Pablo Contreras · Vibe Coding · Lima, Peru · Remote LATAM/US
Vibe CodingI have been using Claude Code since launch (May 2025). Before that I was already piping context to Claude web with my own FastStruct extension. I build tools that accelerate that flow and publish real prompts. AI is a tool — I review, test and maintain the code myself.
01
Roles where I integrate AI into daily work (current role at Destacame).
No experiences match these filters.
02
FastStruct + CV builder (bypabloc/cv) + Astro monorepo of this portfolio. Tools I use to ship.
VS Code extension that visualizes and documents a project structure in seconds, optimized for AI workflows.
CV builder shipped in a single night with Claude Code (May 12, 2026). End-to-end demo of the modern vibe coding workflow: data → render → deploy.
This very portfolio: an Astro 6 monorepo with 5 independent sites deployed to Cloudflare Pages, optimized for GEO/LLM-SEO.
Public template I use as reference for new projects: .claude/ structure with skills, rules, hooks, devtools CLI, conformance and testing.
No projects match these filters.
Before Claude Code (May 2025) I used the Claude web interface to code. Sharing project context meant copy-pasting file by file manually.
I built FastStruct as a TypeScript VS Code extension: single command that scans the workspace, honors .gitignore, formats structured markdown and copies to clipboard. Configurable exclusion patterns.
5x faster personal workflow when instructing AIs. Published in the Marketplace, still useful even now with Claude Code for cases where I need compact context for sub-agents or external summaries.
I needed my own CV builder to iterate fast on CV data without hand-editing each export. Tight timeline: one night available.
Vibe coding session with Claude Code on 2026-05-12: iterative prompts to define data schema, render, themes and export. Human code review on every commit. No copy-paste without understanding.
Working MVP in a single night, public repo at bypabloc/cv. Demonstrates that Claude Code + human review combo ships product without sacrificing maintainability.
How I work with Claude Code, Cursor and subagents. No gimmicks: real prompts, transparent output, concrete metrics.
Before Claude Code I used the Claude web interface to code and piped context with my own FastStruct extension (published on the VS Code Marketplace). When Claude Code launched in May 2025 I was an early adopter from day one.
I was the one who introduced Claude Code in my current team at Destacame. I built a skills/rules/docs structure to standardize usage across developers, define canonical prompts per task, avoid duplicated discoveries and keep coherence in quality gates.
I regularly build CLIs (Python 3.14 devtools orchestrator, scan/verify/feature-test scripts, etc.) with Claude Code to centralize repetitive team processes. Each CLI is reviewed, tested and maintainable code — not throwaway scripts.
Claude Code as primary driver (91% CSAT, 54 NPS). Cursor for targeted visual refactors. Sub-agents (researcher / Explore / code-reviewer) on long outputs to avoid burning the orchestrator context.
"Explore these 3 files and tell me which components are refactor candidates due to duplication. Write your report to tmp/explore_X.md, respond a single line." → Cuts 80% of orchestrator token consumption.
Public repository I use as template and reference for the workflow: .claude/ structure, skills, rules, hooks, devtools CLI, conformance and testing. Serves as blueprint for new projects and to showcase the full vibe coding workflow.
CV builder built in a single night with Claude Code. Public repo bypabloc/cv. Demonstrates that the Claude Code + human review combo ships product without sacrificing maintainability.
Time spent instructing an LLM drops 5x with FastStruct. Zero "Generated with Claude Code" in commits/PRs via prepare-commit-msg hook that scans and strips AI attribution by company policy.
All code is reviewed, tested (Vitest + Playwright) and maintained by me. Quality gates in pre-commit and pre-push (Biome + tsc + astro check + coverage ≥80% + build). AI attribution forbidden by policy — not in commits, not in PRs, not in docs.
// Disclosure: all code is reviewed, tested (Vitest + Playwright) and maintained by me. AI is a tool, not an author. NEVER in commit attribution.
03
Current stack + domains where I have shipped to production.
04
Email or LinkedIn work best. Typical response under 24h.