Coding Agents & Claude Code

• How prompt caching works in Claude Code (and how to stop wasting tokens) — practical breakdown of where Claude Code burns cache hits and where it doesn't; directly actionable for anyone paying per-token in long sessions.
• The "dumb zone" demo: /rewind + /compact to watch the model get smarter — empirical proof that context degradation starts ~120K tokens; rewinding to a clean state then compacting measurably restores output quality.
• Claude Code plugin enforcing a real dev lifecycle: branch → commits → PR draft — hooks Claude Code into proper git hygiene automatically rather than relying on prompt discipline.
• Open-source guardrail for Claude Code PRs: green CI ≠ proof if the agent edits the verifier — addresses the adversarial loop where an agent passes its own tests by modifying them; critical for unattended agentic runs.
• App Claude drives, not an app that calls Claude — architecture discussion — inversion pattern where the LLM owns the control loop rather than being a subroutine; useful framing for agentic SaaS design.
• Lore — give your coding agent the decisions your team already made — typed Markdown decision log served read-only over MCP to Claude Code/Cursor so agents stop re-litigating ruled-out choices.
• GitHub Copilot harness benchmarked against native model harnesses on SWE-bench — Copilot's harness matches model-vendor harnesses in task resolution while using fewer tokens; harness quality matters as much as model quality.
• /research skill design discussion — background agent saves to .md, continues main task — pattern for non-blocking research delegation inside Claude Code; the verbatim description already works, a skill just canonicalises it.
• Multi-skill invocation behavior: /skill-1 /skill-2 do XYZ — which fires? — behaviour regressed twice in three weeks (both → one → both → one); signals Claude Code's skill dispatch is not yet stable.
• Agent error loops: how to actually handle them — community-sourced recovery patterns for when Claude Code or Cursor spins on the same failure repeatedly.
• MCP server giving an AI agent its own crypto wallet, built with Claude Code — demonstrates MCP as the composability layer for agents needing external capabilities (wallet signing, not just read tools).
• MCP for HuggingFace publish workflows — feedback wanted before expansion — early signal of MCP servers targeting ML publication pipelines; worth tracking for agentic MLOps tooling.

LLM Mechanics: Token Spend, Context & Model Routing

• Claude's effort tiers vs Gemini 5.6's model proliferation problem — Claude's explicit effort ladder (low→ultracode per model) is more navigable than Gemini's opaque model×effort matrix; argument for transparent capability signaling.
• Gemini 5.6: Sol/Terra/Luna × effort levels — how to choose? — cross-model effort routing is becoming the dominant UX problem as model families proliferate; no good answer yet.
• Tokenmaxxing is dead, long live tokenmaxxing — argues that blanket token-increase strategies masked comprehension failures; real fix is targeted context, not volume.
• Keep AI spend flat while token usage grows: better defaults, routing, caching — not spend alerts — Coinbase's approach: default to open-weight models (GLM 5.2, Kimi 2.7) via LLM gateway, opt-up to frontier; cost control through routing not friction.
• 5-hour limits should be context-driven, not time-driven — rate limiting by elapsed time punishes slow tasks; argument for context-window exhaustion as the natural throttle signal.
• Switching from Gemini to Claude: model/effort selection for quick vs. deep tasks — practical community routing heuristics for Haiku/Sonnet/Opus × effort; useful baseline for cost-aware pipeline decisions.

Agentic SaaS & Remote Workflows

• levelsio: coding almost solely on VPS with Claude Code for ~a year — VPS + Claude Code as a device-agnostic, always-on dev environment; /goal runs overnight unattended; normalises server-side agentic coding as production practice.
• levelsio follow-up: does he push directly to prod, preview env, does he review the code? — thread surfaces the real workflow gap: no preview runner mentioned, code review unclear; honest look at how raw the VPS-agent loop is.
• How do you send screenshots to an agent while coding on cloud? — unresolved UX problem for remote Claude Code: visual feedback loop to the agent breaks when there's no local browser; no clean solution yet.
• e-ink smart clock handling multiple agent messages and notifications — ambient agent status display on e-ink; early hardware pattern for monitoring unattended agentic runs without a screen.

Vibe-Coding & Repos

• browser-use/video-use — edit videos with Claude Code, 100% open source — drop raw footage, chat with Claude Code, get final.mp4; handles filler-word cuts, dead space, auto color grading; extends browser-use's agent-drives-tool pattern to video pipelines.