convez.ai
SkillFinanceiroAvançado
tm4gic

Analista de 10-K e Transcrições

tm4gic/10k-transcript-analyst

Compara 10-K e call de resultados para expor silêncios deliberados

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill Claude que lê simultaneamente o 10-K de uma empresa e sua transcrição de call de resultados, gerando um memorando analítico com TL;DR, destaques, análise de gaps (o que o 10-K sinaliza mas o call ignora) e perguntas de follow-up. Identifica sistematicamente o que a administração optou por não discutir.

Instalação

$npx convez-ai tm4gic/10k-transcript-analyst

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 6 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Due diligence para investidores (buy-side)
  • Análise de risco de disclosure e gerenciamento de narrativa
  • Pesquisa de compliance e auditoria interna
  • Identificação de gaps entre filing público e comunicação oral

Veredito

Skill bem arquitetado com escopo fechado e produção rigorosa. O diferencial real — análise de gaps sistemática entre 10-K e call — resolve um problema legítimo de pesquisa financeira. Requer domínio de Python, EDGAR e interpretação regulatória, mas executa com precisão e transparência de fontes.

README

A tradução para o português deste README ainda está na fila.

10-K + Transcript Analyst

A Claude Skill that reads a company's 10-K and its earnings call transcript side-by-side, then produces an analyst memo:

  • TL;DR — what a PM needs to know in 30 seconds
  • Highlights — the 5–8 things that actually mattered
  • Gap analysis — items the 10-K raised that the call did not address (the value-add)
  • Beyond the 10-K — guidance and color from the call with no parallel in the filing
  • Follow-up questions — what a buy-side analyst would put on the next call

Essentially: a junior analyst who actually reads both documents and notices what didn't get said.

Why this exists

Most "earnings summarizers" read the transcript and stop there. The interesting question isn't what management chose to talk about — it's what they chose not to. A 10-K discloses material risks and business facts that the call often steers around. This skill triangulates the two so the silences become legible.

It's also a clean demonstration of what a well-designed Claude Skill looks like: tight scope, real data sources, explicit grounding rules, and an output format with teeth.

Install

# Clone into your Claude skills directory
git clone https://github.com/tm4gic/10k-transcript-analyst.git \
  ~/.claude/skills/10k-transcript-analyst

# Python deps
pip install requests beautifulsoup4 lxml pypdf

Set a contact email for EDGAR's User-Agent policy (required by the SEC):

export EDGAR_UA="your-name (you@example.com)"

Use

Save the earnings call transcript to a file (most IR sites publish a PDF; Seeking Alpha and Motley Fool publish HTML you can save as .html or paste into a .txt). Then, in Claude Code:

Use the 10k-transcript-analyst skill on NVDA with the transcript at ~/Downloads/nvda_fy25_q4.pdf.

Or invoke the ingester directly to inspect the cleaned text first:

python scripts/ingest.py NVDA ~/Downloads/nvda_fy25_q4.pdf
# → out/NVDA_10k.txt
# → out/NVDA_transcript.txt
# → out/NVDA_meta.json

Then ask Claude to generate the memo from those files following SKILL.md.

What you get

See examples/AAPL_FY2024_memo.md for an illustrative output.

The memo follows a fixed structure: TL;DR, Highlights, Gap Analysis (the centerpiece), Beyond the 10-K, Follow-Up Questions, Sources. Every number is cited as either (10-K) or (call) — no fabrication, no hand-waving.

Grounding

The skill is instructed to:

  • Never invent financials
  • Be honest about silence — if the call didn't address a 10-K item, say so plainly
  • Quote sparingly (≤15-word phrases, attributed); paraphrase the rest
  • Mark uncertainty when section extraction is incomplete

See SKILL.md for the full rules.

Limitations

  • Only the most recent 10-K is fetched. To analyze an older fiscal year, modify scripts/ingest.py to pass an accession number.
  • Transcript parsing is best-effort. Speaker detection relies on a Name: ... pattern; unusual formats may need cleanup.
  • Section extraction uses regex against EDGAR HTML — most modern filings parse cleanly, but older or unusually-formatted 10-Ks may not split well. The skill will flag this rather than fabricate.
  • This is a research aid, not investment advice.

License

MIT

Ver o original no GitHub

Bom para

Relacionados