ravenclaude-core

Substrate-neutral pre-execution classifier contract. A single Haiku call emits a `run_config` JSON envelope that right-sizes cardinality knobs + per-phase model tier + reasoning level for multi-phase agentic workflows (the rc-deep-research loop is the first consumer). Workflows read the envelope; substrate adapters (Claude / Codex / Copilot) map tier labels to SKUs. Carved out behind `.ravenclaude/run-config.json` `enabled: false` so adoption is opt-in and rollback is one line.

Universal pre-dispatch model right-sizer. A single Haiku 4.5 forced-tool call evaluates `{subagent_type, description, prompt_head}` at every Agent dispatch / Workflow `agent()` call / `thing-decide.py` seat dispatch and binds (downgrades AND upgrades) the model selection. Carved out behind `.ravenclaude/dispatch-config.json` `enabled: false` so adoption is opt-in and rollback is one line. Tribunal-seat verdicts are shadow forever for MVP (protects the v0.32.0 backbone-diversity invariant).

Score and improve an agent file against a 6-dimension rubric — Mission clarity, Scope sharpness, Capability Grounding alignment, Output-Contract completeness, Escalation paths, Example scenarios. Each dimension scored 1-5 with anchors; includes a remediation template that turns a low score into an actionable PR. Reach for this skill when authoring a new agent, reviewing a PR that adds or modifies an agent, or running a periodic agent-bank audit. Used by `prompt-engineer` (primary) plus `architect`.

Recompute the M/H/G/O/E/I/T/V weighted closeness score (and the observed-vs-inferred quality bar) for a product-analog comparison, reusing the exact formula from the 2026-08-14 analog-repos-gap-fill survey instead of hand-deriving the arithmetic each time. Use when scoring how close a candidate repo/product is to RavenClaude (or any similarly-shaped catalog+governance+ops comparison) on the eight-dimension rubric.

For any consumer project where a CI workflow is meant to *enforce* a property (lint, format, security scan, manifest validation, version pin, layout allow-list, etc.). When working on a CI file or adding a new check, run this skill — every gate must fail on a known-bad input AND pass on a known-good input. A gate that runs without gating, or gates the wrong thing, is invisible from inside a green CI dashboard. Also triggers on user phrases like "audit the CI", "verify the gates", "is this gate real", "does this lint actually fail".

Authors, validates and packages a claude.ai Organization Skill — the intake questions, the trigger and scope exercises, then lint, pack and verify against measured platform constraints. Reach for this when someone wants to publish a Skill to their organization, or when an upload was rejected and the reason is unclear.

Safely retire a local branch by tagging its tip first, then deleting the branch. The sanctioned escape hatch from guard-destructive.sh's blanket `git branch -D` block — turns silent-loss-of-unmerged-work into a reversible, audit-logged operation. Use when an agent or maintainer wants to clean up old feature branches that represent abandoned approaches, stale plans, or superseded work.

Point at a website home page and extract every logo variant plus the brand 'schema' (design tokens — colors, typography, radii) into a ready-to-apply brand kit, so HTML documents/reports you generate for that project match the brand. Use when the user says things like 'pull the logo and brand from this site', 'extract the brand schema from <url>', 'make my reports match <company>'s style', or 'grab the colors and fonts from this homepage'.

Route ONE well-defined job (single-file, tests, summary, mechanical refactor) off Claude via cheap-lane-delegate.sh (bounded, returns). Opt-in cheap_lane: advise|agent. NOT for quota-escape, host-switch, fresh window, or 'pass remaining work to Grok' → session-handoff.

One-off Claude orchestration — route this task through Claude's brain even when the host CLI is Copilot/GPT/Grok. The `orchestrator:` knob in comfort-posture.yaml sets the always-on default; this skill is for on-demand invocation.

Remove finished agent worktrees, prune their branches, and surface anything still in flight. Run at the end of a multi-agent session and weekly as hygiene.

For Claude sessions running in a consumer project (any project where ravenclaude-core is installed). Triggered when you discover a cross-domain finding worth contributing back to the RavenClaude marketplace — either spontaneously, or when the user says "contribute this," "save to RavenClaude," "this is worth keeping," or similar. Walks through the qualifying check, picks the right shape (lesson, best-practice, or both), and formats a copyable staging submission the user can drop into RavenClaude/docs/staging/incoming/.

Open a pull request for the current branch using the project's standard template. Verifies the branch is green, summarizes the diff against main, and pushes only after the user confirms.

For any script that produces a file the repo COMMITS and CI DIFFS — generated HTML, lockfiles, JSON snapshots, fixture files, documentation rendered from source. The output must be byte-identical regardless of who regenerated it or on what OS. This skill catches the specific failure modes that turn a freshness gate into a paper tiger that fails forever — OS-dependent path separators, locale-dependent encodings, nondeterministic ordering, drifting timestamps. Triggers on phrases like "freshness check failing", "regenerated locally but diff doesn't match", "works in CI not on Windows", "generator produces different output", and on any review of a script whose output is committed.

Route a yes/no decision through the command-review tribunal (the Thing) for a binding verdict instead of pausing the human. Use before asking the user any yes/no question, and in the post-PR decision review. The tribunal auto-decides rule/fact-derivable calls and defers genuine preferences (and irreversible/high-blast calls) back to the human.

Author a Vega-Lite / Deneb / SVG spec for a stated intent on any surface (web vega-embed, react-vega, Evidence, Observable, Power BI Deneb, Tableau extension/SVG, SVG-in-DAX). Six-step method: pick grammar → bind data → encode → wire interactivity → test null/empty → verify via render loop. Ships a surface-agnostic spec-patterns library. Mandatory security audit (no data.url, no remote loader, no SVG script) enforced by lint.py (Gate 101). Complements the visual-feedback-loop (render referee) and pbir-layout-engine (coordinate linter). NOT for coordinate/layout arithmetic (pbir-layout-engine) or render-loop orchestration (visual-feedback-loop).

When a tracked host tool (Claude Code, Copilot CLI/Chat, Codex CLI, Cursor, Gemini CLI, Aider, Windsurf — read live from host-support.json) ships a new version, scan the repo for citation-marked drift, auto-apply safe mechanical fixes, and queue judgment calls for review. Reach for this skill on "a host tool updated, what needs to change" or via the SessionStart nudge.

Capture a reference website's full design schema (spacing/type/grid/elevation/component recipes) and apply it to a target — cloning the craft while swapping in the target's own brand and structurally blocking the reference's identity from leaking. Fidelity is browser-gated; trade dress is not cleared here.

Bind THIS repo to one of your claude.ai/design design-system projects in one step — lists your projects, lets you pick, and writes .ravenclaude/design-project.json so every session's capability banner surfaces the link and the agent can read/edit the project. Use when starting design work in a repo that has no .ravenclaude/design-project.json yet (or to re-point it). Pairs with the built-in /design-sync skill (which does the actual component sync) and the DesignSync tool.

Enforces a per-PR file-count and LoC budget. PRs touching >5 files or >400 LoC route to architectural review before merge. Counters T-1 (premature abstraction) + T-4 (going off-script) — Copilot CLI degrades sharply past 10 files per file, and uncontrolled diff growth is the strongest predictor of off-pattern mutations.

Use this skill when the user wants to create a new agent and has a clear business goal but isn't fluent in the agent's target domain. Walks them through producing a strong brief by filling in `templates/agent-brief.md` from their plain-language description, then iterating once or twice before building or dispatching the agent. Triggers when the user asks for "a new agent that does X" without already supplying the technical spec, or when they've filled the brief template and left blanks where they didn't know.

Auto-discover the consumer's environment posture by probing installed CLIs (pac / az / aws / gcloud / gh) with read-only commands at session start, decoding JWTs for role/scope claims, and assembling a draft `.ravenclaude/environment-context.md` for save/edit/skip. Streamlines proposal 2026-05-22-001's permission-awareness mechanism.

Onboards a non-Claude-Code coding agent (GitHub Copilot CLI / OpenAI Codex CLI / Cursor / Aider / Devin Desktop) to this repo at session start. Use when this repo will be operated on by an agent that does not natively read CLAUDE.md. Routes through AGENTS.md, then layers the per-host wiring, version floors, and failure-mode mitigations that host actually needs.

The FORGE gated-planning pipeline that /forge runs: depth-scaled gates that turn any idea into a two-panel-reviewed, critic-checked, tiebroken, routed plan. Mimics Claude Code Ultraplan's deep-plan loop and improves it with cross-model divergence, a fact-verification gate, a correlated-error critic, and a binding tribunal for conflicts.

Score a consumer repo's GitHub-development protocol against the shipped gold-standard catalog and produce a remediation queue ranked by leverage. A 10-row core rubric (mapping 1:1 to the P3 exemplar-repo catalog + the P4 Actions-hardening rules) plus 3 agent-operability rows for the agent-as-primary-GitHub-operator bar — each row is a check the agent runs against .github/workflows/*, the repo's .git, and its ruleset, a pass/partial/fail verdict, and the shipped fix (an /init-agent-ready template or a knowledge file). Proportional banding (dynamic denominator). Use to measure a repo against the best-of-the-best bar before calling its CI/branch-protection done. Honest scope: it measures STRUCTURAL COVERAGE, not taste.

Run a periodic staleness sweep over all `plugins/<plugin>/knowledge/*.md` files and any decision-tree sections — flag entries past their `last-verified` window, categorize by re-verification effort (Tier 1-5 per Researcher schema), produce a remediation queue with named re-verifiers. Reach for this skill on the Researcher's weekly cadence OR before a marketplace release. Used by `deep-researcher` (primary) plus the maintainer.

Surface knowledge-file staleness across all plugins as a Structured Output block. Wraps the `knowledge-health.py` script — sweeps `plugins/*/knowledge/**.md`, groups files by bucket (stale / due_soon / untracked / fresh), and returns a remediation queue. Used by the release checklist and the `ravenclaude doctor` health check. Read this skill before answering "is our knowledge layer current?", "which knowledge files are stale?", or before tagging a marketplace release.

Claude-Code-session-state surfacing contract for the Mímir dashboard tab. Documents the five reachable on-disk sources under `~/.claude/` and `<project>/.claude/`, the encoded-path algorithm + reverse-decode fallback, the hard scrubbing/torn-write/staleness/worktree/server-parity invariants, the per-card source map, and the honest empty-state contract for in-process-only fields. Read this before authoring or modifying `_read_mimir` (in both `serve-dashboards.py` copies), the `/__mimir` endpoint, the `#/mimir` generator tab, or the Gate 49 render fixture.

Create an isolated git worktree under .claude/worktrees/ for a sub-agent to work in. Use this before dispatching any coder agent so that parallel work cannot collide.

Deterministic layout-arithmetic linter for dashboard/report page definitions. Runs seven checks — no-overlap (AABB), within-canvas, equal-gap, column-alignment, and three PBIR-specific invariants (no-empty-binding, bounded theme-override count, visualType/displayOption schema validity) — against a page JSON, a page directory, or a fixture set. Stdlib-only, no network, exit-coded for CI. The load-bearing technical core under the data-viz-designer agent; usable standalone to lint a Power BI PBIR page layout.

Diagnose and improve a Claude Code project's permission setup — the `.claude/settings.json` allow/ask/deny rules and the `.claude/settings.local.json` personal overrides. Use when settings.local.json has grown bloated (>20 rules, many one-shot), when permission prompts feel too frequent or too coarse, or before sharing a project's settings.json with a team. Companion to the Claude Code core `/fewer-permission-prompts` skill (which is the data-driven sweep); this skill is the design discipline that decides what should and shouldn't be allowlisted.

Pre-release checklist for shipping any plugin update through this marketplace — plugin.json + marketplace.json + architecture.md version-mirror discipline, .repo-layout.json glob coverage for new dirs, CLAUDE.md skill / template tables synced, JSON validation, prettier check, audit-gates meta-test, and the consumer migration-note rule. Includes the exact bash commands to run (Windows PowerShell + bash). Reach for this skill at the end of any plugin PR or before merging a release-candidate branch. Used by the maintainer (primary) plus `project-manager`.

Run a ready-made CONTROL probe alongside any negative result — http/dns/file/cmd. A negative is not a diagnosis until a positive control on the same subsystem has been observed. Makes the right action one line instead of an idea you have to have.

Rewrites an ambiguous-but-single-domain (domain_count<=1) user prompt via emit_optimized_prompt, or drafts an advisory multi-domain dispatch plan via emit_dispatch_plan (domain_count>=2), before the turn runs — surfacing constraints, missing context, and wild assumptions instead of silently guessing or fragmenting. Phase 5 screens/formats either generator's output before injection. Companion generators to prompt-optimizer-gate.sh's Tier-1 classifier.

Curated, applied prompt-pattern catalog used across this marketplace — decision-tree traversal pre-action prior, alternate-methods-before-blocked, Structured Output Protocol `---RESULT_START---` block, scenario-retrieval inline prior, escalation-by-mandatory-phrasing, citation-aware research, environment-context preamble, orchestrator-worker reinforcement, agent-scenario-authoring frontmatter, claim-grounding/source-honesty marker. Each pattern includes when to use it, what it composes with, an example block, and the failure mode it prevents. Reach for this skill when authoring a new agent, when an existing agent shows a behavior gap that a known pattern would close, or when the `prompt-engineer` is consulting on an agent revision. Used by `prompt-engineer` (primary).

Reversibly pseudonymize names/entities in text BEFORE sending it to a model, and restore them in the reply. Deterministic denylist (reliable) + structured PII + optional NER; local vault; residual scan. Pseudonymizes, does NOT anonymize.

Deep research harness — fan-out web searches, fetch sources, adversarially verify claims, synthesize a cited report. Includes an inline substrate adapter that reads .ravenclaude/run-config.json once at startup; when enabled:false (the default) all agent() calls are byte-identical to the pre-port baseline (a behavioral invariant — no gate currently CI-enforces it).

Iterate an artifact until it measurably passes a BOUNDED rubric — the Convergence Engine. Objective/deterministic gates run BEFORE any model judge, the judge is a DIFFERENT model than the author (never self-grade), the stop decision is a model-free predicate, and it emits the BEST iteration (keep-best/regression-revert) under a hard iteration cap + model-call budget. The engine NEVER claims perfect — verdicts are rubric-pass | capped | plateaued | budget-exhausted with an honest residual-gaps list. Use to drive any artifact (code, prose, a visual report, an agent file) to a defensible bar; for the agent-file case it delegates to agent-quality-rubric, and for visual surfaces it composes with visual-feedback-loop.

Build a website or dashboard by prompting Claude and watching it render live, then commit it back — by COMPOSING existing surfaces (Claude Code on the web + a self-contained-HTML constraint + a GitHub Pages/githack branch preview + a PR), not by building a custom backend. Use when someone wants the "pick a repo, prompt, see it build, ship it" loop. Covers the static/self-contained case fully; says when a real data-connected dashboard needs the heavier Tier-1 path.

Whole-repo systematic bug sweep — deterministic chunking + a content-hash cache + a Workflow-orchestrated multi-dimension review fan-out (correctness / security / concurrency / resource-leaks / error-handling / performance / ci-cd-actions-security / dead-code-simplification), cross-model-checked at higher tiers, merged and verified before any fix is proposed. Distinct from Claude Code's own built-in /code-review, which reviews a diff/PR/branch — this skill reviews the whole tree, including GitHub Actions workflows and CI gate definitions.

Meta-skill that keeps all agents, skills, and knowledge files current and honest — Daily Quick Check + Weekly Deep Research modes. Categorization schema (Tier 1 Consensus / 2 Strong-but-Contextual / 3 Divergent / 4 Emerging / 5 Deprecated). Includes 90-day staleness checks for decision trees AND `.ravenclaude/environment-context.md`.

For the RavenClaude marketplace maintainer. Walks every file in docs/staging/incoming/ one at a time, runs a security sweep + topic-expert analysis on each, presents each with a keep/update/deny prompt, promotes approved submissions to their canonical location (docs/memory-bank/lessons-learned.md for lessons, docs/best-practices/<slug>.md for best-practices), and deletes denied ones. Trigger when the maintainer says any of "check for updates", "check the staging queue", "anything in staging", "review submissions", "drain the queue", or "review staged contributions" — also fires on the slash command /review-staged-contributions.

Gate an unattended scheduled routine's produced diff/PR through a two-panel, cross-model tribunal (mirrors /forge-pipeline's divergent-panel + tiebreak shape, applied after the fact) before it is finalized: approved as-is, sent back for a bounded revision round, or escalated to a human. Use at the end of any routine whose run contract ends in opening or updating a PR (plugin-discovery, research-cadence).

Run the project's full quality gate — format check, lint, typecheck, unit tests, integration tests — in order, fail fast, summarize. Use before reporting any non-trivial change as complete.

Consult the unverified scenarios bank (`plugins/<plugin>/scenarios/*.md`) before answering plugin-domain questions. Glob + tag-filter + recency-weight, surface top 2-3 with mandatory unverified-scenario preamble ("Based on N unverified scenarios from YYYY-MM tagged [scope] — verify in your environment"). Secondary source — never replaces canonical knowledge files.

Discover great-but-low-visibility ideas — the fringe excellence mainstream search buries under popularity. Given a seed (a name, repo, post, or topic), traverse its citation/collaborator graph + the periphery sources, score each candidate on depth × novelty × practitioner-grounding MINUS a popularity penalty, dedup against the idea-board + installed plugins, and emit a ranked shortlist that routes the best finds into /forge. Use when you spot a promising fringe idea (often on LinkedIn / X / GitHub) and want the cluster around it, or to periodically sweep a niche for under-the-radar excellence.

Write a run-dir handoff and start a NEW interactive session (unbounded TUI — never a headless print-flag). Use for /handoff, context full, fresh window, quota host-switch, 'pass remaining work to Grok'. NOT for one bounded Grok/Copilot job → cheap-lane-delegation.

Hand a mid-flight finding or a small, well-scoped task to the RIGHT peer Claude Code session — the one already working in the relevant worktree/branch — via ListAgents/SendMessage, instead of paging the human or letting the finding go stale until that session's next Stop. Use when you discover something relevant to another live worktree; NOT for routine sub-agent dispatch (spawn-team) or starting a new peer session.

Translate `.ravenclaude/comfort-posture.yaml` into `.claude/settings.json` permission rules. Owns the (category, level) → permission-rule emission table that the `/set-posture` slash command applies. Read this skill when authoring or reviewing the translation logic, when a category is missing rules, or when a level's behavior surprises you.

Team Lead dispatch playbook. Pick the surface first (slash command vs skill vs specialist agent vs orchestration shape — Step 1.25), then whether to delegate (Step 1.5), then which agents and order. Load whenever choosing skill vs agent vs slash, weighing delegation, or about to dispatch more than one agent. Keeps routing consistent; platform description-match alone is not the router.