bolivian-peru prompt library
Free, production-grade AI skills & explainers — battle-tested on real software, not toy demos.
A growing, open collection of skills (reusable AI-agent prompts in the
Anthropic skill format) and explainers.
Each one is something we actually built and shipped — extracted into a portable form so
you can drop it into Claude Code, Cursor, Copilot, or any LLM agent and get the same
result on your codebase.
Every skill here follows two non-negotiable principles:
- Deep, real analysis — no mock data. Skills instruct the agent to read your actual
code/config and verify every claim against a real file. Nothing is invented or
placeholdered. If a value can't be verified, it isn't shown.
- No fluff. Concrete steps, real code, the hard-won gotchas. Optimized for an agent
to execute, not to read nicely.
Skills
| Skill | What it does |
|---|
| system-map-visualization | Deeply analyzes any codebase, then builds a live, animated 3D WebGL "system map" of the whole ecosystem + a static PlantUML architecture diagram, both from one code-verified data file. Real-time telemetry, clickable nodes, guided end-to-end journeys. No backend of its own — one self-contained static HTML file. |
| niche-pitch-engine | Runs a verified B2B outreach pipeline for any niche: industry-mechanics research, registry/signal-driven target shortlist, per-company dossiers, two-step pitch emails in the target language, adversarial claim-by-claim QA, and a persisted approval queue behind a hard send gate (opt-outs, dedup, SQL-enforced caps, mandatory legal footer, human approves every send). Truthful-only by construction — no invented facts, no fake social proof. |
| agentic-chat-interface | Builds a production-grade streaming agent chat front-end: SSE-streamed conversation, live tool-call timeline, collapsible "thinking" panel, stop/retry/abort, graceful failure. Framework-light (React + fetch + Streams API + Tailwind) — no chat library, no SSE library. Decoupled from any backend by one 10-event SSE contract. |
| agent-ops-board | Autonomous company engine: authority, control plane, verified outcomes, board as record. Runs a company or multi-sided ecosystem (customers, suppliers, partners, resellers) with one top authority (owner or approver groups), one coordinator agent and bounded executors: a decision-rights matrix and earned-autonomy ladder, kill switch, per-loop pauses and outcome circuit breakers, an idempotent side-effect state machine that never blindly retries, reviewed bulk batches, verification at the source, owner briefs, and a paced board helper (export / health / post with write-then-read-back). |
| brotli-decoder-cve-fix | Takes a server from "a 9.2 critical Brotli bug hit Node.js" (CVE-2026-111497, a.k.a. Node2Shell / HTTP2Shell) to a verified answer and a closed exposure: checks the advisory against primary sources, inventories every runtime that bundles Brotli (Node.js, Bun, Python, curl, proxies), proves with a safe probe whether attacker-controlled Brotli reaches a decoder (Express 5 / body-parser inflates br request bodies before auth), blocks it at the nginx edge in minutes, then patches runtimes with zero-downtime restarts and verifies the running processes. Includes a read-only check.sh and drop-in nginx guard files. |
More coming. Watch / star the repo.
Social previews
Each skill has its own preview card (<skill>/assets/og.png). GitHub folder links always show the repository's
preview, so share a skill with its page instead, which carries that skill's card:
https://bolivian-peru.github.io/bolivian-peru-ai-skills/<skill>/ (for example
agent-ops-board).
Cards and pages are generated by node assets/og/build.mjs from assets/og/skills.json; add a row there for a new skill.
How to use a skill
Each skill is a top-level folder containing a SKILL.md (the prompt) plus any helper files.
Pick whichever path fits your tool:
Claude Code / Claude Agent SDK — drop the skill folder into your project's
.claude/skills/ (or ~/.claude/skills/). Claude discovers it by the description: in the
frontmatter and invokes it when your request matches.
Cursor / Copilot / Windsurf / any chat — open the SKILL.md, paste its contents into
the chat as context, and add a one-line ask, e.g. "Follow this skill to build a system map
of THIS repo." The agent reads your code and produces the artifacts.
Plain API — send the SKILL.md body as a system/developer message, then ask the model
to apply it to your repository (give it file access via your tool layer).
The skills are tool-agnostic. They assume the agent can read files and run shell commands
in your project. They never require our infrastructure.
Why these exist
These come out of real production work. When something gets solved cleanly enough to
generalize, the reusable core is extracted and published here — free, MIT. Use it, fork it,
ship it. No attribution required.
If a skill saved you time, a ⭐ is appreciated — it tells us which ones to keep investing in.
Contributing
PRs welcome — see CONTRIBUTING.md. The bar: it must be something real you
shipped, it must analyze real systems (no mock data), and it must be concrete enough that an
agent can execute it end-to-end without hand-holding.
License
MIT © 2026 bolivian-peru — see LICENSE. Do whatever you want.
