Directive Shorthand — the world calls it Lingo.
A tiny shorthand of directives — short, sensible tokens typed right into any message — that tell any AI agent exactly how to reply, what to remember, and how to work. Any agent. Any platform. Any toolkit. The vocabulary belongs to you, not to a vendor: learn it once, and it works everywhere an agent speaks Lingo.
If you can dream it, you can direct it — in under ten characters. That's not a slogan; it's the design: the grammar grows by whole domains of behavior, so anything an agent can do can be steered the same few-character way.
Thirteen shorthand directives · three families · zero memorized shorthand
The Problem
You steer today's AI the way you'd steer a stranger: long instructions, repeated every time, in every app — or a settings page you have to go and find. Nothing you learn in one place works in another. There has never been a standard for the simplest act in the AI age: telling the machine how you want it.
Every assistant hides your preferences in a differently named page, with a different format and a different character limit. Your instructions don't travel — an entire cottage industry exists just to re-paste the same prose into each one.
Slash commands are real — inside one app. A command in one chat tool means nothing in another, and they trigger that app's actions, not your intent about how the reply itself should look, sound, or be checked.
People already invent shorthand — ELI5, TL;DR — because the need is real. But folklore has no registry and no guarantees: every model interprets it differently, and it covers a sliver of what an agent actually does. It steers text. It cannot steer memory or work.
Try the Lingo
This is the whole vocabulary — the complete Directive Shorthand 1.0 registry. Tap tokens to stack them into one message, and watch two things happen: the meaning compose in plain English, and the fastest form derive itself. Nobody memorizes the short forms — stop typing when the word can only be one thing, and it fires.
re: the reply — how the answer arrives
mem: the memory — what the agent keeps about you
do: the work — how the task gets done
——That message tells the agent toTap some tokens above — your stack's meaning composes here, in plain English.
The Three Families
Three families, thirteen tokens — that's the whole of 1.0. The prefix names the part of the agent you're steering, and each token owns exactly one axis inside its family, which is why stacks can never contradict themselves.
How the answer arrives, how long it is, how it talks, how it's shaped.
Axes — Channel voice · show · Length brief · deep · Register plain · Structure recap · steps · vs
| Token | Shortest | What it does |
|---|---|---|
| re:voice | vo | Reply as a voice note.Mode re:voice on makes it sticky until off. |
| re:show | sh | Answer visually — diagram, chart, map, or image; the agent picks the medium that fits. |
| re:brief | br | The short version: verdict first, essentials only. |
| re:deep | de | The full version: research-grade, evidence and sources shown. |
| re:plain | pl | Plain, jargon-free language.Mode Bare colors one reply; on sticks. |
| re:recap | rec | Catch me up — state, decisions, what's open. |
| re:steps | st | A numbered procedure, no essay: one action per step. |
| re:vs | vs | Side-by-side comparison, trade-offs named. |
What the agent remembers about you — saved deliberately, corrected cleanly, ranked by what matters.
Axes — Operation save · fix (last wins) × Priority pin · recall stands alone — a read never stacks with a write
| Token | Shortest | What it does |
|---|---|---|
| mem: | — | The bare prefix saves what the message names. The family's most frequent act costs zero letters — and the agent confirms exactly what it saved. |
| mem:pin | pi | Save it at special priority: surfaces first, never pruned quietly. |
| mem:fix | fi | This message corrects a memory — replace the old version, confirm the change. |
| mem:recall | re | Answer from what's actually remembered — and admit plainly when nothing is. |
How the task itself is done — invisible in the reply's shape, visible in its evidence. That's precisely why it isn't under re:
| Token | Shortest | What it does |
|---|---|---|
| do:check | ch | Verify the claims against live sources first — and say what was checked. |
| do:draft | dr | Produce the artifact, take no external action. Nothing is sent, posted, or paid. |
In the Wild
Drawn from the specification's own conformance examples — the tests a fresh agent must pass to claim it speaks Lingo. Each card is one ordinary message and what the token inside it makes happen.
"Book the venue and get me the quotes. re:voice"
The agent
Replies with a voice note containing the answer. Nothing is saved — a bare token steers one reply and evaporates.
"re:voice on — and did the venue reply?"
The agent
Answers in voice, writes "voice replies on" into its memory, and confirms the save. Tomorrow's token-free "Any news?" still gets a voice note — until re:voice off clears it, also confirmed.
"How do the two providers compare on price? re:show"
The agent
Answers with a chart — the medium that fits a price comparison — plus at most a one-line takeaway. An essay with a picture attached is a failed re:show.
"Did the venue confirm? re:vo/br"
The agent
A short voice note. vo and br are the shortest prefixes that can only mean voice and brief — derived on the spot, never memorized. An ambiguous prefix would not fire at all.
"My sister's name is Ana. mem:" …later… "Correction — Ana, one n. mem:fix/pin"
The agent
Saves the fact and confirms it. Later, replaces it with the correction and marks it pinned — corrections at top priority, confirmed both ways.
"Write the landlord email about the leak. do:draft" · "Is this supplement safe with my meds? do:check"
The agent
Produces the email and stops — it is not sent. And before answering the safety question, verifies against live sources and states what it checked. Stack them — do:check/draft — for verified content, never executed.
How It Works
Full words come first, so a beginner guesses the Lingo cold. Everything that makes it fast is a rule, not a list — learn the rule once and it applies to every token that will ever exist.
Directives compose across axes: channel + length + register + structure all apply at once. Within one axis, the last mention wins — so a stack can never contradict itself.
Stop typing when the word can only be one thing — the rule network engineers have typed on Cisco gear for decades. An ambiguous prefix never fires and is never guessed: you just add a letter.
Forgot the vocabulary? Type just the prefix — re: — and the agent lists the whole family, one line per token, in your language. The menu is a behavior of the agent, not a manual you have to find.
The Name
Directive Shorthand
The formal name — the title on the standard, the name in citations and conformance claims. It follows the lineage of the great shorthand systems: modifier + Shorthand, one singular method, named for what it shortens. The directives are the intents; this is the notation that compresses them.
Lingo
What everybody actually calls it. Standards live two lives — ECMAScript is the specification, JavaScript is what the world says — and a vocabulary people talk about needs a name that survives a sentence: "just use Lingo, it's easier." So: Directive Shorthand on the title page, Lingo everywhere else. Learn the Lingo.
shorthand directives
The individual commands the system defines — re:voice, mem:pin, do:check — each one a full directive ("reply in a voice note", "save this at top priority", "verify before you answer") compressed into a few characters. The system is the shorthand; the commands are shorthand directives.
Open + Supercharged
Directive Shorthand is published openly, in full, for anyone to adopt — that is what makes it a standard rather than a feature. Every conforming agent speaks the same Lingo with the same meanings. What differs is how much power stands behind each command — and that is AgentForge's home ground.
Any agent, on any platform, that implements the specification gives you the full experience of the standard: the same thirteen tokens, the same grammar, the same stacking and type-ahead, the same discovery behavior — tested against the published conformance examples.
A plain implementation steers with words alone: re:voice formats a reply, do:check asks for care, mem: writes to whatever memory the host provides. Correct — and complete, as far as the spec reaches.
AgentForge toolkits don't just honor the Lingo — they back it with built-in access. When a command needs the world, the toolkit reaches it through real, programmatic tools instead of good intentions.
The grammar — families, stacking, type-ahead, discovery — plus five tokens: voice · show · brief · deep · plain. Any agent passing Level 1 gives you the essential Lingo on day one.
Structure (recap · steps · vs), the memory family (mem:), and the work family (do:check · do:draft) — ratified separately, so a minimal agent can be honestly conformant without them.
The registry is append-only: a ratified token never changes meaning, and a retired name is never reused. Grammar changes need a major version; new tokens are minor ones. Private experiments live in a reserved x- space where they can never collide with the standard.
Both documents are published in full — the spec agents load, and the research memo that records why every decision was made. Nothing about this standard is a black box.
International
The canonical tokens never change, anywhere on earth — like the ON/OFF on every device you've ever owned. Languages don't fork the standard; they join it, one authored pack at a time.
The canonical set is the interchange form — semantics every agent resolves identically. It always works, in every country, pack or no pack. Translate it per language and portability dies; keep it symbolic and it travels like an airport code.
Each language gets a pack file — a full native vocabulary designed for its speakers' own guessability, with its own shorthand space. Authored once, thoughtfully, by someone who speaks the language — because a literal translation can say the opposite of what a token does. Launch packs: English · Spanish · German · French · Polish · Dutch — six packs natively covering 13 of the 17 countries in the reference footprint, with Swedish, Norwegian, and Finnish as the named fast-follow tier.
Type the bare prefix and the registry comes back in your language, from your pack — names you can guess, explained the way you'd explain them. Where a pack has a gap, the canonical token still works, and the pack says so honestly.
Origin
Directive Shorthand began with one token — a way to say "answer me in a voice note" without writing a paragraph about it — and a sharper insistence behind it: the commands must live in the toolkit, so every agent knows them always, instead of living in one agent's memory of one user. The research that followed verified the opening: fragments everywhere, a standard nowhere. So the design was finished, written as a specification with a conformance test, and published in full — spec and research alike — on the day it was designed.
Authored by Michael Nash · AgentForge · First published 2026-10-18
The specification, the design record, and this page carry that line deliberately. Priority is a feature of a standard: the dated, public, attributed record is how a specification's authorship stays a matter of fact.