- marketing-psychology eval 4: BJ Fogg assertion did not match expected_output
which lists Goal-Gradient Effect. Fixed.
- sales-enablement eval 2: all 6 categories assertion contradicted expected_output
which only categorizes the 3 given objections. Fixed.
- ad-creative eval 5: TikTok hard limit corrected to recommended (80 chars
recommended, 100 max) per SKILL.md.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fixes 5 issues identified by independent Codex review:
- product-marketing-context: match auto-draft workflow, section flexibility
- marketing-psychology: replace phantom models with actual SKILL.md models
- ad-creative: correct RSA pinning guidance to match skill
- free-tool-strategy: boundary test now defers to related skill (page-cro)
- paywall-upgrade-cro: boundary test references only related skills
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Each skill now has 5-8 evals covering:
- Core framework usage with realistic prompts
- Casual trigger phrase variants
- Sub-type and section-specific coverage
- Boundary tests (skill deferral to related skills)
- Structured assertions for grading
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Following Anthropic skill-creator guidance that Claude undertriggers
skills, make descriptions pushier across all 32 skills:
- Add casual/frustrated user phrases
- Add implicit need triggers where users need the skill but dont name it
- Add catch-all sentences explaining when to use
- Add missing cross-references between related skills
- Ensure consistent format across all descriptions
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The .agents/skills/ directory is the emerging cross-agent standard adopted by
OpenAI Codex, Cursor, Windsurf, and npx skills. All 32 skills now check
.agents/product-marketing-context.md first with .claude/ fallback for older
setups. Install paths in README updated accordingly. Bumps all skills to 1.1.0.
Closes#50
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Free, local-first voice synthesis studio (MIT license) powered by Qwen3-TTS.
Voice cloning from short samples, local REST API, no per-character costs.
4-5x faster on Apple Silicon via MLX.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix 3 wrong character counts in ad copy examples (all off by +1)
- Fix TikTok ad text: 80 recommended / 100 max (was 100 recommended)
- Fix LinkedIn description: add 300 max (was missing)
- Fix Gemini API model name: gemini-2.5-flash-image (was gemini-2.0-flash-exp)
- Fix Cartesia pricing: $0.03/min (was $0.0085/sec)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Covers ElevenLabs (voice cloning, best quality), OpenAI TTS (cheap at scale),
Cartesia Sonic (40ms latency), PlayHT, Resemble AI, WellSaid Labs, Fish Audio,
and cloud providers. Includes comparison table, decision tree, and voice+video
layering workflow with ffmpeg.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ByteDance's video model with native audio, up to 2K resolution, multimodal
inputs (12 references), and OpenAI-compatible API. Estimated 10-100x cheaper
than Sora 2 per clip, making it ideal for high-volume ad production.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Covers image generation (Nano Banana Pro, Flux, Ideogram), video generation
(Veo, Kling, Runway, Sora, Higgsfield), and code-based video (Remotion).
Includes decision tree, cost comparison for 100+ variations, platform image
specs, and recommended hybrid workflow for scaled ad production.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Inspired by Anthropic growth marketing team workflows. Covers generating
headlines, descriptions, and primary text at scale across Google Ads, Meta,
LinkedIn, TikTok, and Twitter/X with platform character limits, performance
based iteration loops, and batch generation workflows.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
