Claude Code Prompt Observatory
ftvision/agent-prompt-observatory
Histórico visual da evolução do system prompt do Claude Code com diffs inline
Ver no GitHubAgente que fornece uma interface interativa para explorar a evolução do system prompt do Claude Code ao longo de 289 versões publicadas. Visualiza mudanças em linha do tempo, estrutura de componentes (ferramentas, seções de prompt, mensagens) e correlaciona alterações com lançamentos de modelos Claude. Ferramenta de pesquisa para entender como o comportamento do agente foi refinado ao longo do tempo.
Instalação
npx convez-ai ftvision/agent-prompt-observatory1 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
- Pesquisa e análise de como prompts evoluem com lançamentos de modelos
- Compreender mudanças estruturais no sistema de agentes Claude Code
- Comparar comportamentos de diferentes versões de prompts históricos
Veredito

Ferramenta especializada e bem executada para pesquisadores e engenheiros de prompts que querem entender a arqueologia técnica do Claude Code. Excelente para análise comparativa, mas requer interesse específico em evolução de prompts.
README
A tradução para o português deste README ainda está na fila.
Claude Code Prompt Observatory
Live: https://ccprompt.feitong.phd/
A research interface for exploring how Claude Code's system prompt has evolved across every published version. Prompt captures come from marckrenn/claude-code-changelog (the rolling cc-prompt.md plus its commit history); historical versions through 2.1.112 were originally captured from cchistory.mariozechner.at before that archive went offline. A Python analyzer parses each capture into a structured model, and a Vite static UI renders that model as two views.
Views
Structure — single-version inspector. The current prompt rendered as a 3D layered model, stacked top-to-bottom in Anthropic's prompt-cache order: Tools at the top, System Prompt sections in the middle, User Prompt slabs at the bottom. Click any slab to read its prose / schema / metadata in the side panel.
Evolution — temporal stream graph across the entire 289-version range. Each row is a component (User Message total, System Prompt sections, every individual tool); the bar at each version's column has height proportional to character count. Empty cells mark releases where a section was absent. A model-release axis above the version axis labels Claude 4 → 4.7 launches with vertical guides threading through the matrix, so you can see when prompt changes correlate with model releases. Click any component name to expand a change-log timeline below it; click any change-log entry to see the line-level diff of that version against its predecessor.
Project layout
claude-code-analyzer/ Python package + `cca` CLI (parser, exporter, tests)
data/raw/ Captured *.md prompt files, one per version
data/versions_meta.json {version, release_date} index for the corpus
ui/ Vite app (single-page, no framework)
src/views/ structure.js / evolution.js + per-view CSS
src/components/ version-picker
src/data/loader.js fetch wrapper for ui/public/data/*
public/data/ JSON output the UI reads at runtime
scripts/fetch_versions.py pulls new captures from upstream
.github/workflows/
update-data.yml cron: fetch + re-export + commit
deploy.yml on push: vite build + GitHub Pages deploy
Data flow
Two stages, both fully scripted:
scripts/fetch_versions.pypolls the upstreamversions.json, downloads any new.mdcaptures intodata/raw/, and updatesdata/versions_meta.json.cca export data/raw --output-dir ui/public/dataparses each capture into structured components (system_message sections including the prompt preamble, tools with prose+schema, user_message xml_tags) and writes:meta.json— version index with release datesstructures.json— per-version skeletons (titles + char_counts, no prose)components/{version}.json— full prose/schema text per component, used for inline diffsdiffs.json— pairwise structural deltas
The UI loads meta.json + structures.json upfront, then fetches components/{version}.json lazily when the user opens a diff.
Continuous update + deploy
.github/workflows/update-data.yml runs daily at 06:00 UTC:
- Runs
fetch_versions.pyto pick up new upstream captures. - Runs
cca exportsoui/public/data/reflects the new corpus. - If anything changed, commits
data/+ui/public/data/to master and pushes.
The push triggers .github/workflows/deploy.yml, which runs npm ci && npm run build in ui/ and publishes ui/dist/ to GitHub Pages via actions/deploy-pages.
End-to-end, a new upstream version becomes a deployed page update with no human in the loop.
Run locally
Need: Node 20+, Python 3.11+, uv.
Pull the data:
uv run python scripts/fetch_versions.py
Generate the UI's structured data:
cd claude-code-analyzer
uv sync
uv run cca export ../data/raw --output-dir ../ui/public/data
Run the UI:
cd ui
npm install
npm run dev # opens http://localhost:5173/
For a production-style preview:
npm run build && npm run preview
Run the analyzer tests:
cd claude-code-analyzer
uv run pytest
Architecture notes
- No framework. The UI is plain JS, CSS, and HTML wired together with Vite. ~1200 lines of view code.
- OKLCH everywhere. Color tokens (
--viz-user,--viz-system,--viz-tools) drive both the 3D slabs in Structure and the bars in Evolution; one source of truth. - The diff is line-level LCS with a fallback to a linear naive diff above 2M cells, so a pathological 5000-line schema diff can't lock the tab.
- Hardened error paths. Both top-level data load and individual section render failures show recovery UI with retry, not a blank screen.
- Forced-colors mode preserves the data viz by opting bars + legend swatches out of color forcing while still adopting
Highlight/Marksystem colors for selection states.
Credit
Prompt captures: marckrenn/claude-code-changelog (current source — extracts the prompt from the published npm bundle on each release) and cchistory.mariozechner.at (historical source for versions through 2.1.112). Without those archives this project would have nothing to look at.