AI
LLM Context Limits and the Smart Zone
Every LLM starts sharp and gets dumber as the context fills. Attention relationships scale quadratically with tokens, and Pocock puts the degradation point at roughly 100k regardless of whether the model advertises 200k or 1 million. Past that threshold, decisions get sloppy. The fix is small tasks that stay inside the smart zone, not longer contexts.
“They have a smart zone and a dumb zone.” — Matt Pocock
Compacting squeezes a session into a summary and carries it forward. Pocock prefers clearing entirely. Every fresh start returns to the same clean system prompt, which makes behavior predictable and keeps sediment from accumulating across sessions.
Alignment Before Code: The Grill Me Skill
The “specs to code” approach passes a requirements doc to AI, then fixes the doc when output is wrong. Pocock tried it and rejected it.
“because the code is your battleground.” — Matt Pocock
His alternative: start every feature with /grill-me. He ran the live grill me session on gamification feature (16:04) live, watching a sub-agent spend 93.7k tokens on Opus before firing its first question: “Points economy: what actions earn points and how much?” Sessions run 40 to 100 questions. The output is not a plan, it is a shared understanding with the model.
From Grilling Session to PRD
After the grilling session, Pocock had 25k tokens of alignment conversation. He ran /write-a-prd, which produced a destination document with a problem statement, 18 user stories, and implementation decisions. He did not read it. LLMs summarize well, and after a grilling session the shared design concept is already in place, so the PRD is just that concept written down.
The skill also proposes modules to modify. Pocock checked that list, confirmed the gamification service was the right deep module to build, and got a local markdown issue file in seconds.
Vertical Slices and the Kanban Board
AI defaults to horizontal coding: all database work first, then API, then front end. That delays feedback until phase three, when you finally have an integrated system to test. Pocock’s fix is vertical slices. Each issue cuts through all layers and leaves something visible at the end, like “award points for lesson completion visible on dashboard.”
The /prd-to-issues skill generates independently grabbable tickets with blocking relationships. Two tasks with no shared dependency can be grabbed simultaneously by separate agents, turning the kanban board into a parallelizable directed acyclic graph.
The AFK Agent Loop: Handing Off the Night Shift
Planning is human-in-the-loop. Implementation can run AFK. Pocock demonstrated the ralph-once.sh AFK run (57:40), a bash script that reads all issue markdown files, grabs the last five commits for context, and runs Claude with permission-mode except-edits. One iteration produced a gamification service with 284 passing tests.
TDD keeps the agent honest. The /red-green-refactor skill tells Claude to write a failing test first, confirm red, then write code to pass. Implementation and review should never share a context: the reviewer in the dumb zone misses bugs the implementer would catch fresh.
Codebase Architecture as the AI Quality Ceiling
Shallow modules, many small files with tangled dependencies, are hard to test and hard for agents to reason about. Deep modules, from John Ousterhout’s A Philosophy of Software Design, expose a small interface and hide complexity inside. One test boundary wraps the whole thing. Pocock’s /improve-codebase-architecture skill scans a repo and surfaces consolidation candidates.
“without that you just end up with slop” — Matt Pocock
Sandcastle, a TypeScript library Pocock spent a week building, handles parallelization. The Sandcastle parallel agent walkthrough (90:32) showed a planner selecting issues, spinning up Docker worktrees per issue, running Sonnet for implementation and Opus for review, then merging all branches with a third agent.
Q&A
Now that 1 million token context windows exist, do we ever actually want to use them? Pocock said the smart zone is still around 100k; larger windows mostly shipped more dumb zone, which helps retrieval tasks but not coding. ▶ Watch (37:24)
How do you deal with agents producing more code than you can review, and how do you parallelize properly? He said more code review is unavoidable when you delegate implementation to agents, and he does not yet have a good answer for keeping pull requests small while running multi-issue loops. ▶ Watch (58:58)
How do you get the agent to code the way you want, using the libraries and standards you want? Pocock distinguishes push (coding standards baked into the reviewer prompt so they are always present) from pull (standards available as skills the implementer can fetch when needed). ▶ Watch (1:27:41)
Notable Quotes
They have a smart zone and a dumb zone. Matt Pocock · ▶ Watch (11:01)
because the code is your battleground. Matt Pocock · ▶ Watch (13:39)
without that you just end up with slop Matt Pocock · ▶ Watch (1:13:29)
Key Takeaways
- Keep LLM context under ~100k tokens; clear and restart rather than compact to avoid sediment degrading output quality.
- Run a grill-me session before writing any code to reach a shared design concept — alignment beats planning documents.
- Structure issues as vertical slices with blocking relationships so agents get cross-layer feedback fast and can run in parallel.