Recipes for agentic work in K-12 — working builds and ideas worth stealing, shared by the practitioners who shipped them. Start with the foundations, then pick a recipe.
What committee members actually run, grouped by solution — open a group to see who's doing it and how. Every claim here comes from a member's real setup; members add their own commentary over time.
One small login app holds the district's only OAuth redirect and issues a shared cookie carrying identity ONLY — every app re-resolves roles from the database per request, so revoking access takes effect on the next click. 11 apps on it.
Google-side auth (Firebase / Google) in front of her Google Cloud tooling.
Runs Google Firebase in his stack alongside Google Cloud and AWS.
Puts his idea-catcher portal behind Google SSO so submissions stay internal to district staff.
Ties the district AI chat platform to Google Workspace for role-based access across staff and students.
Google OAuth in front of a Supabase role/permission model with row-level security.
Commentary from the folks above lands here as members annotate the cookbook.
Google answers "who are you"; Active Directory governs what you may see. Never put roles in the shared cookie — identity in the cookie, authorization re-resolved from the database on every request.
Commentary from the folks above lands here as members annotate the cookbook.
Every agent gets its own delegated Google Workspace account, scoped like the staff member it assists — agents never borrow a human's credentials. Proven in the OpenClaw executive-cabinet rollout.
Commentary from the folks above lands here as members annotate the cookbook.
Runs OpenClaw agents in sandboxed AWS micro-VMs (Agent Core), with S3, EventBridge, and Lambda handling storage and scheduling.
AWS-hosted Claude builds, balancing in Gemini calls to manage image-generation costs; staff-facing tools first, which keeps the data-privacy picture simpler.
Self-hosted or AWS, plus a Vercel VPS — a Linux and open-source advocate.
Commentary from the folks above lands here as members annotate the cookbook.
Microsoft 365 throughout, with a production help desk platform and 150+ repositories — heavily invested in the DevOps and security side of making internal tools publicly accessible.
Commentary from the folks above lands here as members annotate the cookbook.
Runs all three clouds — Google, AWS, and Microsoft — with C# on the Microsoft side and JavaScript for internal apps.
Commentary from the folks above lands here as members annotate the cookbook.
BigQuery and Vertex AI pipelines on Google Cloud, plus an open-source Claude agent in Docker on a Mac Mini.
Shifting district hosting toward Google Cloud.
Built his idea-catcher portal on Firebase, with the agent-council logic in Google AI Studio.
District AI chat platform in GCP with a Firebase backend, moving from per-seat subscriptions to token-based pricing.
Google Workspace and Apps Script today, standing up a BigQuery warehouse on Google Cloud next.
Commentary from the folks above lands here as members annotate the cookbook.
District web apps on Windows IIS plus Linux/Python apps on district infrastructure — every app pool runs as its own AD service account with Integrated Security, so no app stores a SQL password. Boring, well-understood, already trusted.
Commentary from the folks above lands here as members annotate the cookbook.
A home lab of Linux servers under Proxmox — a mix of VMs and containers behind his agent systems.
Personal task agents on a Mac Mini (Perplexity's Personal Computer); exploring local model hosting through a regional service center (WNYRIC / NY BOCES).
A district GPU rig (RTX 3070) serving Llama 3.1 8B behind a token-gated proxy, plus a Hyper-V rack host — the standing rule is that anything touching student PII runs local, and only scrubbed summaries reach a hosted model.
Commentary from the folks above lands here as members annotate the cookbook.
A Python MCP server sits between agents and the warehouse (Redshift / PostgreSQL, PowerSchool) and enforces role-based access — a principal's agent only retrieves that principal's building. Agents ask approved tools; nothing queries production directly.
Connects agents to Google Workspace and his Zoho CRM through MCP servers; a master course database lives in Obsidian, with Markitdown parsing documents on the way in.
A Data MCP that limits results by the authenticated caller's role, now feeding live Atrium dashboards — plus write tools for OneSync and Roster Server so rules can be changed by asking rather than by hand.
Reverse-engineers vendor CLIs and APIs to reach data the district already pays for when the vendor offers no interface.
Commentary from the folks above lands here as members annotate the cookbook.
An IP-whitelisted server reaches the SIS database directly, and nightly / hourly exports land in a local Microsoft SQL Server that hosts all district data. Agents reach it only through an MCP server whose stored procedures re-resolve the caller's scope on every call.
Commentary from the folks above lands here as members annotate the cookbook.
Moving de-siloed data out of Google Sheets into a BigQuery warehouse, backed by a Gates Foundation grant.
BigQuery with dbt-style data prep feeding her Vertex AI, Gemini, and Claude Code workflows.
A BigQuery warehouse with dbt and Dagster, building out the structures from API calls to event streams.
Commentary from the folks above lands here as members annotate the cookbook.
Dump district information into an organized structure and let AI be the middle layer — the semantic-intelligence approach he floated on the May call.
Runs a constant collection of course data into the Obsidian master database, with agents crawling the web for education updates.
Commentary from the folks above lands here as members annotate the cookbook.
Topics the group hasn't dug into yet — named so you know they matter, empty until we have real practice to report.
How we scope what agents may see and do, and when a human signs off.
no group content yetHow a build moves from one district to another without starting over.
no group content yetWhat agent-built, student-facing tools must meet — and how we verify it.
no group content yetFilter by what you're after, click a card for the full recipe — prerequisites, data needs, build steps, and a prompt you can hand your own coding agent.
Nothing matches those filters — try clearing one.
Shared skills and agent files — communal by design, stamped with who maintains each one. Grab what's useful, contribute what you've built.
Autonomous build-to-done: implements, runs the full Definition of Done (build, lint, typecheck, full test suite, Playwright with screenshots), opens a PR with visual evidence, then watches CI and AI reviewers and fixes every round until clean.
Clarifies intent, researches in parallel, designs the approach, and emits a task breakdown plus a machine-checkable Definition of Done, optionally filed as GitHub issues that /lfg can drive to done.
Parallel development with git worktrees: create, list and remove them, plus a clean command that prunes worktrees, deletes merged branches, and closes orphaned issues.
Configures a project's verification gate and review agents by writing .psd/verify.json (build, lint, test, typecheck and E2E commands, strictness, and which AI reviewers are active).
Strips the showing-off from any artifact (code, prompt, skill, doc, landing page) by testing it against a blunt intended user who only cares about getting the job done.
Multi-LLM parallel research: decomposes a question, sends the parts to several models, and synthesizes the results.
LLM Council research pattern: several frontier models answer independently, peer-review each other, and a final pass synthesizes the result.
A decision council for important choices: seven structured perspectives, adapted from de Bono's Six Thinking Hats, deliberate before a recommendation.
Writes and edits technical content in Google developer-documentation style, distilled from the full Google style guide, including its word list.
Builds presentations using principles from Garr Reynolds, Nancy Duarte, Guy Kawasaki, Seth Godin, and TED.
Organizational strategic planning for K-12 school districts using a research-backed four-stage process.
Converts a PDF to clean Markdown, describing image content as text so the document is ready for AI tools.
Generates images with Google's Gemini image model using workflow-based prompting.
Local text-to-speech on a Mac using MLX and the Kokoro model, with no cloud calls.
Generates narration, podcast audio and voice-overs from text with the ElevenLabs API.
Read-only export and archive of DocuSign envelopes, templates and signed documents, built for migrating to Documenso.
Tracks Washington State K-12 education bills by pulling everything from the education committees through the legislature's SOAP API, with no keyword filtering that could miss a bill.
Generates self-contained single-page HTML artifacts (reports, specs, explainers, interactive editors) with deliberate, non-generic design.
Drop it in your project repo and point your coding agent at it — it studies the repo, interviews you, and packages a ready-to-review cookbook submission: recipe, prereqs, description, sample prompt, screenshot, and a profile update.
Your agent updates YOUR practitioner profile, recipes, and shelf skills straight in the group's knowledge base, using your own Coda token. Only rows whose Credit is you, ever; the public site rebuilds nightly.
Distribute the group's skills as a Claude Code plugin — one install gives a district's builders the shared toolkit, with the OpenClaw harness routing models.
The cookbook grows one shipped build at a time — and it's self-serve. The recipe-prep agent digs through your project's repo and packages a ready-to-review submission (recipe, prereqs, description, sample prompt, screenshot). The update skill lets your coding agent update your own profile and recipes directly, with your own Coda token — the site rebuilds nightly. Or just email Luke and we'll write it up together.
or email Luke →