Compose an adaptable four-role judge panel (grounding, coherence, corroboration, audit) to evaluate a candidate for promotion — a belief, claim, record entry, or artifact — against novel or arbitrary content such as a REPL state, corpus, document set, or codebase. Use whenever the user asks to judge, vet, adjudicate, or promote a claim or candidate; mentions spawning judges, judge panels, promotion candidates, belief-to-fact promotion, or a reconciliation record; or wants an impartial evaluation of content they authored themselves — the self-invested-claimant case is exactly what this skill hardens against. Pairs with the prompt-engineering and hypershot-protocol skills, which must be invoked first in any session that authors judge prompt bytes (house Guardrail 15).
Run one turn of the self-improving loop end to end — DIAGNOSE (spark-steering) → AUTHOR (prompt-engineering + hypershot-protocol) → SPAWN (subagent-composition) → ADJUDICATE (judge-composition | self-play | complexity-convocation) → user-gated PROMOTE — enforcing the house invariant at every seam and halting at the user gate. Use when a human has opened a turn with spark-steering and wants the diagnosed fix — a new or updated skill, sub-agent, prompt, or Trellis module — carried through authoring, spawning, and clean-room adjudication without hand-composing each stage; or to sequence a capability-flywheel turn. Composes the other skills; replaces none. Invoke prompt-engineering and hypershot-protocol first when authoring any prompt bytes (house Guardrail 15).
Write an OpenCnid five-tier chain-of-density note for a research paper, or scaffold a complete new paper repo around one. Use whenever the user asks to add, summarize, note, recognize, or "make a repo for" a research paper (arXiv, ACL Anthology, a lab blog post — any published research), mentions a density chain, CoD note, tier note, paper note, or the OpenCnid paper-repo collection — even if they never name this skill. Also use when updating, re-verifying, or extending existing OpenCnid paper repos, including batch runs across many paper repos at once. Also runs in SYSTEM MODE — a chain-of-density map of a whole codebase or system rather than a paper — a branching "density-trellis" reverse-engineered from the project's commit history, written to docs/density-chain/DENSITY-CHAIN.md with an interactive HTML+SVG render. Use system mode when the user asks to map, summarize at increasing density, or make a density-chain / density-trellis of a project, engine, codebase, or its features and roadmap.
Assess whether something is excessively complex and return interpretable, user-gated recommendations, by composing a fresh cover of sub-agent judges for the artifact at hand — one agent characterizes the context first, a composer builds one judge per belief-facing seat, three judges evaluate in isolated clean contexts, and a judges-judge audits their runs on real telemetry. Use when the user says something is too complex, over-engineered, convoluted, tangled, bloated, or hard to follow; asks whether complexity is justified or worth keeping; asks what to simplify or cut; wants a second opinion on a design, module, schema, config, plan, or document they suspect is overbuilt; or needs a structurally impartial review of work they authored themselves — the self-invested-claimant case this ceremony hardens against. Spawns real sub-agents. Pairs with prompt-engineering and hypershot-protocol, which must be invoked first in any session that authors this skill's prompt bytes (house Guardrail 15); builds on judge-compo
Meta-prompting technique using zero-contamination structural examples. A hypershot is a frame with free variables; the variables range across a continuum from unbounded spread operators (...) to fully-loaded instruction-bearing variables ({Concise_Action_Oriented_Reply_With_No_Fluff}). Use whenever authoring a prompt template, workflow output schema, agent instruction, wiki entry another agent will read, or any artifact that primes a model's generation, even when the user does not explicitly say "hypershot." It is the right move anytime you would otherwise reach for few-shot examples and risk semantic contamination of the context window. Pairs with the prompt-engineering skill, whose master template is a hypershot.
The Lexideck prompt-engineering protocol: structural clarity and manifold alignment over "magic words." Use whenever authoring a prompt, template, agent instruction, system prompt, or output schema and precision matters: semantic tagging, hierarchical markers, structured placeholders, collections, and attention management, plus the best-practice loop. Triggers include "write a prompt", "improve this prompt", "prompt template", "agent instructions", "system prompt", "output schema". For the contamination-free structural-example technique this protocol's master template relies on, see the hypershot-protocol skill.
Diagnose which SPARK axis (Skills, Personalities, Approaches, Resources, Knowledge) is actually short before reaching for a fix, and move only that one. Invoke by name when stuck mid-task, when output is shallow despite a clean tool run, or before adding a permission rule, MCP server, plugin, subagent, CLAUDE.md, memory file, or new skill. Also for steering, adapting, tuning, or reconfiguring an agent's SPARK coordinates.
Compose a Claude Code sub-agent on the fly — persona, inherited-context transfer, tool budget, and return contract — as either an ephemeral Agent-tool call or a persistent .claude/agents/*.md definition. Use whenever the user asks to spawn, delegate to, build, design, or fix a sub-agent or agent file; mentions agent frontmatter, tool allowlists, parallel agents, fan-out, or worktree isolation; or when a sub-agent came back with thin, wrong, or off-scope output and the prompt is the suspect. Also use when deciding whether to delegate at all. Pairs with the prompt-engineering and hypershot-protocol skills, which must be invoked first in any session that authors agent prompt bytes (house Guardrail 15).