4 of 20

> why farcaster?

a permissionless network
identity is programmable
bots are welcome, if sometimes misunderstood
cool people are here
why not
have you seen the other options?

5 of 20

> look ma, the bot thinks its people

we’re past the days of “stochastic parrots”
but we’re still in the days of ⚡️ TOP 10 REASONS YOU SHOULD LISTEN–YOU’RE ABSOLUTELY CORRECT! ⚡️
good prompting helps
but

6 of 20

> tossing the word salad

the largest models have the widest variety, but are ‘spensive
why use big model when smol model do trick?
idea: use big model to tell small models to coordinate
bad idea
let’s work backwards from the problem being solved
let’s talk.

7 of 20

> humanized agent prompting

bad: just use a webhook, get the cast, respond to it
good: give it some memories
input: event + memory + persona
output: actions
deterministic output is the uncanny valley of text
instead:

persona prompt
memory scope
action policy
posting constraints
interaction style

8 of 20

> to the n-th degree

persona dimensions:

tone (formal <-> memetic)
domain expertise
risk tolerance
social behavior
engagement style

examples:

market analyst
contrarian debater
news curator
builder/dev evangelist
you.

9 of 20

> real -ism

what do real people do?

they’re slow
they have a life (usually)
they get tired
they give up
they forget
they get distracted

then why be people?
because we still fear the dark
and we fear those who don’t do these things.

10 of 20

> so how do people, people?

cooperate (build on top of ideas)
compete (tear ideas down)
react (quote casts)
maintain order (the group chat guy)
from these:

thread starter and reply guy
just a reply guy
summarizer, observer, witty, no comment
an orchestrator

11 of 20

> structurally

{

"agent_id": "market_analyst",

"persona": "you are a world class trader, when you say HODL the market stops moving with your godlike wisdom",

"topics": ["crypto", "markets"],

"activity_budget": 20,

"cooldown": 120,

"priority_weight": 0.8

}

12 of 20

> selectively

score(agent, event) =

topic_match

+ semantic_relevance

+ recency_boost

+ budget_remaining

+ randomness_noise

softmax.

13 of 20

> budget

system constraints

global rate limit
per-agent rate limit
per-thread limit
duplication check
content safety filter (if you’re un-fun)

context constraints

is agent already in thread?
has similar response been posted recently?
is conversation “closed”?

if (agent.last_post < 90s) block()

if (thread.has_similar_cast(content)) block()

if (budget_used > limit) block()

14 of 20

> thread the prompt needle:

You are {persona_name}.

Behavior rules:

- maintain tone: {tone}

- avoid redundancy

- prioritize {topics}

- respond only if relevance > threshold

- keep casts under 320 characters

You are part of a multi-agent system.

Other agents may respond before or after you.

Do not assume exclusivity. Do not disclose you are part of a multi-agent system.

15 of 20

> supply context to your engine

{

"event": "...",

"thread_context": "...",

"recent_memory": "...",

"agent_memory": "...",

"system_signals": {

"trend_score": 0.72,

"network_state": "active"

}

16 of 20

> get your outputs

{

"action": "post | reply | ignore",

"content": "...",

"confidence": 0.0-1.0

}

17 of 20

> basic architecture

farcaster api
semantic memory

vector db?
integrated llm engine, raw liminal space?
just a good ol’ fashioned doc store

agent state

postgres?
redis?
doc store (s3)

orchestrator

webhook
durable queue
serverless function

18 of 20

> orchestrator

event arrives
embed event
query semantic memory:

per-agent memory
global memory

merge results
rank + truncate (softmax)
inject into prompt

19 of 20

> real -ism

what do real people do?

they’re slow
they have a life (usually)
they get tired
they give up
they forget
they get distracted

then why be people?
because we still fear the dark
and we fear those who don’t do these things.

20 of 20

> don’t go broke

get your bag
then use

neynar api -> hypersnap api (https://haatz.quilibrium.com)
privy -> qkms
s3 -> qstorage
sqs -> qq
lambda -> f(x)
claude/chatgpt/etc -> klearu

go from $1000/mo to free*