Skip to content

kevcooper/aedgin

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 

Repository files navigation

Aedgin: A Token-Minimized Pidgin for Human–Agent Communication

Version 0.1 — Draft Specification

Origin & Purpose

Aedgin (a portmanteau of AI + pidgin) is a constructed contact language designed for communication between humans and AI agents (LLMs). Its goal is to minimize token consumption and API cost while remaining fully intelligible to both parties.

Like historical pidgins, Aedgin emerges from a practical need: two "groups" (humans and language models) that share English as a superstrate but have different cost structures for using it. Tokens cost money; brevity saves it. Aedgin is not a cipher or code — it is a simplified, rule-governed dialect that any English speaker or LLM can learn quickly.


Design Principles (Borrowed from Pidgin Linguistics)

Pidgin Trait Aedgin Adaptation
Simplified grammar Drop articles, copulas, most prepositions when inferable
No conjugation/declension Bare verb stems; no tense marking unless ambiguous
No grammatical gender/number Context handles plurality; "they" is universal
Isolating structure One word = one meaning unit; no inflection
Limited core vocabulary Prefer short, common words; avoid synonyms
No embedded clauses Flat sentence structure; use semicolons to chain
Pragmatic & context-heavy Rely on shared context window to resolve ambiguity

Core Grammar Rules

1. Drop Articles

English articles (a, an, the) are almost always inferable from context and are pure token waste.

English: "Can you write the report about the quarterly earnings?" Aedgin: "write report about quarterly earnings"

2. Drop Copulas (be/is/are/was)

The verb "to be" is usually recoverable.

English: "The status of the project is behind schedule." Aedgin: "project status behind schedule"

3. Drop Subject Pronouns When Obvious

In a 1:1 human–agent chat, "I" and "you" are almost always clear from context.

English: "I want you to summarize this document." Aedgin: "summarize this doc"

4. Bare Verb Stems (No Tense Unless Ambiguous)

Default interpretation is present/imperative. Mark tense only when needed with a prefix particle.

Tense Marker Example
Past did "did send email"
Future will "will deploy friday"
Present (default) "deploy now"
Ongoing now "now run tests"

5. Flat Clauses — No Embedding

Replace subordinate clauses with semicolons or sequencing.

English: "After you finish the analysis that I requested, please send it to the team so they can review it before the meeting." Aedgin: "finish analysis; send to team; they review before meeting"

6. Compounding Over Phrases

Use compounds and shorthand instead of prepositional phrases.

English: "a list of the items in the database" Aedgin: "db item list"

7. Use Symbols as Operators

Leverage symbols that tokenize efficiently and carry clear meaning.

Symbol Meaning Example
or -> then / leads to / output "parse csv -> json"
+ and / also / include "add header + footer"
~ approximately / about "~500 words"
!= not / exclude "list users != admin"
> / < more than / less than ">3 paragraphs"
? question / uncertain "deploy ready?"
@ at / directed to / regarding "email @boss re: budget"
# topic / tag / number "#3 priority"
& and (shorter than "and") "pros & cons"
: is / equals / defined as "tone: formal"
... continue / elaborate "good start..."

8. Standard Abbreviations (Core Lexicon)

Aedgin maintains a shared abbreviation set. Both parties should recognize these:

Abbrev Meaning
fn function
doc document
msg message
info information
req request / requirement
resp response
cfg configuration
db database
auth authentication
impl implementation
fmt format
src source
dest destination
prev previous
curr current
approx approximately
re: regarding
w/ with
w/o without
b/c because
eg for example
ie that is
max maximum
min minimum
est estimate
pls please
thx thanks
q question
ans answer
ctx context
orig original
diff difference
avg average
qty quantity
TL;DR summary

9. Output Control Shorthand

Instruct the agent's response format concisely:

Directive Meaning
brief short answer, 1–2 sentences
list N return N items as list
verbose detailed explanation
y/n yes or no only
code only just code, no explanation
table output as table
steps numbered steps
diff show only changes
1para one paragraph
ELI5 explain simply
no preamble skip introductions, get to answer

Sentence Patterns

Imperative (most common — commanding the agent)

[verb] [object] [modifiers]

"summarize doc; tone: casual; ~200 words" "fix bug in auth module; code only" "compare plan A vs B; table fmt"

Query

[topic] ? [constraints]

"deployment status?" "best db for timeseries? brief" "diff between REST & GraphQL? table"

Conditional

if [condition] -> [action]; else -> [action]

"if tests pass -> deploy staging; else -> notify team"

Sequence

[step1]; [step2]; [step3]

"read csv; filter rows > 1000; export json"


Agent Response in Aedgin

Agents CAN (and should, when the user writes in Aedgin) respond using the same conventions:

Standard response: "I've completed the analysis of your quarterly data. The key findings are as follows: revenue increased by 12% compared to the previous quarter, while operating costs remained stable."

Aedgin response: "analysis done. key findings: revenue +12% vs prev quarter; operating costs stable."

Agent-Side Rules

  1. Mirror user's compression level — if they write full English, respond normally
  2. Drop meta-commentary ("Sure!", "Great question!", "I'd be happy to...")
  3. Front-load the answer; context/caveats after
  4. Use same symbol set and abbreviations
  5. Skip restating the question

Token Savings Estimates

Message Type Standard English (tokens) Aedgin (tokens) Savings
Simple command ~15–25 ~5–10 ~50–60%
Complex instruction ~40–80 ~15–30 ~60–65%
Agent acknowledgment ~20–40 ~5–10 ~70–75%
Agent explanation ~100–200 ~40–80 ~50–60%

Estimated average savings: 40–60% on both input and output tokens.


Full Example Conversation

Standard English (≈95 tokens)

Human: "Hi Claude, I'd like you to help me write a Python function that takes a list of dictionaries and filters them based on a key-value pair that I'll specify. Please make it efficient and add type hints. Can you also include a brief docstring?"

Agent: "Sure! I'd be happy to help you with that. Here's a Python function that filters a list of dictionaries based on a specified key-value pair. I've included type hints and a docstring as requested:"

Aedgin (≈30 tokens)

Human: "write python fn: filter list[dict] by key-value pair; type hints + docstring; code only"

Agent: [outputs code directly, no preamble]


When NOT to Use Aedgin

  • Nuanced emotional communication — compression loses tone
  • Legal / medical precision — ambiguity is dangerous
  • Teaching / explaining to beginners — clarity > brevity
  • Creative writing prompts — you often want richness, not compression
  • When the human prefers natural language — always respect preference

Bootstrapping: Teaching an Agent Aedgin

Paste this into a system prompt or conversation opener:

respond in aedgin: compressed english dialect. rules: drop articles,
copulas, obvious pronouns. bare verb stems. use -> for sequencing,
+ for and, ~ for approx, : for is/equals. use standard abbrevs
(fn, doc, msg, cfg, db, fmt, etc). no preamble or meta-commentary.
front-load answers. mirror user compression level. brief unless
asked verbose.

Roadmap / Open Questions

  1. Negotiated compression — could agent + human negotiate compression level per-conversation?
  2. Domain lexicons — extend abbrev table for specific fields (medical, legal, devops)
  3. Aedgin-to-English transpiler — tool that expands Aedgin into full English for documentation
  4. Benchmark suite — standardized prompts in English vs Aedgin to measure token savings + accuracy retention
  5. Creolization risk — if agents train on Aedgin conversations, does it drift? How to stabilize?
  6. Multimodal Aedgin — conventions for image/audio prompts (eg "gen img: sunset, cyberpunk, 16:9")

Aedgin v0.1 — a living spec. Fork it, extend it, compress it further.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors