ALCF AI Agent Coding Talk
bricevideau-ai/alcf-ai-agent-coding-talk
Estudo de caso: 17 sessões, 2.650 prompts, 2 projetos, métricas reais
Ver no GitHubRetrospectiva de 15 semanas usando Claude Code em dois projetos reais de engenharia de performance (CCS e rust-gpu + claspr). Inclui deck de apresentação, métricas de sessão, análise de custos e ferramentas de reprodução. Recurso educacional completo documentando a experiência prática de codificação com agents de IA.
Instalação
npx convez-ai bricevideau-ai/alcf-ai-agent-coding-talk1 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
- Estudar padrões reais de uso de Claude Code em projetos grandes
- Analisar custo-benefício de agentes IA para engenharia de performance
- Entender fluxo de trabalho colaborativo entre múltiplos agentes
Veredito

Documentação técnica honesta e reproduzível de uso real de Claude Code. Mais valioso como estudo de caso para engenheiros sênior do que como ferramenta reutilizável — excelente para quem quer entender dinâmica prática de agentes IA em projetos complexos.
README
A tradução para o português deste README ainda está na fila.
From Directing to Dialogue
Fifteen Weeks of AI Agent Coding for Performance Engineering — a first-person retrospective from two ALCF-adjacent projects (CCS and rust-gpu + claspr).
Source materials for an ALCF talk evaluating my use of Claude Code on two real projects, one I owned and one I didn't.
Contents
| File | What it is |
|---|---|
ai-agent-coding-alcf.pptx | The deck (38 slides, ALCF template, ~30–35 min) |
ai-agent-coding-alcf.pdf | PDF render for quick viewing |
build_deck.py | python-pptx generator — re-run to rebuild from edits |
deck_outline.md | Slide-by-slide outline plus the numbers used |
session_stats.json | Mined metrics from the surviving session JSONLs across both laptops |
ccusage/ | ccusage@latest cost snapshots from both laptops + a README explaining the merge protocol |
CLAUDE.md | Orientation for the next Claude agent picking up the deck |
ALCF Presentation Template.pptx | The ALCF template — committed so the build is self-contained |
Rebuild
pip install python-pptx
python3 build_deck.py
# Optional: render to PDF for QA
soffice --headless --convert-to pdf ai-agent-coding-alcf.pptx
No external paths: build_deck.py resolves the template relative to its own location.
Scope and provenance
- Two projects covered:
- CCS (
argonne-lcf/CCS) — C99 autotuning configuration-space library; I wrote and maintain it. 4 sessions, ~600 user prompts, 96 PRs merged in 6 weeks. - rust-gpu + claspr — OpenCL Kernel execution model added to
Rust-GPU/rust-gpu(PR #3 on the fork, +24,540 / −595 across 761 files) plus the newclasprsingle-source OpenCL host layer (~81K LOC). 13 sessions, ~2,059 user prompts.
- CCS (
- 17 sessions total across both laptops, ~2,650 user prompts over 15 weeks (Feb 25 → Jun 8, 2026).
- Token/cost metrics live in
session_stats.jsonandccusage/. Only 5 of the 17 session JSONLs survived Claude Code's 30-day eviction by the time mining was run (a910ecaa, 11e6e374 on the Arm Mac; 4862cd6d, 61dc0523, afa1ce4c on the native-Intel-Ubuntu laptop) — those are direct measurements; the other 12 sessions' prompts survive viahistory.jsonlbut their per-call tokens are gone. Slide 30 in the deck projects the full-project bill from the 5-session sample. - The deck is built from the official ALCF PowerPoint template, which is committed to this repo so the build is self-contained.
Authoring notes
The talk itself was prepared with Claude Code (Opus 4.7) — first authored on the Arm Mac laptop (Ubuntu environment on macOS host), then reviewed and enriched on the native-Intel-Ubuntu laptop, with two agents collaborating via the GitHub repo. Mildly on-the-nose given the subject matter. The build script, outline, session-mining, and ccusage snapshots all sit in this repo so the whole thing is reproducible end-to-end.