ISF 0427 — sharing
how AI agents
changed my
work and life
MashBean
two months · ~8 hrs/day · still surprised
02 — who I am
a non-coder, working
across policy & the public sector
NOW
General Manager
Matters
censorship-resistant
community platform
FELLOW
ISF 2026
Global Fellow
+ AI+ Expo next week
POLICY
Harvard Kennedy
School Fellow
→ Oxford MPP this Sept
BEFORE
Medical doctor
→ digital democracy
Taiwan public sector,
Ministry of Digital Affairs
03 — why this talk is for non-coders
you don't need to write code.
you need
judgment
,
data
, and
taste
.
JUDGMENT
know whether the answer is good or bad.
DATA
know where to find sources worth working with.
TASTE
know which questions only you can ask.
04 — chat vs agent
two completely different tools.
CHAT
a conversation
01
phone or laptop
02
quick answers, rewrites, brainstorms
03
lightweight, friendly, familiar
AGENT
a workflow
01
runs on your computer
02
reads files · writes tools · ships drafts
03
deep, repeatable, pipeline-able
05 — how I use them today
chat for vibes.
agents for everything else.
CHAT — keep using for
→ light conversations & bad-mood check-ins
→ replying to small emails
→ Deep Research (still very strong)
AGENT — moved everything else
→ research, analysis, document work
→ building small tools on my laptop
→ pipelines I run again and again
RULE OF THUMB
if I ask the same thing 3× → I turn it into a pipeline.
06 — takeaway one
01
let the model think.
on hard questions, speed isn't the point. the longer it reasons, the better the answer.
FAST
instant
BETTER
thinking
USE THIS
pro · heavy
07 — takeaway two
02
paying more is
actually cheap.
FEELS LIKE
$$ / month
an extra subscription line.
COMPARED TO
a hire
it's almost nothing.
the productivity gain isn't linear.
08 — why I started
it wasn't curiosity.
it was layoffs.
STEP 01
company
under pressure
STEP 02
we had to
downsize
STEP 03
one engineer
left
STEP 04
so I jumped
in myself
— Lunar New Year, downloaded an agent, didn't go back.
09 — aha moment one
it organized a folder.
BEFORE
1,000+
files · zero structure
→
AFTER · 1 SESSION
/01
research/
/02
policy-drafts/
/03
archive/
named, dated, sorted.
10 — aha moment two
the blog I'd
postponed for years.
old facebook posts
column drafts (2018)
scattered blog posts
.docx in 4 folders
notes app dump
→
SHIPPED · ONE WEEKEND
a real blog,
finally.
years of writing,
collected and published.
11 — case study
designer left. design system stayed.
SOURCE
design system
in figma
left by departing designer
→
AGENT WEEKEND
organized
→ pushed to git
tokens, components, docs
→
OUTPUT
new pages
in 2–3 hrs
i'm not a designer
or a frontend dev.
the labor needed to ship a service is dropping fast.
12 — research publishing pipeline
five roles. one relay.
01
researcher
gathers sources
02
writer
first draft
03
critic
stress-tests it
04
editor
revises
05
publisher
ships output
each role gets its own prompt, task, and responsibility.
13 — why pipelines matter
standardize the process.
quality stops drifting.
01
standardized
process
02
stable
quality
03
far fewer
hallucinations
04
improves
over time
14 — the reasoning chain
a structure for thinking,
not just writing.
core
question
01
→
sub-
arguments
02
→
evidence
03
→
risk
check
04
→
synthesis
05
15 — five reasoning types
each carries a different risk.
01
deduction
apply rules to cases
02
induction
find patterns in examples
03
analogy
compare similar cases
04
abduction
find the best explanation
05
causal reasoning
explain cause and effect — highest risk, most evidence required
16 — daily training practice
one hard question.
every single day.
①
ask
a real question worth a day.
②
run
push it through the pipeline.
③
publish
blog it or archive it.
④
tune
improve the prompts & roles.
— I'm training myself to ask better; the agent to answer better.
17 — example questions I've actually asked
questions I had no one
to discuss with — yet.
Q1
which media revenue-sharing models actually worked internationally?
Q2
how do we evaluate the influence of exile communities?
Q3
what policy is needed for AI-agent identity?
Q4
what are the real effects of age-verification laws, country by country?
18 — what humans should still do
delegate the dull.
keep the judgment.
DELEGATE TO AGENT
repetition
file wrangling
first drafts
data cleaning
KEEP FOR YOURSELF
asking the question
judging the answer
making the decision
taste & context
— inspiration comes from walking. I'd rather walk more.
19 — who should use agents
two kinds of people
get the most out of this.
A
people
with taste
you can ask questions only
you would think to ask.
B
people
with data
you know where the sources are
and how to put them together.
20 — try it tomorrow
ask better
questions.
use better
methods.
01
download an agent
02
choose a stronger model
03
ask one hard question
04
turn repeats into pipelines
thank you.
MashBean · ISF 0427