Constraints over Context

To get reliable output from an AI agent, make the desired behavior easy or mandatory by construction. Context helps the agent understand; constraints make it comply.

This is the thesis behind Sigil UI, Kevin's agent-native design-system work, and a broader operating philosophy for agent systems.

The Failure of Context Alone

Reference libraries, screenshots, examples, and long prompts are useful, but they remain advisory. An agent can understand a reference and still produce its own button, spacing, radius, and copy because the implementation surface is unconstrained.

That is the "AI-generated aesthetic": not a style, but the absence of enforced decisions. The model falls back to the median.

Constraints Make Drift Hard

In a constrained system:

  • components already exist
  • tokens define visual identity
  • rules block forbidden patterns
  • skills teach the local procedure
  • doctors verify the result
  • the agent composes instead of inventing primitives
Reference-only:   context -> agent -> new components -> hope
Constraint-first: spec -> tokens/components -> agent composes -> verify

The agent does not need to have taste. The system has taste. The agent needs to operate the system.

Sigil as the Concrete Case

Sigil UI implements this as DESIGN.md -> tokens -> components -> doctors:

  • 519 tokens across 33 categories encode identity.
  • 46 presets make identity swappable.
  • 350+ components consume variables.
  • CLI doctors catch incomplete presets, contrast failures, and component convention drift.
  • Skills tell agents how to edit the system.

That is materially different from giving the agent a screenshot and saying "make it like this." The design has become an operating environment.

Generalization

The same pattern appears across Kevin's stack:

Domain Context-only failure Constraint-first fix
UI agent sees references, still writes generic Tailwind tokens, components, taste rules, visual doctors
Code style agent reads preference, forgets under pressure lint, formatters, type checks, hooks
Git safety agent knows not to wipe work, still runs destructive commands hooks and explicit blocked commands
Wiki maintenance agent remembers to update docs sometimes doctor, freshness check, log/index requirements
Tool routing agent picks favorite tool capability map and trigger evals

Design Rule

Use context for judgment and constraints for invariants. If a failure would be expensive or likely to recur, do not merely describe the correct behavior. Encode it.

Concept Position

Field Value
Concept family Brain, memory, and retrieval
Concept owned To get reliable output from an AI agent, make the desired behavior easy or mandatory by construction. Context helps the agent understand; c...
Category map Concept System Map

Timeline

  • 2026-07-01 | Concepts category refresh added this page to the Brain, memory, and retrieval family, linked it to Concept System Map, and kept it standalone because it owns this reusable mental model: To get reliable output from an AI agent, make the desired behavior easy or mandatory by construction. Context helps the agent understand; c... Source: User request, 2026-07-01
  • 2026-06-18 | Expanded with Sigil's current 519-token implementation, cross-domain examples, and the context-vs-constraint decision rule. Source: User request, 2026-06-18; Sigil UI
  • 2026-06-13 | Referenced in Build, Don't Buy and Sigil UI external validation: reference tools supply context; Sigil supplies constraints. Source: User, 2026-06-13
  • 2026-05-31 | Page created from the Reticle/Sigil MANIFESTO.md. Extracted the "constraints over context" principle for agent-native systems. Source: reticle-ui/MANIFESTO.md, 2026-05-31