# Don't Use Next.js Without These 7 Skills > [!question] Lingering Questions > - What exactly counts as "success" in Supabase's ~60% MCP-only figure — task completion, absence of a security flaw, or something else? The video gives no methodology. > - Does the "bottleneck is context, not capability" claim replicate outside a sponsored Supabase segment, or is it vendor marketing dressed as a finding? > - Is there a practical ceiling on how many skills one project can stack before they degrade each other's effectiveness? > [!tip] Article Ideas > - "Agents are lazy about reading reference files" as a design constraint for any agent-facing documentation (CLAUDE.md, skill files, runbooks) — cross-domain beyond Next.js. > [!link] Cross Reference > - [[3. Resources/Permanent Notes/Context Not Constraints Is What Unhobbles an Agent]] — same "bottleneck is context not capability" claim, independently arrived at from a different source > - [[3. Resources/Permanent Notes/Harness Design Can Matter More Than Model Choice for Agent Performance]] — skills-as-harness-design is a specific instance of this general claim > - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — on-demand context loading as an emerging pattern in AI-agent tooling > [!quote] Vocabulary > - *skill* — per this source, "a few files with some text about a particular topic" an AI agent loads into context on demand, rather than a permanent system-prompt addition > - *cached components / `use cache`* — Next.js's unified caching directive; the migration target of Skill 2 > - *static shell* — the portion of a page render that's identical for every visitor and can be served instantly while personalized content streams in ## Bib capsule A screen-recorded YouTube walkthrough surveys skills.sh (a Vercel-built leaderboard of AI-agent "skills") and picks roughly seven the presenter considers must-haves for a 2026 Next.js + Supabase stack: browser-based self-verification, cached-components migration, a static-shell optimizer, partial pre-fetching, React composition patterns, Supabase's own agent skills, and two design-QA skills (Impeccable, shadcn UI). ## Author's thesis AI coding agents benefit more from small, composable, on-demand-loaded "skill" files than from either a bloated system prompt or bare tool/MCP access alone — the bottleneck for agent output quality is context delivery, not underlying model capability. ## Key arguments (≤3) 1. **On-demand context beats always-on context** — a skill is markdown the agent pulls in only when its name/description matches the task, avoiding permanent prompt bloat while still making the guidance available. 2. **Tool access (MCP) alone underperforms tool access plus a skill** — Supabase's cited internal data shows ~60% success on tasks like RLS policy generation with MCP alone, rising "significantly" once a skill is layered on, because agents default to stale training data rather than calling an available docs-fetch tool. 3. **Browser-based verification catches a defect class code checks structurally cannot** — the "next dev loop" skill demo caught a real contrast bug (build/type-check-clean, visually broken) by giving the agent an actual browser to screenshot and inspect. ## Claim atoms ### Skills load context on demand, not always-on > [!quote] ID:rw_gv7KrPTCW_8_skill_definition > A skill is basically just a few files with some text about a particular topic, so that if you prompt your AI agent... it can load that text into the context. Opening framing (~0:22). Core reusable mental model behind every skill discussed afterward; generalizes past Next.js to any AI-agent-skill design decision. ### Browser-based agent verification catches contrast bugs code checks miss > [!quote] ID:rw_gv7KrPTCW_8_contrast_catch > The browser check caught one real polish issue. The muted copy was a little too light. ~1:22–7:00, next dev loop skill demo. Concrete evidence that visual/browser verification finds a defect class (contrast, layout) that build/type checks structurally cannot. ### MCP tools alone underperform skills for correct behavior > [!quote] ID:rw_gv7KrPTCW_8_context_not_capability > MCP alone is not enough without workflows or guidelines, which is often described in skills... So, the bottleneck is context, not capability. ~15:32–17:00, Supabase agent-skills segment. Vendor-reported (partnered/sponsored segment) but mechanistically plausible: agents skip available docs-fetch tools and default to stale training data unless a skill nudges them otherwise. Treat the ~60% figure as unverified. ### Skill authors should inline essentials, not defer to reference files > [!quote] ID:rw_gv7KrPTCW_8_inline_essentials > They do recommend keeping the essentials directly inside skill.md. The agents are lazy about reading reference files. Put critical knowledge where they can't miss it. ~17:00–18:00, closing note on the Supabase segment, framed as general skill-authoring advice. Directly actionable for anyone authoring internal docs/skills meant for AI agent consumption. ## Permanent candidates - **Write agent-facing docs assuming the agent won't fetch the reference file** — strongest, most portable claim; not tied to Next.js at all. Promoted below. - The "bottleneck is context, not capability" framing overlaps with an existing vault Permanent note ([[Context Not Constraints Is What Unhobbles an Agent]]) — connect rather than duplicate. ## Processing checklist - [x] Check whether an existing Permanent note already covers the "context not capability" claim — yes, [[Context Not Constraints Is What Unhobbles an Agent]]; linked instead of duplicated. - [ ] Treat Supabase's ~60%-vs-higher figure as vendor-reported (partnered segment) until independently verified. - [x] Connect into the relevant MOC ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — on-demand skill loading as a concrete instance of the vault's broader agent-tooling-design thread - [[3. Resources/Readwise/Articles/Don't Use Next.js Without These 7 Skills]] — source document - [[3. Resources/MOCs/AI Agent Safety MOC]] — applies-in: the MOC lists this note under Persistent instruction & steering.