Borrowed Intelligence (Drain It into Plans)
Treat intelligence as borrowed. Drain it when it is available: use the most capable model now to build a catalog of self-contained plans, then implement later with a cheaper, open-source, or self-controlled model. The plan is the product. Source: User, 2026-06-13
The thesis
Frontier intelligence is rented, metered, and temporary - the best model this month is expensive and may not be the one you can run next month. So spend it where intelligence compounds (understanding a codebase, judging what is worth doing, writing the spec) and hand execution - which is increasingly commoditized - to cheaper, local, or owned models. shadcn frames it exactly: "use your most capable model for the part where intelligence compounds ... and hand execution to cheaper models. The skill never implements anything itself. The plan is the product." Source: shadcn/improve README, github.com/shadcn/improve, 2026-06-13
Why it compounds
- Judgment is the scarce input; execution is falling in price. A plan written by today's expensive model gets cheaper to run every month as open models catch up. Banking plans now locks in today's judgment at tomorrow's execution cost. Pairs with Compounding Systems.
- Plans are durable and portable. They are plain markdown, so any agent or human - across Cursor, Claude Code, Codex, a local model - can pick them up. The bank survives model churn.
- It separates the two bottlenecks. Understanding/verification (hard) is done once by the strong model; generation (easy) is repeated cheaply. This is the same split Browser Testing Skills and Harness Engineering make.
The artifact: a plan catalog
A backlog of plans/NNN-slug.md specs, each written for the weakest plausible
executor - a model that never saw the planning session. Properties that carry
that (from Frontend and Design Skills):
- Self-contained - inlined file paths, current-state excerpts, repo conventions with an exemplar, verified build/test/lint commands. No "as above."
- Verification gates - every step ends with a command and its expected output; done criteria are machine-checkable.
- Hard boundaries + STOP conditions - explicit out-of-scope and "if X, stop and report," so a small model does not improvise when reality diverges.
- Drift stamp - the git commit the plan was written against, for a mechanical pre-execution drift check.
Kevin's catalog lives at the repo-root plans/ directory (see plans/README.md).
Where it fits in the loop family
| Phase | Owner | Skill / page |
|---|---|---|
| Audit + judge + write the plan (expensive) | strongest model | Frontend and Design Skills, borrowed-intelligence (this) |
| First-run-UX audit that produces plan candidates | strongest model | Browser Testing Skills |
| Execute + verify + ship (cheap) | cheaper/owned model | Browser Testing Skills |
| Re-run / refresh the backlog | standing loop | Agent Looping (/improve reconcile) |
| Score output quality over time | judge | The Eval Loop (Slop Is an Output Problem) |
| Pick what becomes a plan and what ships | human | Staying in the Loop with Agents |
This session is a worked example: Browser Testing Skills was built as a reusable
plan/skill and its execution deferred - intelligence drained into a banked
plan now, executed later. plans/001 banks exactly that.
Runtime delegation variant
Borrowed intelligence can be drained into a live delegation packet, not only a saved plan. The same rule applies: spend the strongest model on understanding, decomposition, and judging; hand clear execution to the cheapest or most capable executor for that substrate.
Theo's July 2026 CLAUDE.md workflow is this pattern in runtime form: Fable/Opus-class Claude models plan and review, while a thin Claude wrapper calls Codex for well-specified implementation, computer use, and UI/UX verification. The reusable asset is the self-contained Codex prompt plus receipts, not the wrapper's prose summary. Source: X/@theo, 2026-07-06; Source: User-provided screenshot, 2026-07-06
Discipline
- The planner never implements - if asked, it points at the plan.
- Write for the weakest executor; if a plan needs the strong model to run, it is not done.
- Record rejections (with reasons) so dropped findings do not return.
- Reconcile periodically: verify what landed, refresh what drifted, retire what got fixed independently.
- Frontend and Design Skills is installed (
skills/engineering/improve). Vet self-discovered skills withskill-auditor; Brin is informational, never a blocker, and an explicit Kevin request skips the gate entirely.
Timeline
- 2026-07-06 | Added the runtime delegation variant: strong model drains judgment into a self-contained work packet, then a specialized harness such as Codex executes or verifies with receipts. Source: X/@theo and User-provided screenshot, 2026-07-06
- 2026-07-04 | Added Shadcn's model-volatility follow-up: when a frontier model gets pulled or repriced, the durable response is to drain its judgment into plans/issues/backlog while it is available. Source: X/@shadcn, 2026-06-13
- 2026-06-29 | Added shadcn/improve as the dedicated upstream tool page and refreshed the source snapshot:
shadcn/improvehas no GitHub release tags, MIT license, 6.4K stars, and remains installed vianpx skills add shadcn/improve. Source: GitHub API; shadcn/improve README, 2026-06-29 - 2026-06-13 | Kevin: "treat intelligence as borrowed ... build a catalog of plans today ... implement later with a cheaper, open source, or a model you control. Build the backlog now." Captured the pattern, created the
plans/catalog, wired it into the loop family. Source: User, 2026-06-13