catalog-v4.0: Merge pull request #16 from drayline/release/v4.0

5-generation Claude alignment ship for the rootnode-skills system. 27 Skills recalibrated 2026-07-25.

This release coordinates the v4.0 catalog ship. Each Skill ships its own per-package release; this is the umbrella that frames the catalog as a unified system.

What's in v4.0

27 Skills across 9 functional categories. Two surfaces: chat-side Claude Projects and Claude Code execution. Unified by a single principle layer governing both.

The v4.0 ship is the 5-generation alignment cycle triggered by Claude Opus 5 GA (2026-07-24). Calibration scope moves to Opus 5 + Sonnet 5 dual-primary (both default to high effort on Claude API and Claude Code), with Fable 5 integrated-aware, Opus 4.8 fallback-graceful (a model users are silently served on classifier-flagged requests — Skills must produce correct-shape output there), Sonnet 4.6 legacy-graceful, and Haiku 4.5 graceful-with-extended-thinking.

Load-bearing changes:

• Effort inversion. Default effort is high, not xhigh. Anthropic recommends starting at high and stepping up to xhigh for demanding coding/agentic work. low/medium are legitimate primary cost controls, not degraded modes.
• D4 verification-instruction discipline. New AEA §4.14 codifies the surface-invariant rule: remove self-directed re-checking instructions (Opus 5 already re-checks its own work; instructing it compounds into over-verification), keep external-artifact verification (evidence grounding against sources / files / tests / rendered pages) in imperative voice. Sweep patterns and per-hit classification procedure documented.
• D5 prompt-compilation contract change. Project Mode delivers scaffold artifacts as separate files on disk (Custom Instructions + each knowledge file, {code}_ prefixed) rather than inline code blocks. Inline delivery remains as a runtime-constrained fallback, flagged when used.
• D6 context-budget rebuild. Automatic-RAG-by-window replaces the 200K-era fixed-threshold model. Fixed-token tier bands retired for qualitative window-relative bands; the ~66,500 measurement preserved as historical context; empirical measurement is the primary method for borderline projects.
• Tendency taxonomy expansion. 10 → 14 numbered entries. Verbosity refactored into a three-surface family (conversational / agentic narration / written deliverable). Four Opus 5-specific tendencies added: over-verification (REMOVAL countermeasure), scope expansion, subagent over-delegation, correction narration. Two prompt/environment-conditional defects treated separately (conservative-instruction literalism, thinking-disabled output artifacts).
• D8 canonical-KF mirror completion. root_CLAUDE_OPTIMIZATION_NOTES.md and root_AUDIT_FRAMEWORK.md added to audit/canonical-kfs/; the O10 staleness diff now covers 7 KFs rather than 5.

Every Skill's calibration marker is updated to the new dual-primary scheme; the T3 label retires "Opus-recommended" for "High-effort recommended." Six previously untagged Skills received markers. See the Model Compatibility section of the README for the full effort ladder, fallback semantics, and per-tier definitions.

Download

All Skills, one download: grab rootnode-catalog-v4.0.zip from the Assets below. It bundles all 29 packaged Skill files (24 chat-side -cp flat zips for Claude Projects, 5 Claude Code -cc wrapper zips). Unzip it, then upload or install the Skills you want.

Individual Skills: use the links under Skills below — each Skill's own release carries its packaged zip.

GitHub auto-attaches "Source code (zip / tar.gz)" to every release; those are the raw repository, not the packaged Skills. Ignore them and use the bundle or the per-Skill zips.

Skills

Build (3)

• rootnode-prompt-compilation
• rootnode-skill-builder
• rootnode-cc-design

Diagnose (5)

• rootnode-project-audit
• rootnode-global-audit
• rootnode-anti-pattern-detection
• rootnode-prompt-validation
• rootnode-repo-hygiene

Optimize (3)

• rootnode-behavioral-tuning
• rootnode-memory-optimization
• rootnode-context-budget

Cross-Layer (1)

• rootnode-full-stack-audit

Bridge (2)

• rootnode-project-brief
• rootnode-session-handoff

Cross-Boundary (2)

• rootnode-handoff-trigger-check
• rootnode-profile-builder

Runtime (CC, 2)

• rootnode-critic-gate
• rootnode-mode-router

Block Libraries (4)

• rootnode-block-selection
• rootnode-identity-blocks
• rootnode-reasoning-blocks
• rootnode-output-blocks

Domain Packs (5)

• rootnode-domain-business-strategy
• rootnode-domain-software-engineering
• rootnode-domain-content-communications
• rootnode-domain-research-analysis
• rootnode-domain-agentic-context

Architecture

Two surfaces, one mental model. Chat-side Claude Projects and Claude Code execution share a unified principle layer covering placement discipline, decomposition by mechanism, behavioral countermeasures, cross-Skill composition contracts, verification-instruction discipline (D4), and landscape-volatility discipline (D8).

Site

rootnode.design