Claude Skills - Biblioteca de Skills Validados
viktor-milev/claude-skills
Skills validados para otimização de prompts e análise de investimentos em Claude
Ver no GitHubColeção de skills prontos para uso em Claude, desenvolvidos e validados para trabalho analítico real. Inclui otimizadores de prompts, arquiteto de prompts e ferramentas de análise de investimentos, todos estruturados como arquivos SKILL.md autocontidos que integram nativamente ao Claude Desktop e web.
Instalação
npx convez-ai viktor-milev/claude-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Otimizar e arquitetar prompts complexos com rigor analítico
- Estruturar fluxos de prompt multi-etapas para problemas complicados
- Pressure-test teses de investimento e trades através de múltiplas perspectivas
- Transformar ideias em rascunho em prompts finalizados e prontos para usar
Veredito

Conjunto bem-executado de skills para otimização de prompts e análise de investimentos, com ênfase em rigor e aplicação prática. A biblioteca destaca-se pela abordagem de triage em otimização de prompts e pela metodologia de cinco-advisors em análise de teses, ambas validadas em uso real. Útil principalmente para usuários que trabalham com prompting estratégico ou análise quantitativa em Claude.
README
A tradução para o português deste README ainda está na fila.
claude-skills
A small library of Claude skills I've built and validated for my own work, published openly.
Each skill is a self-contained SKILL.md file with YAML frontmatter that Claude loads on demand. Drop the folder into your skills directory and Claude will pick it up.
Skills
Prompting
| Skill | What it does |
|---|---|
prompt-optimizer | Rigorous-mode prompt optimizer. Triages inputs (OUT OF PURVIEW / BORDERLINE / IN PURVIEW), decides single-turn vs. chained architecture, and rewrites IN PURVIEW prompts against a 9-dimension diagnostic. |
prompt-optimizer-flow | Flow-mode prompt optimizer. Turns rough, dictated, stream-of-consciousness input into a finished, paste-ready Claude keystone prompt in a single pass — no placeholders, no clarifying questions, no diagnostic theater. |
prompt-architect | Generative counterpart to the optimizers. Classifies a task against a 10-bucket taxonomy, decides single-turn vs. chained architecture, and produces a technique-grounded v1 draft (or a full chain) when you have no prompt written yet. |
Investment
| Skill | What it does |
|---|---|
investment-council | Pressure-tests an investment thesis, trade, or mispricing claim through five orthogonal advisors (Bear, Bull, Base Rate Skeptic, Market Structure Analyst, Risk Manager) who analyze independently, peer-review anonymously, and return a verdict tuned to decision (take/size it) vs. analysis (is it mispriced). |
Installation
Claude Desktop / Code:
- Clone or download this repo.
- Copy the skill folder you want (e.g.
prompt-optimizer/) into your Claude skills directory:- macOS/Linux:
~/.claude/skills/ - Windows:
%USERPROFILE%\.claude\skills\
- macOS/Linux:
- Restart Claude. The skill will load when its trigger conditions are met.
Claude.ai (web):
Skills can be uploaded to a Project. Open the skill folder, copy the contents of SKILL.md, and paste it into a new Project's custom instructions, or upload the file directly if your plan supports skill uploads.
Why these exist
Prompting
prompt-optimizer
Most published prompt optimizers solve the wrong problem. They optimize for polished prompts when what matters is the quality of the output. And they're miscalibrated at both ends of the difficulty spectrum — over-engineering simple prompts and under-architecting complex ones, especially missing the case for staged, multi-prompt structures.
The prompt-optimizer skill here is the result of a long iteration cycle (A/B tested against alternatives, 80% pair-level win rate at v3, refined further in v4) that taught me prompt optimization is mostly about triage: deciding what to leave alone, what to lightly polish, and what to rebuild.
prompt-optimizer-flow
The rigorous optimizer above assumes you already have a written prompt and the time to audit it. Flow mode is the other half of the same problem: you're dictating, mid-thought, and you want to start a Claude session now. Most optimizers break that moment — they interrogate you with clarifying questions, or hand back a template full of [fill this in] placeholders that doesn't actually run. Flow mode does the opposite: it strips the dictation noise, infers the intent aggressively, preserves your own voice, and returns a finished, paste-ready keystone prompt in one response. The discipline is in what it refuses to do — no scoring, no change log, no questions — so the prompt lands before the thought is gone.
prompt-architect
The two optimizers above both assume a prompt already exists — flow mode drafts one fast, rigorous mode audits one you've written. prompt-architect is the step before either: you have a genuinely complex task and no prompt in mind. It runs in reverse — given the problem, it works out which prompting technique and which architecture the task actually calls for. It classifies the task against a 10-bucket taxonomy, then makes a call most optimizers never surface: should this be a single prompt at all, or a staged chain with handoffs and checkpoints? The output is a technique-grounded draft you can run as-is, or hand to prompt-optimizer for a rigorous audit — the generative front end to the same pipeline.
Investment
investment-council
A compelling investment thesis usually fails for one of two reasons, and the second is the dangerous one. The first is sycophancy: the analysis tells you what you want to hear. The second is technical incompleteness — the thesis sounds airtight because it never had to clear a base rate, name who is on the other side of the trade, account for its factor exposure, or size itself against the book it is joining. investment-council convenes five orthogonal advisors — Bear, Bull, Base Rate Skeptic, Market Structure Analyst, Risk Manager — who analyze independently, peer-review each other anonymously, and hand back a verdict tuned to the question you actually asked: should I take this position, or is this asset mispriced. It live-verifies the thesis's load-bearing facts before it convenes — the single place it most often earns its keep — and it is calibrated for an institutional book, not a retail buyer who needs protecting from their own enthusiasm.
More skills coming soon
The other skills in this repo will follow the same pattern — built for real work, validated before publishing.
Maintenance and updates
Each skill is subject to periodic maintenance reviews and updates, as more capable models are being released by Anthropic. Please see the Changelog for further information. Current skills are optimized for Opus 4.8
License
MIT. Use them, fork them, modify them, ship them in your own tools. Attribution appreciated but not required.
Contact
@bizarcoin on X.