Workshop · 2026-07-17 · John Lindquist
Agenda — Agentic Software Factory
Eleven sections, one system. Each section below links its guide (skim it, lift the prompts), a session replay you can present from, and a Socratic dialogue that argues the idea out loud.
Start here
References & Resources Every tool in this workshop, with 81 evidence-backed alternatives — so you can build the same system out of the pieces you already have. Open the referenceThe eleven sections
-
How Agentic Work is Evolving
An agent re-audits last month's field-guide directives against local evidence — conversation history, prototypes, and ~2,000 Oracle sessions — to score which laws held and which evolved.
-
Configuring Your Environment
Two files in
.pi/rewire every shell command the agent runs —rmquarantines,gitrefuses force pushes,linearself-assigns — and every override explains itself on stderr at the moment the agent acts. -
Custom zsh Functions
One- and two-letter zsh functions are the cheapest agentic layer: wrap flag-heavy agent CLIs once, compose the wrappers into resume, fork, and account-switching families, then build
pirfrom scratch live. -
Prompting with Images
An annotated screenshot replaces a paragraph of ambiguous prose — the CleanShot arrow does the pointing, the words typed on the image carry the intent, proven end-to-end by real sessions from local history.
-
Prompt Builder
prompt-builder (
p) gives the prompt its own composing surface — skills, file search, slash commands, cross-session history — then hands off to Pi, Codex, or Claude, and handoff mode turns any shell function into a launch target. -
mdflow
A repeatable agent job becomes a markdown file in your repo — context preloaded with backtick inlines, ambient config stripped, behavior eval-gated, changes proposal-first — and mdflow (
md) is the control plane. -
Why Pi, Herdr, and Ghostty
An agentic terminal is a stack, not an app: Ghostty draws, Herdr exposes every pane over a unix socket, Pi drives that same socket, Karabiner makes it all one keystroke — ending in keeper agents and planned swarms.
-
Variations (9x Image Mockups)
Image generation is the cheapest way to explore design space: nine wildly different mockups in minutes (Codex
$imagegen, agy), pick a direction with your eyes, hand the winner back to an agent as a reference image. -
This or That — RLHF
You can't articulate your taste, but you can pick between two things all day: an Elo duel harness turns votes into a preference record your agent reads — it's how this very page theme was chosen (3 rounds, 130 votes).
-
Maximum Intelligence and Video
Oracle bundles up to ~900k tokens of context and drives a real logged-in browser into ChatGPT and Gemini — subscription-priced frontier thinking for the hard questions, and the only path to a model that actually watches your videos.
-
Debugging with Markers
A press-to-capture recorder pins one sentence of human testimony to a bounded ring buffer of app events and exports a redacted payload any agent can debug from — a complete marker loop in under 300 dependency-free lines.
For your agent
The machine copy of this agenda — every section with its guide, replay, and dialogue URLs plus its own agent setup file — is agent.md. Paste this into your agent:
Fetch https://ember-bodhi-scrr.here.now/agent.md?raw — the machine-readable agenda for the Agentic Software Factory workshop. List the sections, then follow the per-section agent guide for the one I name and set it up in my current project.
The whole workshop as one flow
flows/workshop.md is an mdflow flow that bundles the entire workshop into a single agent run: it preloads every section's machine guide plus all eleven Session Replays and Socratic Dialogues, fetched live from these sites the moment you run it — always current, nothing pasted. Give it a task and your agent answers with the full workshop in context:
mkdir -p flows && curl -sf "https://ember-bodhi-scrr.here.now/flows/workshop.md?raw" -o flows/workshop.md
npx -y mdflow flows/workshop.md --_task "Walk me through the workshop, then set up the section I pick" --engine claude
The --engine flag picks your agent CLI (claude, codex, gemini, pi, …); drop it to use your mdflow default. It's a heavyweight load by design — the full replay and dialogue set rides along — so use it when you want the whole workshop reasoned over, and the per-section prompt above when you only need one setup.