Agent Skills: Asymmetric Wins

Asymmetric Wins

An asymmetric win is a refusal that gives back more complexity than it costs in product capability.

The spine:

Preserve the product sentence.
Refuse the small promise.
Delete the code family.

The usual shape: trade a small amount of fidelity, compatibility, modes, or reproducibility to collapse a much larger implementation graph.

This is not arithmetic and not a quota. Do not remove arbitrary features. The job is to find the one small promise that owns a disproportionate code family, then decide whether refusing that exact promise leaves the product sentence intact.

Do not say "good enough" and ship drift. Name the product sentence first. Keep the workflow, safety, accessibility, and recognizable product feel. Refuse only the promise that was forcing a second system, and keep it when the evidence says the loss is load-bearing.

Compose With

• one-sentence-test detects the opportunity (the surface audit surfaces the convenience feature that forces a second product sentence). This skill owns the decision.
• refactoring counts callers, fixtures, state branches, styling branches, docs paths, and other code-family evidence before the refusal is executed.
• frontend-design owns visual direction, accessibility, brand, and whether a UI detail is load-bearing before pixel fidelity is refused.
• cohesive-clean-breaks executes the resulting breaking change, wave ordering, and old-path deletion.
• greenfield-clean-breaks and radical-options link here instead of re-deriving the refusal move.

When To Run

Run this pass when a design adds or preserves:

a fast path beside the canonical path
a provider-specific SDK wrapper beside a standard protocol
a fallback parser for an old shape
a second transport for one environment's nicer UX
a compatibility alias nobody explicitly asked for
an option that only preserves an old mental model
a partial reflection API that makes callers ask which surfaces are real
pixel-perfect reproduction of an old UI as a refactor requirement
exact snapshot, fixture, layout, animation, or responsive-state compatibility
a hand-drawn HTML diagram or component tree that reproduces structure already
  owned by a UI primitive
a rare mode whose tests, docs, branches, and state outlive its value

Domain Manifestations

The same move shows up in different clothes:

UI refactor
  Refuse exact pixel or markup reproduction only when the core workflow,
  hierarchy, accessibility, important states, and product feel survive.

Code refactor
  Refuse legacy aliases, duplicate helpers, fallback parsers, and old call
  shapes when callers can move to one canonical path.

Tests and reproducibility
  Refuse exact old snapshots, fixture quirks, or transitional behavior when the
  product contract is clearer than the old artifact.

Design artifacts
  Refuse full HTML facsimiles and exhaustive component trees when a native
  primitive, screenshot, compact sketch, or state table preserves the same
  decision.

Architecture
  Refuse keeping both mental models alive. A half-old, half-new system is
  usually the expensive promise.

Procedure

1. Name the product sentence that must remain true.
2. List candidate refusal points: fast paths, old shapes, rare modes, provider
   exceptions, compatibility aliases, fallback parsers, exact reproduction,
   partial reflection, hand-reproduced UI structure.
3. For each candidate, list the code family it forces: methods, adapters,
   unions, error variants, tests, docs branches, UI states, styling branches,
   fixtures, screenshots, migrations, local markup, custom CSS, diagram upkeep.
4. Pick the candidate with the largest code family, not the most visible name.
5. Ask who loses what if that behavior is refused.
6. If the loss is a small convenience and the deletion removes a second shape,
   refuse the behavior and write that refusal into the spec.

The rule is evidence-seeking, not dramatic: if the product sentence survives and the code family disappears, refusal is the default recommendation. Keep the feature when the user loss is load-bearing or when the "deletion" would only move complexity somewhere harder to see.

Decision Template

Use this shape in specs and design notes:

Product sentence:
  ...

Candidate refusal:
  ...

Code family it deletes:
  ...

User loss:
  ...

Decision:
  Refuse it / keep it because ...

UI Refactor Template

Use this when the promise is visual or interaction fidelity:

Product sentence:
  ...

Must preserve:
  workflow, information hierarchy, accessibility, important states,
  recognizable product feel, inspectable state, reviewable intent

Can refuse:
  exact spacing, old breakpoints, one-off hover states, incidental animation,
  pixel-perfect empty/loading/error states, duplicate responsive layouts,
  locally reproduced HTML when a shared primitive owns the same contract,
  exhaustive component trees when a smaller artifact preserves the decision

Code family it deletes:
  ...

Replacement artifact:
  shared primitive / screenshot / compact sketch / state table / full tree

Decision:
  Refuse it / keep it because ...

Pixel, markup, and screenshot fidelity are load-bearing when the exact detail carries comprehension, accessibility, trust, brand, or a regression-sensitive state. Otherwise, exact reproduction is a promise like any other: keep it only when it earns the code family it forces.

UI Artifact Ladder

Use the smallest artifact that preserves the design decision:

1. Shared primitive or existing component
   Use when the UI library already owns the structure, spacing, states, and
   accessibility contract.

2. Screenshot or generated image
   Use when visual appearance matters more than exact markup.

3. Compact sketch or state table
   Use when the decision is hierarchy, state, or flow.

4. Partial tree
   Use when parent-child structure is the point under review.

5. Full HTML diagram
   Use only when exact DOM shape is the product contract or the bug.

Do not make agents reproduce a full HTML tree to prove a design unless the tree is the thing that must stay stable. Prefer the natural primitive, then verify the states that matter.

Worked Example: Social Sign-In

Product sentence:
  All social sign-in routes through the API-hosted page via OAuth 2.1 PKCE.

Candidate refusal:
  Browser SPAs can use Google GIS for a roughly 1-second sign-in.

Code family it deletes:
  signInWithIdToken
  OIDCProvider narrowing
  per-app GIS helpers
  GIS blocked-browser UI
  SocialSignInUnavailable
  provider-specific SDK scaling for Apple and Microsoft
  two social sign-in docs branches
  two social sign-in test paths

User loss:
  Google sign-in is a few seconds slower in browser SPAs.

Decision:
  Refuse it. The UX loss is small; the second auth shape is permanent.

The product still has social sign-in. It refuses one fast path so one invariant can own every provider and environment.

For narrative context, see docs/articles/20260504T160541-asymmetric-wins-support-fewer-features-to-collapse-complexity.md.