Frontier

The law this skill runs on: a perfectly consistent system can still be two years stale. Everything else in the suite consumes a standard (/card, frontend-slides) or enforces it internally (harmonize); nothing looks OUTWARD to ask whether the standard itself is current — on ANY surface. /frontier is that outward eye, generalized across surfaces: for a surface, it measures your work against the live external frontier and pulls the frontier toolchain toward you. It SETS the standard; harmonize converges the parts onto it. Frontier is a moving target, so this skill hardcodes NO tools, versions, or exemplars — method in the body, the per-surface trust-map in ~/.claude/frontier/surfaces.md (each surface: its benchmark exemplars + frontier-tool sources + the artifacts it reads + a last-verified date), the live web as the source of truth.

Three hard invariants:

• Audit is read-only; adopt/install is the dangerous half. Default audit mutates NOTHING (mirrors /raze audit, /hone audit). Adopting a frontier tool is system-mutating — the additive twin of raze's deletion — so it inherits raze's contract: propose-on-approval, per-item (no bulk yes), pinned versions, branch-only, build-green after each, reversible.
• Respect the surface's stack, absolutely. Read the surface's real config first (package.json / requirements.txt / CLAUDE.md / the relevant module). Never adopt what the stack does not want; in a constrained surface (e.g. this repo is stdlib-first) REFUSE to install and downgrade to audit-only, saying why. DELEGATE the actual adoption to the surface's executor (design → vercel:shadcn; model-stack → the claude-api skill + provider config; etc.) — never re-roll it.
• Live-first, never a baked-in snapshot, per surface. Discover each surface's frontier by querying its dated sources at run time; re-fetch the benchmark exemplars live each run, never from a remembered impression. A frozen list is the rot this skill exists to fight.

/frontier (no surface — list + triage)

Read ~/.claude/frontier/surfaces.md; print the configured surfaces, each with its last-verified date and a one-line staleness flag (newest source older than ~30 days = stale). Recommend the surface most worth auditing now. Mutates nothing. If the file is missing, create it from the seed (design fully specified + the proposed surface stubs) and say so.

/frontier audit (default mode)

Benchmark ONE surface against its 2026 frontier; write nothing.

1. Pin the surface. From surfaces.md, load the surface's definition: what it covers, its benchmark exemplars, its frontier-tool sources, and which of YOUR artifacts it reads (design → ~/.claude/DESIGN_SYSTEM.md, /card renders, the repo UI; model-stack → your model-routing module + requirements.txt + the claude-api reference; etc.). Fan the reads to a scout; only the surface signatures return to main.
2. Build the benchmark rubric, live. WebFetch each exemplar's current state and distill 3-6 concrete, observable signatures. Judge by the surface's declared axes — design = product-craft (Apple/Linear/Figma) AND emotional/motion/warmth (Pixar/poolday); other surfaces define their own axes in surfaces.md. "Current" is operationalized with checkable signals, never vibes.
3. See the real artifact. For visual surfaces, render via the /card headless-Chrome path (--headless=new, forced reduced-motion, real viewport — per the render-verification lesson) and Read the pixels; gaps live in the render, not the markup. For non-visual surfaces, read the live config/code/behavior.
4. Score + prioritize. Each finding: WHERE (exact artifact/element) -> WHAT IS BEHIND -> FALLS SHORT OF (which exemplar) -> the CONCRETE 2026 UPGRADE -> EFFORT (S/M/L) -> SEVERITY (dated / acceptable / frontier). Rank by severity x reach / effort. Tag OUTDATED (a stale take on an evolved pattern) / MISSING (a whole dimension absent — name structural blind spots as first-class rows, e.g. DESIGN_SYSTEM.md has no warmth/motion vocabulary vs Pixar/poolday) / CURRENT.
5. Emit + route. Print the ranked upgrade list + the single highest-leverage move. Hand pure internal-consistency drift to harmonize (not this lane). Offer plan for the tool-dependent items and a proposed standard-file delta (e.g. a DESIGN_SYSTEM.md rewrite) as a diff. Honest-empty: a surface already at the frontier gets "keep," never a manufactured critique.

/frontier plan

The dry-run of the dangerous half. Run audit , then resolve the current frontier toolchain for the gaps (WebSearch/WebFetch the surface's dated sources for this week's state; cross-reference the existing stack) and emit an ADOPTION PLAN — one row per item: WHAT (+ pinned version) · WHY (which gap) · HOW (exact command/change) · FOOTPRINT (dep type, bundle/transitive, lockfile delta) · BLAST RADIUS (files/configs touched) · REVERSAL (the exact undo). Adopt NOTHING. Every candidate clears the anti-slop gate (maintained Execute ONLY pre-approved items from a prior plan. Per-item approval (no bulk yes — one wrong adoption is the unrecoverable case); pin every version (never @latest into a lockfile); one adoption = one commit, branch-only; run the surface's build/test/lint (per /run) green between each — a red check halts, reverts that one item, reports. Highest-bar items (anything registering an endpoint/MCP server, or a global/system install) get an explicit extra confirm. Refuse-list: secrets/env, CI/deploy config, anything outward-facing. Where adoption changes a standard file (tokens replace raw hex, a motion axis appears), propose the edit as a diff, write only on approval, and hand enforcement to harmonize. Report adopted (version-verified) / skipped / rolled-back + the standard-file diff.

/frontier refresh [surface]

Re-query the sources in surfaces.md (one surface, or all), re-date them, and report what ENTERED and LEFT the frontier since last time (the diff is the value). Any mode auto-suggests a refresh when a surface's newest source is older than ~30 days. This is also the on-ramp to ADD a surface: append its block (what it covers · benchmark exemplars · frontier sources · the artifacts it reads · executor) to surfaces.md.

Rules

• Setter, not enforcer: /frontier sets each surface's standard from the outside world (and may propose rewriting the standard file); internal-consistency drift across artifacts is harmonize's lane — hand it off, don't fix it here.
• No hardcoded tools/versions/exemplars in this file — method only; the regenerable ~/.claude/frontier/surfaces.md holds the dated per-surface specifics, the live web is truth. A baked-in snapshot is the failure mode this skill exists to fight.
• Adoption safety = raze's additive twin: propose-on-approval, per-item, pinned, branch-only, build-green-per-item, reversible, refuse-list; respect the surface's stack absolutely (constrained/stdlib surface → audit-only). DELEGATE adoption to the surface's executor; never re-roll it. Make the COST legible (footprint) so approval is informed, not reflexive.
• Judge by the surface's declared axes (design = product-craft + emotional/motion/warmth; others define their own in surfaces.md). A missing whole dimension is a finding, not an oversight to skip.
• Honest-empty: already-frontier surfaces and an already-current stack get "keep / nothing to adopt" — a manufactured upgrade is worse than none (raze/hone discipline).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: yiyaw-lab
• Source: yiyaw-lab/agent-armor
• License: MIT
• Homepage: https://yiya.dev

Install and usage instructions live in the source repository linked above.