- Browse curated Gen 0 Skills and completed evolution runs. Deploy
- any Skill directly or fork-and-evolve it with your own
- specialization.
+ Browse curated Gen 0 Skills and completed evolution runs. Deploy any Skill directly or
+ fork-and-evolve it with your own specialization.
- Production-ready Skills you can deploy immediately or fork as the
- starting point for evolution.
+ Production-ready Skills you can deploy immediately or fork as the starting point for
+ evolution.
View
@@ -243,9 +230,7 @@ export default function AgentRegistry() {
Completed Runs
-
- Community Evolutions
-
+
Community Evolutions
{completed.length} runs
diff --git a/frontend/src/components/AppShell.tsx b/frontend/src/components/AppShell.tsx
index c067298..701fe45 100644
--- a/frontend/src/components/AppShell.tsx
+++ b/frontend/src/components/AppShell.tsx
@@ -201,10 +201,7 @@ export default function AppShell() {
GitHub
-
+
/llms.txt
diff --git a/frontend/src/components/AtomicLineageView.tsx b/frontend/src/components/AtomicLineageView.tsx
index 21f71d0..1660f96 100644
--- a/frontend/src/components/AtomicLineageView.tsx
+++ b/frontend/src/components/AtomicLineageView.tsx
@@ -17,9 +17,7 @@ interface AtomicLineageViewProps {
function parseCompositeSections(compositeMd: string): Map {
const out = new Map();
// Strip frontmatter if present.
- const fmStripped =
- compositeMd.match(/^---\s*\n[\s\S]*?\n---\s*\n([\s\S]*)$/)?.[1] ??
- compositeMd;
+ const fmStripped = compositeMd.match(/^---\s*\n[\s\S]*?\n---\s*\n([\s\S]*)$/)?.[1] ?? compositeMd;
const lines = fmStripped.split("\n");
let currentHeading: string | null = null;
let currentBody: string[] = [];
@@ -27,10 +25,7 @@ function parseCompositeSections(compositeMd: string): Map {
const m = line.match(/^##\s+(.+)$/);
if (m) {
if (currentHeading) {
- out.set(
- normalizeHeading(currentHeading),
- currentBody.join("\n").trim(),
- );
+ out.set(normalizeHeading(currentHeading), currentBody.join("\n").trim());
}
currentHeading = m[1];
currentBody = [];
@@ -93,11 +88,7 @@ function stripFrontmatter(md: string): string {
* SKILL.md inline (no fake diff). A user can see exactly what each parent
* contributed without being told lies about what "changed."
*/
-export default function AtomicLineageView({
- nodes,
- edges,
- genomes,
-}: AtomicLineageViewProps) {
+export default function AtomicLineageView({ nodes, edges, genomes }: AtomicLineageViewProps) {
const [expandedId, setExpandedId] = useState(null);
// Partition nodes into composite (generation 1, engineer_composite) and
@@ -118,9 +109,7 @@ export default function AtomicLineageView({
return { composite: undefined, parents: [] as LineageNode[] };
}
const parentIds = new Set(
- edges
- .filter((e) => e.child_id === composite!.id)
- .map((e) => e.parent_id),
+ edges.filter((e) => e.child_id === composite!.id).map((e) => e.parent_id),
);
const parentNodes = nodes.filter((n) => parentIds.has(n.id));
return { composite, parents: parentNodes };
@@ -135,9 +124,7 @@ export default function AtomicLineageView({
// side-by-side "what this parent contributed" view when a parent card is
// clicked. The composite genome is the one with meta_strategy === 'engineer_composite'.
const compositeSections = useMemo(() => {
- const compositeGenome = genomes.find(
- (g) => g.meta_strategy === "engineer_composite",
- );
+ const compositeGenome = genomes.find((g) => g.meta_strategy === "engineer_composite");
if (!compositeGenome) return new Map();
return parseCompositeSections(compositeGenome.skill_md_content);
}, [genomes]);
@@ -195,27 +182,23 @@ export default function AtomicLineageView({
- Atomic composites are assembled{" "}
- from many independent parents, not mutated from one.
- Running a line-by-line diff between a parent (e.g. the
- heex-and-verified-routes variant) and the composite would produce
- nonsense — the parent's ~30 lines all appear "removed" and the
- composite's ~300 lines all appear "added", because they're about
- completely different topics. The traditional Lineage Diff Viewer
- is reserved for molecular{" "}
- mutation chains where parent → child is a real edit.
+ Atomic composites are assembled from many
+ independent parents, not mutated from one. Running a line-by-line diff between
+ a parent (e.g. the heex-and-verified-routes variant) and the composite would produce
+ nonsense — the parent's ~30 lines all appear "removed" and the composite's ~300 lines
+ all appear "added", because they're about completely different topics. The traditional
+ Lineage Diff Viewer is reserved for{" "}
+ molecular mutation chains where parent →
+ child is a real edit.
Instead, this view shows a{" "}
- contribution view:
- click a parent card below to see its SKILL.md side-by-side with
- the matching section in the composite. The composite's body is
- organized by dimension — the heex-routes parent contributes the
- ## HEEx + Verified Routes{" "}
- section, the streams parent contributes the{" "}
- ## Streams + Collections{" "}
- section, and so on. You can read both sides and see exactly
- what survived into the final skill.
+ contribution view: click a parent card
+ below to see its SKILL.md side-by-side with the matching section in the composite. The
+ composite's body is organized by dimension — the heex-routes parent contributes the
+ ## HEEx + Verified Routes section, the streams parent
+ contributes the ## Streams + Collections section, and
+ so on. You can read both sides and see exactly what survived into the final skill.
@@ -231,9 +214,7 @@ export default function AtomicLineageView({
{compositeGenome?.frontmatter &&
typeof compositeGenome.frontmatter === "object" &&
"name" in compositeGenome.frontmatter
- ? String(
- (compositeGenome.frontmatter as { name: string }).name,
- )
+ ? String((compositeGenome.frontmatter as { name: string }).name)
: composite.id}
@@ -274,9 +255,7 @@ export default function AtomicLineageView({
}`}
>
- No matching section found in the composite body for
- this parent's dimension. The parent may have
- contributed cross-cutting rules rather than a single
+ No matching section found in the composite body for this parent's dimension.
+ The parent may have contributed cross-cutting rules rather than a single
named section.
)}
@@ -343,10 +321,9 @@ export default function AtomicLineageView({
Top: the parent variant's full SKILL.md. Bottom: the matching{" "}
- ##{" "}
- section from the composite where this parent's guidance was
- distilled. Sections may have been compressed, reworded, or
- merged during Engineer assembly.
+ ## section from the
+ composite where this parent's guidance was distilled. Sections may have been
+ compressed, reworded, or merged during Engineer assembly.
- Select a parent from the list on the left to see its
- SKILL.md alongside the matching section in the composite.
+ Select a parent from the list on the left to see its SKILL.md alongside the matching
+ section in the composite.
{runDetail.status}
@@ -255,8 +248,7 @@ export default function AtomicRunDetail({
"This run has 1 generation. A re-evolution would produce Gen 2 seeded by this Gen 1 composite + newly-spawned alternatives."
}
>
- Gen {competitionScores?.generation ?? 1} /{" "}
- {competitionScores?.total_generations ?? 1}
+ Gen {competitionScores?.generation ?? 1} / {competitionScores?.total_generations ?? 1}
{variants?.length ?? "—"} dimensions
@@ -287,9 +279,7 @@ export default function AtomicRunDetail({
@@ -587,8 +558,8 @@ function DimensionsOverview({
Per-Dimension Fitness vs Baseline
- Each dimension's evolved fitness compared to raw Sonnet on the same challenges.
- Copper = skill fitness. Gray = raw Sonnet baseline.
+ Each dimension's evolved fitness compared to raw Sonnet on the same challenges. Copper =
+ skill fitness. Gray = raw Sonnet baseline.
@@ -599,11 +570,7 @@ function DimensionsOverview({
Foundation
{foundations.map((d) => (
-
+
))}
)}
@@ -615,11 +582,7 @@ function DimensionsOverview({
Capabilities
{capabilities.map((d) => (
-
+
))}
)}
@@ -633,7 +596,8 @@ function DimensionsOverview({
Benchmark Tier Breakdown
- Raw Sonnet performance by challenge difficulty tier on {benchDetail.total_challenges} challenges.
+ Raw Sonnet performance by challenge difficulty tier on {benchDetail.total_challenges}{" "}
+ challenges.
@@ -648,16 +612,18 @@ function DimensionsOverview({
{benchDetail.tiers.map((t) => (
-
- Empirical knowledge about building skills for AI coding agents.
- Every finding is backed by measured data from 867 controlled
- experiments, not theory or intuition.
+ Empirical knowledge about building skills for AI coding agents. Every finding is backed
+ by measured data from 867 controlled experiments, not theory or intuition.
- This run has no competition bracket data available. The seed run
- either pre-dates the{" "}
-
- [competition_scores]
- {" "}
+ This run has no competition bracket data available. The seed run either pre-dates the{" "}
+ [competition_scores]{" "}
learning-log format, or its payload was malformed during export.
@@ -83,38 +75,37 @@ export default function CompetitionBracket({
- Each dimension ran a 1-round{" "}
- competition: the pre-existing seed variant vs. one freshly-spawned
- alternative. Both variants were scored on{" "}
-
- {challengesPerVariant} sampled challenges
- {" "}
- drawn from the dimension's 108-challenge training pool (27 held-out
- for future test-time evaluation). The variant with the higher{" "}
+ Each dimension ran a 1-round competition:
+ the pre-existing seed variant vs. one freshly-spawned alternative. Both variants were
+ scored on{" "}
+ {challengesPerVariant} sampled challenges{" "}
+ drawn from the dimension's 108-challenge training pool (27 held-out for future test-time
+ evaluation). The variant with the higher{" "}
mean L1 fitness won.
- Scoring criteria per
- challenge: the L1 scorer runs regex checks against the generated
- output file. Each challenge defines{" "}
-
- must_contain
- {" "}
- required substrings,{" "}
-
- must_not_contain
- {" "}
- forbidden substrings, and a set of dimension-specific anti-pattern
- detectors. The fraction of passing checks is weighted by each
- check's weight, and the weighted pass rate becomes the score.
+ Scoring criteria per challenge: the L1
+ scorer runs regex checks against the generated output file. Each challenge defines{" "}
+ must_contain required
+ substrings,{" "}
+ must_not_contain{" "}
+ forbidden substrings, and a set of dimension-specific anti-pattern detectors. The
+ fraction of passing checks is weighted by each check's weight, and the weighted pass
+ rate becomes the score.
- Limitations for this run:{" "}
- this is generation{" "}
+ Limitations for this run: this is
+ generation{" "}
{generation}/{totalGenerations}
{" "}
- (single-round). {baselineRan ? "A baseline run on the full challenge pool was completed before this competition." : "No baseline pass over the full 108-challenge pool was run before sampling."} Higher reviewer layers (L2 trigger accuracy, L3 trace, L4 comparative, L5 trait attribution) were not exercised. Real production runs would layer these on top for richer signal.
+ (single-round).{" "}
+ {baselineRan
+ ? "A baseline run on the full challenge pool was completed before this competition."
+ : "No baseline pass over the full 108-challenge pool was run before sampling."}{" "}
+ Higher reviewer layers (L2 trigger accuracy, L3 trace, L4 comparative, L5 trait
+ attribution) were not exercised. Real production runs would layer these on top for
+ richer signal.
+ Generating skill variants — each variant is a complete SKILL.md package with
+ scripts, references, and examples. This typically takes 1-2 minutes per
+ dimension.
- Generating skill variants — each variant is a complete SKILL.md package with scripts,
- references, and examples. This typically takes 1-2 minutes per dimension.
-
- Deploy autonomous populations into adversarial environments.
- Watch them compete, mutate, and survive to forge the ultimate
- cognitive skillsets.
+ Deploy autonomous populations into adversarial environments. Watch them compete,
+ mutate, and survive to forge the ultimate cognitive skillsets.
@@ -117,28 +113,16 @@ export default function EvolutionDashboard() {
Your Runs
@@ -180,11 +164,9 @@ export default function EvolutionDashboard() {
Backed by Research
- SKLD didn't emerge from a vacuum. The techniques below come from
- active ML and applied-AI research; SKLD composes them for a
- specific artifact (Claude Agent Skills) with a controlled
- benchmark. Each source is mapped to what we took and what we
- didn't.
+ SKLD didn't emerge from a vacuum. The techniques below come from active ML and
+ applied-AI research; SKLD composes them for a specific artifact (Claude Agent Skills)
+ with a controlled benchmark. Each source is mapped to what we took and what we didn't.
Guo et al. · arXiv:2309.08532 ↗
@@ -232,9 +214,8 @@ export default function EvolutionDashboard() {
Reflective mutation + Pareto selection
- Trace-informed diagnosis before mutation, multi-objective
- Pareto fronts. The two techniques that make evolution
- intelligent instead of a random walk.
+ Trace-informed diagnosis before mutation, multi-objective Pareto fronts. The two
+ techniques that make evolution intelligent instead of a random walk.
Agrawal et al. · github.com/gepa-ai ↗
@@ -255,9 +236,8 @@ export default function EvolutionDashboard() {
Joint multi-component optimization
- Mutating one agent component (a prompt, a tool, a parameter)
- often requires matching changes elsewhere. SKLD applies this
- to the interdependent parts of a SKILL.md.
+ Mutating one agent component (a prompt, a tool, a parameter) often requires matching
+ changes elsewhere. SKLD applies this to the interdependent parts of a SKILL.md.
- Accumulated lessons inject into every mutation prompt, so the
- population never re-discovers failures it already explored.
- SKLD promotes these to the public Bible.
+ Accumulated lessons inject into every mutation prompt, so the population never
+ re-discovers failures it already explored. SKLD promotes these to the public Bible.
Imbue Research · imbue.com ↗
@@ -301,9 +280,9 @@ export default function EvolutionDashboard() {
Trigger accuracy + A/B comparator
- Skills have two reliability problems: activation (does it
- trigger on the right queries?) and execution (does it produce
- good output?). SKLD's L2 and L4 layers borrow this directly.
+ Skills have two reliability problems: activation (does it trigger on the right
+ queries?) and execution (does it produce good output?). SKLD's L2 and L4 layers borrow
+ this directly.
- You can't assert output == expected for LLM
- behavior. Execution traces answer what Claude did,
- which is what SKLD's L3 layer scores against.
+ You can't assert output == expected for LLM behavior. Execution traces
+ answer what Claude did, which is what SKLD's L3 layer scores against.
diff --git a/frontend/src/components/FitnessExplainer.tsx b/frontend/src/components/FitnessExplainer.tsx
index e5bc944..87e8ee7 100644
--- a/frontend/src/components/FitnessExplainer.tsx
+++ b/frontend/src/components/FitnessExplainer.tsx
@@ -11,24 +11,22 @@ export default function FitnessExplainer() {
What does "fitness" mean?
- Fitness is the average of the L1
- deterministic checks across all sampled challenges for a
- variant. Each challenge defines a list of required substrings
- (must_contain) and forbidden ones
- (must_not_contain); the score is the fraction of checks
- that pass, weighted by the challenge's own weight.
+ Fitness is the average of the{" "}
+ L1 deterministic checks across all sampled
+ challenges for a variant. Each challenge defines a list of required substrings (
+ must_contain) and forbidden ones (must_not_contain); the score is
+ the fraction of checks that pass, weighted by the challenge's own weight.
- 1.0 = every required
- substring was present and zero forbidden substrings appeared.{" "}
- 0.5 = half the checks
- passed. The composite skill's fitness is the average across all 12
- dimension winners.
+ 1.0 = every required substring was present and
+ zero forbidden substrings appeared. 0.5 = half
+ the checks passed. The composite skill's fitness is the average across all 12 dimension
+ winners.
- Higher reviewer layers — trigger accuracy (L2), trace analysis (L3),
- comparative pairwise (L4), trait attribution (L5) — are part of the
- production pipeline but were not run for this showcase.
+ Higher reviewer layers — trigger accuracy (L2), trace analysis (L3), comparative pairwise
+ (L4), trait attribution (L5) — are part of the production pipeline but were not run for this
+ showcase.
- Real evolution runs consume API budget, so SKLD.run is currently
- invite-only. Enter your code below, or request an invite and we'll
- email you one.
+ Real evolution runs consume API budget, so SKLD.run is currently invite-only. Enter your
+ code below, or request an invite and we'll email you one.
- Submitting does not automatically grant access. We'll review
- and email a code if we can onboard you.
+ Submitting does not automatically grant access. We'll review and email a code if we
+ can onboard you.
- {error && (
-
- {error}
-
- )}
+ {error &&
{error}
}
{/* Mobile entry selector — visible below lg */}
- What this is. A
- production-ready Claude Agent Skill for writing{" "}
- {family}. It enforces a
- component-forward architecture — parent LiveView as a thin
- coordinator, LiveComponents own stateful regions — covers the{" "}
- {numDims - 1} most important capability areas (HEEx templates,
- function components, streams, forms, mount lifecycle, event
- handlers, PubSub, navigation, auth) and ships with a 14-rule
- anti-patterns catalog that includes mechanical grep-based
- detectors a reviewer can run directly.
+ What this is. A production-ready Claude Agent
+ Skill for writing {family}. It enforces a
+ component-forward architecture — parent LiveView as a thin coordinator, LiveComponents own
+ stateful regions — covers the {numDims - 1} most important capability areas (HEEx
+ templates, function components, streams, forms, mount lifecycle, event handlers, PubSub,
+ navigation, auth) and ships with a 14-rule anti-patterns catalog that includes mechanical
+ grep-based detectors a reviewer can run directly.
- How it was built.{" "}
- {numDims} dimensions of the skill were evolved in parallel. For
- each, a pre-existing seed variant competed against a freshly-spawned
- alternative on 2 sampled challenges. The higher-scoring variant
- became the dimension's winner. An Engineer then assembled the{" "}
- {numDims} winners into one coherent composite, resolving{" "}
- 3 conflicts between capability stances (mount vs.
- async-load discipline, event funnel vs. PubSub direct handling,
- composite granularity). The final composite weighs ~20KB of
- Markdown across {numDims} named sections plus the detector
- catalog.
+ How it was built. {numDims} dimensions of the
+ skill were evolved in parallel. For each, a pre-existing seed variant competed against a
+ freshly-spawned alternative on 2 sampled challenges. The higher-scoring variant became the
+ dimension's winner. An Engineer then assembled the {numDims} winners into one coherent
+ composite, resolving 3 conflicts between capability stances (mount vs.
+ async-load discipline, event funnel vs. PubSub direct handling, composite granularity).
+ The final composite weighs ~20KB of Markdown across {numDims} named sections plus the
+ detector catalog.
Quality signal.{" "}
- {fitness.toFixed(3)}{" "}
- average fitness across{" "}
- {numChallenges * 2} test runs
- (L1 deterministic checks — required-substring presence +
- forbidden-substring absence).{" "}
- {perfectFitnessCount} of{" "}
- {numDims - 1} capability dimensions hit a perfect 1.000. The{" "}
- {seedWinnerCount} dimensions
- where the pre-existing seed beat the spawned alternative
- (heex-and-verified-routes, streams-and-collections,
- navigation-patterns) flag weak scorer discrimination on those
- facets — the production pipeline would layer L2 trigger accuracy,
- L3 trace analysis, and L4 comparative pairwise review on top to
- get richer signal. Bottom line:{" "}
-
- publishable as a showcase
-
- ; {spawnWinnerCount}/{numDims} wins for the Spawner are a healthy
- signal that atomic evolution isn't just elitism.
+ {fitness.toFixed(3)} average fitness across{" "}
+ {numChallenges * 2} test runs (L1 deterministic checks
+ — required-substring presence + forbidden-substring absence).{" "}
+ {perfectFitnessCount} of {numDims - 1} capability
+ dimensions hit a perfect 1.000. The {seedWinnerCount}{" "}
+ dimensions where the pre-existing seed beat the spawned alternative
+ (heex-and-verified-routes, streams-and-collections, navigation-patterns) flag weak scorer
+ discrimination on those facets — the production pipeline would layer L2 trigger accuracy,
+ L3 trace analysis, and L4 comparative pairwise review on top to get richer signal. Bottom
+ line: publishable as a showcase;{" "}
+ {spawnWinnerCount}/{numDims} wins for the Spawner are a healthy signal that atomic
+ evolution isn't just elitism.
- {installable.length} installable{" "}
- {installable.length === 1 ? "file" : "files"}
+ {installable.length} installable {installable.length === 1 ? "file" : "files"}
{" "}
- ship in the downloadable zip and are loaded by Claude at
- runtime. {metadata.length} additional metadata files are
- preserved for auditing the evolution process but are NOT part
- of the deployable package.
+ ship in the downloadable zip and are loaded by Claude at runtime. {metadata.length}{" "}
+ additional metadata files are preserved for auditing the evolution process but are NOT
+ part of the deployable package.
{item.kind} ·{" "}
- {item.present
- ? "present in this package"
- : "not generated by this run"}
+ {item.present ? "present in this package" : "not generated by this run"}
This run decomposed the skill into{" "}
{winners} dimensions, spawned{" "}
- {startingVariants} candidate variants,
- competed them on{" "}
- {challenges} challenges, and
- assembled the winners into{" "}
- one composite skill at{" "}
-
- {fitness.toFixed(2)} average fitness
-
- .
+ {startingVariants} candidate variants, competed
+ them on {challenges} challenges, and assembled
+ the winners into one composite skill at{" "}
+ {fitness.toFixed(2)} average fitness.
- Click a family name to see per-challenge detail, tier breakdowns, and
- score distributions.
+ Click a family name to see per-challenge detail, tier breakdowns, and score distributions.
@@ -200,7 +191,8 @@ export default function SkldBenchFamily() {
Score Distribution
- How many challenges scored in each composite range — most challenges cluster around 0.5–0.8
+ How many challenges scored in each composite range — most challenges cluster around
+ 0.5–0.8