# Harrison Halperin — full site corpus > Machine-generated companion to llms.txt. Every visible project and > post on https://harrisonhalperin.com with a one-line blurb. > Regenerated by scripts/org/emit-evidence.py — do not hand-edit. Curated overview: https://harrisonhalperin.com/llms.txt Evidence pack (JSON): https://harrisonhalperin.com/evidence.json MCP (read-only): https://harrisonhalperin.com/api/mcp — manifest https://harrisonhalperin.com/.well-known/mcp.json Sitemap: https://harrisonhalperin.com/sitemap.xml ## About - About — https://harrisonhalperin.com/about - Work index — https://harrisonhalperin.com/work - Blog index — https://harrisonhalperin.com/blog - Knowledge graph — https://harrisonhalperin.com/graph - Fit (JD → evidence brief) — https://harrisonhalperin.com/fit ## Work - Agent CI/CD — No single identity can both write code and merge it. Edit agents only propose — branch protection and a separate merge actor decide what ships. https://harrisonhalperin.com/work/agent-cicd - Agentic Evaluation SDK — LLM interactions become structured evaluation cases — captured, scored with independent metrics, and mapped to a deployment risk tier that can gate a release. https://harrisonhalperin.com/work/agentic-eval - Agentic Project Management (Atlas) — A pull request with an open BugBot thread needs someone to fix it, reply to it, and shepherd it through checks to merge. Atlas automates that loop — detect, fi… https://harrisonhalperin.com/work/atlas - AI Hub — Every model interaction takes one governed path — access brokered centrally, inputs masked, and a ledger entry for each request. https://harrisonhalperin.com/work/ai-hub - AI inventory optimization — Generative AI applied to a concrete operations problem: where to hold warehouse inventory, and why. https://harrisonhalperin.com/work/ai-inventory - Automating Cloud Cleanup — Nothing in this tool can destroy anything. Orphaned Terraform Cloud resources get a signed, dependency-ordered destruction plan built from read-only evidence —… https://harrisonhalperin.com/work/destruction-bundles - AWS Artificial Server Quotas Observability Platform — Custom checks for the AWS limits that native tooling doesn’t monitor — security groups per VPC, access keys per user — surfaced as metrics that warn before gro… https://harrisonhalperin.com/work/quota-monitoring - BRIEF.md — An open standard for how AI agents discover, read, and hand off project context — plus the schemas and tooling that make it real. https://harrisonhalperin.com/work/briefs - Clear Diagrams — A manifest-driven plugin that regenerates layered Mermaid architecture docs from live repository evidence — plus preview-safe HTML and split sources so large d… https://harrisonhalperin.com/work/clear-diagrams - Connect Layer — A Cursor plugin that runs the diligence phase of a third-party OAuth integration — cited Terms-of-Service and pricing research, a registration checklist, a ver… https://harrisonhalperin.com/work/connect-layer-plugin - Constellation — A read-only-first project mapper: seed it with one ticket, wiki page, or repo, and it builds a tiered, evidence-backed connection graph — then a multi-page hub… https://harrisonhalperin.com/work/constellation - Context Compression — Converts long-form documentation into a compact, structured format that preserves decisions, interfaces, and constraints for language-model agents. Optimizes f… https://harrisonhalperin.com/work/context-compression - Developer Activity Ledger — A local-first pipeline that gathers a developer's scattered activity — commits, IDE and CLI chat sessions, ticket references, manual notes — and synthesizes it… https://harrisonhalperin.com/work/dev-activity-ledger - Diagram Generator — A pipeline that compiles a JSON relationship dictionary deterministically into Mermaid and SVG — when the architecture changes, the diagram is regenerated, not… https://harrisonhalperin.com/work/diagram-generator - Diagram-First — A Claude Code plugin that makes diagrams the agent's default output for plans, recaps, and system documentation. A router skill decides when structure should b… https://harrisonhalperin.com/work/diagram-first - Drive Mode — The same voice-first, multi-operator pairing idea, built twice: first as a daemon behind the Claude Code CLI, then as a Cursor/VS Code extension for a closed-c… https://harrisonhalperin.com/work/drive-mode - Happy Healthy Games — I thought game design would be fun to think about with a friend, so we vibe-coded this Godot 4 party game in one night — a signal-driven architecture with a he… https://harrisonhalperin.com/work/happyhealthygames - MeTime — A full-stack app that reads my calendar, categorizes each event with a small Claude agent, and renders the week as a Sankey diagram. https://harrisonhalperin.com/work/metime - Modular Orchestration System — A local multi-agent orchestration system. Agents coordinate through an explicit, inspectable loop — plan, act, review — rather than one opaque prompt. https://harrisonhalperin.com/work/mos - Outfra — Outfra is a design factory for local-LLM hardware — describe a budget, a workload, and a form factor, and its agents return parametric CAD, a priced bill of ma… https://harrisonhalperin.com/work/outfra - Publish Guardian — Publish Guardian gates the release of private repositories to public OSS behind a mechanical secret/PII scan, an LLM review lane, a 45-item OSS checklist, and… https://harrisonhalperin.com/work/publish-guardian - Quant H2 — A solo, AI-assisted quantitative trading system spanning four repos: a single-ticker research pipeline from ingestion through backtesting, a control plane that… https://harrisonhalperin.com/work/quant-h2 - Roler.ai — Coordinates three independently-deployable repositories — a Python backend, a React frontend, and a Terraform infrastructure baseline — inside one multi-root C… https://harrisonhalperin.com/work/roler-ai - Sensitive data audit — A discovery pipeline that locates sensitive data across enterprise object storage and reports where it lives — file, column, match count — never the values the… https://harrisonhalperin.com/work/sensitive-data-audit - Steganalysis toolkit — A toolkit for measuring how fragile or resilient hidden-information schemes are — LSB, invisible Unicode, visible watermarks — across text and images. https://harrisonhalperin.com/work/steganalysis - Terragram — A pipeline that reads architecture diagrams and emits Terraform that already satisfies org guardrails — opened as a reviewable pull request. https://harrisonhalperin.com/work/terragram - Thoughtbuilder — thoughtbuilder specifies an event-sourced control plane that turns agent chat reasoning into a versioned, queryable build log scoped to a codebase. The design… https://harrisonhalperin.com/work/thoughtbuilder - Tightrope — When a backend and frontend drift out of contract, this CLI catches it in CI — split repos or monorepo — and returns a machine-readable verdict. https://harrisonhalperin.com/work/tightrope - TLDR Research Ops — A six-stage agent pipeline design that triages a daily tech newsletter, verifies claims against evidence, and produces separate learning and implementation pla… https://harrisonhalperin.com/work/auto-research ## Writing - A picture is worth a thousand tokens — Making diagrams an agent's default output format took more than a skill description: trigger stacks, context budgets, a rendering failure, and a last line of d… https://harrisonhalperin.com/writing/picture-worth-a-thousand-tokens - When the repeating unit is the product — For platforms of many near-identical units, the product is the scaffold: discovery → analysis → generation → validation, with CLI tools collapsing context and… https://harrisonhalperin.com/writing/when-the-repeating-unit-is-the-product - Orchestration engineering: systems that spawn engines — After prompts and context, what compounds is meta-engineering: reusable orchestration that generates domain engines with roles, gates, and encoded expertise. https://harrisonhalperin.com/writing/orchestration-engineering-spawn-engines - Let the scripts judge and the agents fix — The most reliable pattern I keep reaching for: deterministic checks decide, and agents only ever propose. https://harrisonhalperin.com/writing/deterministic-guards-agentic-fixes - Nobody should both write and merge — Separating edit authority from merge authority is the one rule that makes agent-authored code safe to ship. https://harrisonhalperin.com/writing/who-edits-who-merges - One service layer, many surfaces — A browser, a Claude agent, and a background worker should all go through the same code. Anything else drifts. https://harrisonhalperin.com/writing/service-layer-many-surfaces - Compress for agents, not for humans — Context windows are attention budgets. Optimize value-per-token with offline distillation and structured handoffs, not prettier prose for people. https://harrisonhalperin.com/writing/compress-for-agents-not-humans - Why I build orchestration systems locally first — Reproducibility before scale: run agent loops on your own machine before they touch anything shared. https://harrisonhalperin.com/writing/local-first-orchestration - How a diagram-to-infrastructure tool turned into a research habit — Terragram kept losing the thread on long sessions. Fixing that meant studying how models use context, and that research became the reusable asset. https://harrisonhalperin.com/writing/researching-my-way-to-an-orchestration-system - Studying how hidden information breaks — You can’t reason about the robustness of a watermark or a steganographic scheme until you’ve tried, carefully, to break it. https://harrisonhalperin.com/writing/hidden-information-breaks - Agents deserve a real context format — AGENTS.md and SKILL.md solved pieces of agent context. BRIEF.md is my attempt at the missing discovery-and-handoff piece. https://harrisonhalperin.com/writing/a-standard-for-agent-context - Analogical reasoning as an engineering tool — Borrowing the shape of a solved problem speeds up the next one, as long as I stop where the domains diverge. https://harrisonhalperin.com/writing/analogical-reasoning - Dual channels and rule layers beat a longer prompt — When sessions grow, better prompts help less than splitting silent generation from review, and pushing recurring guidance into a cheap rule hierarchy so the wo… https://harrisonhalperin.com/writing/dual-channels-and-rule-layers