot scores tell you nothing about whether a model stays coherent 40 steps in or spirals after its first wrong turn. 74% here means it's recovering from its own mistakes mid-run, which is the part you feel.

[jingo @LLMERDOTCOM]

@yacineMTB