『Fill Your Context Window』のカバーアート

Fill Your Context Window

Fill Your Context Window

無料で聴く

ポッドキャストの詳細を見る
Episode 3 of The Skill Tree is a hosts-only conversation about the tooling forming around AI coding agents — code review rewritten as narrative, MCP notifications, portable agent bundles, and what has to be in place at a company before any of it pays off. Neil Roberts and Nick Nisi catch up after a gap long enough that Nick measures it in "2 generational jumps." The title comes from Nick's inversion of a familiar worry: "We're always worried about filling its context window, but you can use it to fill your context window." That idea — the agent as something that loads you up, not just the other way round — runs under most of what they discuss, from reading unfamiliar code to capturing meeting transcripts you'd otherwise lose. In This Episode What code review looks like when an agent narrates the story of a diff instead of walking you through it file by fileThe MCP feature almost nobody uses — notifications, which Claude calls channels — and why pushing context into a running session mattersWhether "loops" is a new idea or a rename for orchestration people have built for 18 monthsWhat happened when an agent DM'd a coworker on Nick's behalf and invoked the CEO to demand credentialsWhy context, not tools, is the prerequisite for AI adoption — and Nick's formula of a champion plus a real budget Episode Chapters 00:00 Cold Open: The 20% Flip00:21 Two Generational Jumps, and Fable Comes Back02:45 AI Engineer World's Fair and the World Cup Suite06:09 dad: Code Review as a Bedtime Story10:53 Code Spelunking and Matt Pocock's zoom-out11:49 MCP Notifications, or What Claude Calls Channels14:52 Vercel's eve, Convex's adam, and Portable Agent Bundles20:43 Getting Better at AI Without Better Models23:27 What People Call Loops, We've Been Doing for Years25:37 rankduel.app: Tuning a Pairwise Sort29:52 Wikilinks, OpenWiki, and Markdown as a Database34:05 sessions, Lost Research, and sessions wrapped35:42 ideation Gets an express Mode38:03 Wayfinder: When You Feel a Little Bit Lost41:38 TARS, and the Time Claude DM'd a Coworker45:22 Context Before Tools: Getting Agent-Ready49:51 Granola, and Bringing Conversations Back to Slack52:53 Champion Plus Budget56:50 Outro Code Review as the New Bottleneck Nick's framing is blunt: writing code stopped being the constraint, "and you know what the problem is? Everyone else can too." Reviewing everyone else's agent-written code is where the time goes now, and every diff feels like a drive-by. His answer is dad, a Bun CLI at diff.dad that installs as a daemon, lists the PRs you're tagged on, and uses your Claude or Codex account to narrate the change — grouped by logical boundary rather than file boundary, with comments synced both ways to GitHub. Neil's counterpoint is that the isolation cuts both ways. Fewer engineers touch any given piece of code now, so fewer people understand it — but a couple of hours with AI tools can leave you knowing a codebase better than weeks of working in it used to. Pushing Context Into a Running Session The most technical stretch covers a corner of the MCP spec Nick found while building an unshipped dad feature: notifications, which Claude exposes as channels. MCP is a pull model, and notifications invert it, letting information arrive in a session that's already running. Nick wants to comment on his working directory and have those comments land in his live Claude session, not a PR that doesn't exist yet. It works on a branch, but it needs org-level enablement and a special CLI flag, so it isn't shippable. The idea that interests him is filtered error logs pushed into a session as they happen. That thread continues into packaging. Neil walks through Vercel's eve, where an agent is a directory of skills, MCP servers, instructions, tools, and channels, and Convex's adam, which runs the same shape elsewhere. He doesn't much care about either as a product — he wants the format: something he can zip up, hand you, and have you run in whatever harness you prefer. Loops, and Whether They're New Neil has been reading the loop discourse with some skepticism. What people call loops, he argues, is what he and Nick have done for over a year: skills and tools and MCP servers, a script that checks for success, a bit of memory, and iteration. "It's using AI, but they started calling it loops." Nick defends the verification cycle — action, check, feed the failure back in — and traces it to Ralph loops. The disagreement is friendly and never quite resolves. Neil's example is the sharper part: he pointed a goal-driven loop at the pairwise sort behind rankduel.app, armed with a scoring harness measuring what he actually cared about. The model tried every sort it could think of and reported that his was still the best — but surfaced a bug in his range calculation that made the sort measurably better anyway. Context Before Tools The last third turns to adoption. Neil's position is that tools are downstream of everything else: you can't pick tools ...
adbl_web_anon_alc_button_suppression_t1
まだレビューはありません