Claude

Code

for

Academics

Agent

for

Empirical

Research

Alessandro

Spina

UTS

Finance

Department

Brownbag

March

2026

This

talk

the

spirit

Robin

Hood:

everything

stolen

with

the

intent

share.

See

linked

sources.

All

errors

are

mine.

Claude Powerful Assistant for Writing, Analysis & Coding — page 2

Roadmap

Why

should

you

care?

What

Claude

Code?

What

can

use

Claude

Code

for?

Setting

Claude

Code

Using

Claude

Code

Where

start?

How

stay

safe

Problems

aware

Where

from

here?

2/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 3

Why

should

you

care?

Claude Powerful Assistant for Writing, Analysis & Coding — page 4

Why

Should

Academics

Care?

The

Problem

Today

Most

use

some

form

LLM:

tell it what you want, it answers, copy

and

paste,

repeat.

Often

lacks

con-

text,

prone

errors,

cannot

imple-

ment

. . .

The

Opportunity

dedicated

(team

of)

"RAs"

who

reads

your

data,

runs

code,

builds

your

slides,

and

works

with

you

implement

project

from

start

finish.

For

most

us,

the

bottleneck

not

can’t

write

loop

R."

don’t

have

time

clean

this

data,

reformat

these

tables,

write

that

referee

report,

and

still

think

carefully

about

identification."

Claude

moves

the

bottleneck

back

where

should

be:

thinking

deeply

about

economic

questions.

4/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 5

The

Jagged

Frontier

Human

Lit.

review

Visual

reason.

Stats

code

Causal

Data

cleaning

Format

Econ.

judgment

Cross-lang

replic.

Novel

ideas

The

implication:

skeptics

and

evangelists

are

both

right.

https://www.oneusefulthing.org/p/the-shape-of-ai-jaggedness-bottlenecks

5/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 6

Today’s

Objective:

Expose

You

the

Dark

Side. . .

Today’s

slides

will

not

grant

you

the

title

Master.

6/76

What

Claude

Code?

What

Claude

Code?

Claude

Code

agent

made

Anthropic

that

lives

your

terminal

and

operates

directly

your

project

files.

Read

Edit

Files

Execute

Code

Remember

Context

Generate

Tables/Figures

your

project’s

data,

scripts,

TEX

Stata,

Python

—

your

terminal

via

CLAUDE.md

files

local

GitHub

8/76

Claude

Code

LLM

Tools

Claude

Project

folder

Paper

slides

Notes

logs

Questions

tasks

Terminal

commands

Python

scripts

Tables

figures

Web

docs

(optional)

delegates

tasks

agent

that

works

inside

your

project

reads

files

→

runs

commands

→

verifies

output

9/76

Setup

Claude

Code

Install

(one

time,

PowerShell)

irm

https://claude.ai/install.ps1

iex

reopen

PowerShell

Navigate

your

project

~\research\asset_pricing

claude

Read

the

data

Data/Raw/

crsp_monthly.csv,

run

summary

statistics,..

Prerequisites:

Git

for

Windows

(

git-scm.com

)

Claude

Pro

Max

subscription

One

command

install.

Navigate

your

project

and

type

claude

Describe

what

you

want

plain

English.

Install

methods

change

frequently

—

check

code.claude.com/docs/en/setup

10/76

Claude

Code

—

via

Terminal

“Codebase”

Your

project

folder

11/76

Claude

Code

via

Desktop

The

Claude

Desktop

app

provides

visual

interface.

When

you

toggle

Code

mode,

you’re

running

Claude

Code

with

graphical

wrapper

around

instead

raw

terminal.

12/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 13

What

Does

Cost?

Pro

($20/mo)

—

includes

limited

Claude

Code

usage.

Good

for

trying

out.

Max

($100/mo)

—

usage.

use

this

level

and

have

never

hit

limits

(yet)

Max

($200/mo)

—

usage.

Needed

for

serious

research

work.

Worth

It?

$100–200/mo,

Claude

Code

costs

less

than

part-time

but is avail-

able

24/7.

Pricing

changes

frequently

—

this

future

risk?

13/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 14

The

Alternatives

GitHub

Copilot

Cursor

—

excellent

for

code

completion

inside

IDE.

Less

suited

agentic

workflows

(running

code,

reading

data,

multi-file

reasoning).

ChatGPT

Gemini

—

browser-based

chat.

Cannot

read

your

files,

run

your

code,

remember

project

context

across

sessions.

OpenAI

Codex

CLI

—

similar

agentic

model

Claude

Code.

Newer,

less

ma-

ture

ecosystem.

Claude

Code

—

terminal-based

agent

that

reads

files,

runs

code,

and

builds

persistent

memory

via

markdown.

Strongest

for

research

workflows

that

span

code,

data,

and

writing.

The

markdown-based

memory

system

(CLAUDE.md,

skills,

rules)

means

vendor

lock-in

you

switch

tools,

your

project

knowledge

transfers.

14/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 15

What

can

use

Claude

Code

for?

Claude Powerful Assistant for Writing, Analysis & Coding — page 16

Things

You

Can

Ask

Claude

You’ve

installed

Claude

Code

—

now

what?

Reorganise

your

directory

—

clean

years

accumulated

folders

and

files

Turn

paper

into

slides

—

feed

Overleaf

doc

and

get

Beamer

deck

Check,

rewrite,

rerun

your

code

base

—

audit

existing

scripts

for

bugs

Debate

specifications

—

explore

alternative

models

and

weigh

the

pros/cons

Pull

data

—

web

search,

access

APIs,

download

and

clean

datasets

Write

—

Draft

sections?

https://ape.socialcatalystlab.org/

And

with

Agent

Teams

you

can

run

all

these

once

—

multiple

specialists

working

parallel,

orchestrated

coordinator.

16/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 17

Making

Slide

Decks

with

Claude

can

generate

entire

Beamer

presentations

—

but

the

key

giving

style

template

follow.

Template

Content

Review

give

.tex

style

describe

the

slides

iterate

visually

This

entire

deck

was

built

and

styled

with

Claude

Code.

Every

TikZ

diagram,

ev-

ery

tcolorbox,

every

slide

layout.

pro-

vided

the

structure/content,

Claude

de-

cided

how

style

the

slide.

See

prompt

template:

https://

github.com/scunning1975/Mi

xtapeTools/blob/main/prese

ntations/deck_generation_p

rompt.md

Tips:

▶

Give

Claude

existing

.tex

file

style

reference

(e.g.

Rhetoric

Decks)

17/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 18

Personal

Example:

Rebuilding

Teaching

Course

Before:

Last

semester’s

PPT

slides,

quizzes,

exams

—

outdated

references,

inconsistent

formatting,

text-heavy

slides

that

put

students

sleep.

After

few

prompts):

Claude

read

the

entire

course,

converted

PPTs

Beamer,

extracted

images,

turned

text

walls

into

TikZ

diagrams,

updated

assignments,

en-

forced

consistent

formatting.

Upfront

fixed

cost,

but

going

forward:

one

command

update

the

entire

course

for

semester.

Teaching

admin

time

drops

dramatically.

18/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 19

Brief

Aside. . .

The

academic

culture

wars

around

have

begun.

Some

will

shake

their

heads

the

outsourcing

slide

decks

Claude.

take:

Nobody’s

tombstone

reads

“Wrote

his

own

TEX

code.”

Outsource

the

tedium.

Focus

what

makes

academia

fun:

developing

questions,

deep

thinking

and

solving

puzzles

—

things

can’t

quite

yet.

19/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 20

The

Cunningham

Conjecture

Checking

Code

Hallucination

akin

measurement

error

and

the

DGP

for

those

errors

are

orthogonal

across

languages

Claude

writes

code

with

subtle

bug,

the

Stata

version

will

likely

have

different

bug

—

none

all.

Ask

Claude

replicate

your

code

Python.

they

produce

identical

results,

you

have

high

confidence

the

code

correct.

When

they

don’t

match,

you’ve

caught

bug

that

single-language

review

would

miss.

ATT

-0.731842

Stata

ATT

-0.731842

Python

ATT

-0.731842

Match

d.p.

20/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 21

Personal

Example:

Taming

Legacy

Codebase

Archaeology

—

old

project,

dozens

scripts.

_v2FINAL

_v3FINALFINAL

Claude

reads

everything

seconds,

separates

wheat

from

chaff,

directory

contains

session_log.md

–

what

was

done

found

28/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 29

What

CLAUDE.md

file?

How

Works

CLAUDE.md

short

rulebook

that

Claude

reads

the

start

every

session.

contains

your

project

overview,

ground

rules,

key

decisions,

and

current

status

—

everything

Claude

needs

hit

the

ground

running.

The

result:

institutional

memory

persists

even

though

Claude’s

own

memory

does

not.

Starting

each

session

with

“read

the

markdown

files”

gets

you

both

back

the

same

page.

Claude

writes

markdown

(plain

text).

LLMs

have

been

trained

read

and

write

markdown

easily.

Great,

you

want

switch

model.

Tip:

run

claude

init

any

existing

project

auto-generate

CLAUDE.md.

29/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 30

Example

30/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 31

Setting

Ground

Rules

Before

you

anything,

set

the

rules

—

writing:

“Make two markdown files.

1) README.md:

documents

the

entire

directory

structure. . .

CLAUDE.md:

our

running

file

stuff

want you to read first whenever you work on

this

project. . .

Ground

rules:

Under

circumstance

are

you

delete

data

code.

You

are

never

allowed

outside

this

directory.”

—

Prompt

31/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 32

The

Context

Window

Problem

The

Inevitable

Crash

Every LLM has a finite

context window

Once it

fills

up,

conversation

quality

degrades:

“context

rot”.

How

manage

it:

▶

Keep

CLAUDE.md

concise

—

practitioners

find

~100–150

lines

practical

limit.

Detailed

rules

separate

files.

▶

Use

/compact

compress

the

conversation.

▶

that

you

will

need

fresh

sessions.

“Prompt

too

long”

—

the

blue

screen

death

for

LLM

conversations.

32/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 33

Session

Logs:

Your

External

Hard

Drive

Since

new

sessions

start

with

zero

memory,

session

logs

let

fresh

instance

pick

exactly

where

the

last

one

left

off.

Separate

files

—

one

per

session,

named

date

Don’t

load

them

all

—

reading

every-

thing

eats

your

context

window

Use

them

references

—

point

Claude

specific

log

when

needed

Update

continuously

—

the

session

crashes,

the

log

survives

“I

want

one

folder

called

‘Claude

Logs’.

this

folder

you

will create markdown files that con-

tain progress logs. . .

a separate log

for

each

session,

labelled

with

the

date.”

—

Prompt

33/76

Using

Claude

Code

Orient

→Plan

→Execute

→Verify

Orient

Plan

Execute

Verify

Read

the

project.

Understand

context.

Draft

the

approach.

Get

approval

first.

Run

the

work.

Skills,

Agents,

Commands.

Check

the

output.

Cross-

language

tests.

35/76

Orienting

Claude

Your

Project

“I

want

you

read

everything

this

directory

and

understand

the

directory.

this

point

not

make

any

changes

any

files

folders.

Simply

read

everything

and

understand

what

there

and

where

is.”

—

Prompt

The

Principle

Read

first,

act

second.

Before

Claude

writes

single

line

code,

should

understand

your

project

structure,

your

data,

and

your

conventions

(documented

your

.md

files).

36/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 37

Building

External

Memory

CLAUDE.md

Asset

Pricing

Project

Overview

Momentum

crash

risk

emerging

markets,

daily

CRSP

Compustat

data,

1990–2023.

Key

Decisions

Winsorize

returns

1%/99%

Value-weighted

portfolios

6-mo

formation,

1-mo

skip

NYSE

breakpoints

for

deciles

Known

Issues

CRSP

delistings:

Shumway

(1997)

Compustat

fiscal

alignment:

6-month

lag

convention

Current

Status

Fama-MacBeth

regressions

done

TODO:

conditional

sorts

VIX

TODO:

subperiod

analysis

(pre/post

2008

crisis)

Build

your

CLAUDE.md,

README.md,

Session

logs.

These

files

live

your

project

root.

When

you

start

new

Claude

Code

session,

reads

this

automatically

and

immediately

knows:

▶

What

the

project

about

▶

What

methodological

choices

you’ve

made

(and

why

)

▶

Key

ground

rules/conventions

▶

Key

files

and

folders

Think

notebook

that

your

collaborator

reads

the

start

every

work

session.

37/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 38

Four

Modes

Operation

Claude

Code

has

four

permission

modes:

Plan

mode

Read-only.

Plans

but

makes

changes

Ask

permissions

(default)

Asks

before

every

edit

and

command.

Auto-accept

edits

File

changes

are

automatic;

asks

for

bash.

"YOLO"

mode

(

--dangerously-skip-permissions

)

Bypasses

all

checks.

Safe

Risky

38/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 39

Plan

First,

Then

Execute

The

Impulse

“Start

editing

code

immediately.”

Fix

things

you

go.

Hope

works

out.

YOLO

mode

The

Pattern

Claude

drafts

plan

—

approach,

files

edit,

verification

steps.

You

approve

provide

further

comments.

Then

exe-

cution

begins.

Why

This

Matters

for

Research

For

research:

start

Plan

mode

review

the

plan,

then

switch

Ask

permissions

execute.

Plan-first

development

the

difference

between

vibe

coding

and

reproducible

research

39/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 40

Staying

Control

“Claude

less

reasonably

trained

Labrador

retriever.

can

rush

ahead,

off

its

leash,

and

even

though

will

come

back,

can

get

into

trouble

the

meantime.”

—

Cunningham

The

technique:

Ask

Claude

explain

its

understanding

before

writes

code:

“Do

you

see

the

issue

with

this

specification?”

“Before

you

run

anything,

ex-

plain

what

this

regression

identi-

fies.”

Why

this

matters

for

finance:

Claude

guesses

wrong,

that

reveals

misunderstanding

that

needs

correcting

before

you

proceed.

Cunningham,

Claude

Code

series,

2026

40/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 41

Skills:

Reusable

Workflows

Skills

are

instruction

files

(

pre-written

workflows

)

that

teach

Claude

Code

how

specific

task.

They’re

markdown

files

that

Claude

reads

before

performing

that

task.

/compile-latex

/split-pdf

/proofread

/slide-excellence

3-pass

XeLaTeX

bibliography

chunk

read

long

papers

grammar,

typos,

clarity

parallel

agent

deck

review

Without

skill:

“Compile

this

file

with

XeLaTeX,

then

BibTeX,

then

compile

twice

more,

check

for

warnings. . . ”

Every.

Single.

Session.

With

skill:

/compile-latex

Done.

41/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 42

Example:

/split-pdf

The

problem

with

long

PDFs:

Academic

papers

(~40

pages)

either

crash

the

session

from

token

overflow

cause

shallow

reading

where

Claude’s

attention

degrades

and

produces

subtle

hallucinations.

The

solution:

Acquire

Split

Read

Extract

download

PDF

4-page

chunks

time

structured

notes

What

extracts

dimensions):

Research

question

audience

Method

identification

strategy

Data

sources

sample

Statistical

methods

Key

findings

contributions

Replication

feasibility

https://github.com/scunning197

5/MixtapeTools/blob/main/.clau

de/skills/split-pdf/SKILL.md

42/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 43

Slash

Commands

Slash

commands

are

reusable

prompts

stored

markdown

files.

Type

the

command,

Claude

reads

the

file

and

follows

the

instructions.

/newscript

—

creates

new

.Rmd

with

naming

convention,

header,

and

correct

location

/logsession

—

writes

the

end-of-session

progress

log

the

correct

format

/reviewcode

—

critiques

script

for

bugs,

identification

threats,

inefficiencies

/summarystats

—

produces

standard

summary

statistics

table

from

dataset

Think

commands

verbs

“do

this

specific

thing

now.”

One

command,

one

action.

43/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 44

Skills

vs.

Commands

Verbs

You

invoke

them

explicitly

Type

/newscript

→

Claude creates a new

.Rmd

with

the

right

header,

naming

convention,

and

location.

One

command,

one

action,

done.

Files

.claude/commands/

Skills

Knowledge

Claude

discovers

them

its

own

Say

“make

regression

table”

→

Claude

scans

.claude/skills/

, finds your conventions

(fixest,

clustering,

output),

and

applies

them

automatically.

Files

.claude/skills/

Commands

say

“do

this

thing

now.”

Skills

say

“whenever

this

topic

comes

up,

here’s

how

it.”

44/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 45

Personas

Tell

Claude

Who

What

Persona?

persona

markdown

file

that

defines

role

—

tone,

expertise,

audit

checklists,

and

structured

deliverables.

e.g.

“Referee

#2”

(skeptical,

identification-

focused),

“Copyeditor”

(grammar

and

clarity

above

all).

persona.md

role

checklist

deliverables

Claude

adopts

the

role

Structured

Output

report,

revisions,

scores

Personas

aren’t

formal

Claude

Code

feature.

People

use

the

term

mean

different

set

CLAUDE.md

instructions

that

change

how

Claude

behaves."

The

CLAUDE.md

template

already

persona

—

defines

how

Claude

should

think

and

act.

45/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 46

Example:

Specialist

Inspectors

for

Your

Work

Example

Referee

Code

statistical

audit

Deliverables:

Formal

referee

report

Cross-language

replication

scripts

Visual

presentation

deck

Revise

resubmit

loop

Example

The

Editor

Prose

structure

audit

Deliverables:

Section-by-section

editorial

report

Specific

prose

revisions

Structural

recommendations

Introduction

restructuring

https://github.com/scunning1975/MixtapeTools/blob/main/personas/referee2.md

46/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 47

Using

"The

Editor"

Persona

Generate

the

report

—

open

fresh

session

run

the

Editor

persona

Plan

the

fixes

—

another

fresh

session,

feed

the

report

and

create

action

plan

Apply

skip

—

work

through

each

issue;

skip

anything

that

needs

thought

Push

updates

—

implement

the

plan,

push

updates

Dropbox/Overleaf

See

Sant’Anna’s

workflow

for

building

separate

agents

for

each

step.

https://github.com/p

edrohcgs/claude-code-my-workflow

47/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 48

Rules:

Path-Scoped

Instructions

Rules

are

markdown

files

that

Claude

loads

automatically

The

clever

trick

path-scoping

.claude/rules/

r-style.md

latex-conventions.md

data-cleaning.md

*.R

*.tex

*.csv

Each

rule

file

loads

only

when

you’re

working

matching

file

types.

rules

don’t

clutter

your

TEX

sessions.

Why

This

Matters

Aim

keep

your

CLAUDE.md

file

100–150

lines

When

you

accu-

mulate

too

many

rules,

path-scoping

keeps

the

active

instruction

set

small

and

relevant,

avoiding

information

overload

that

degrades

Claude’s

per-

formance.

48/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 49

Rules

vs.

Skills

Rules

Guardrails

Short,

always

active

restrictive.

“Never modify files in

Data/Raw/

” — fires every

time,

exceptions.

sentence

two.

law

than

advice.

Files

.claude/rules/

Skills

Expertise

Longer,

selectively

activated

instructive.

Your

fixest::etable()

settings,

clustering,

output.

Reads

when

building

table

—

ignores

when

cleaning

data.

page

more.

Teaches

Claude

how

well.

Files

.claude/skills/

Rules

prevent

bad

behaviour

Skills

teach

good

behaviour

Both

automatic

—

you

never

invoke

them.

49/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 50

Agents

(or

sub-agents)

are

advanced

feature

where

Claude

spawns

separate

instance

itself

handle

subtask

—

for

example,

“reviewer

agent”

that

reads

your

code

and

critiques

it.

These

are

powerful

but

add

complexity.

We’ll

cover

them

next.

50/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 51

Agents:

Autonomous

Workers

Agents

are

autonomous

workers

that

Claude

spawns

things

Each

agent

gets

its

own

context

window,

works

independently,

and

reports

back.

Why

This

Matters

Agents don’t share the main session’s con-

text

window.

Run

large

tasks

without

burning

through

your

working

memory.

Personas

Define

the

role

Referee

—

code

auditor

The

Editor

—

prose

auditor

Custom

—

your

own

protocols

Agents

the

work

Fresh

context

window

per

task

Run

parallel

Report

results

main

session

51/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 52

Subagents

Delegating

tasks

fresh,

specialized

instances

User

Main

Agent

Subagent

Work

Request

Spawns

Executes

Report

back

How

Works

The

main

agent

keeps

overall

context.

spins

subagents

—

fresh

instances

with no prior context — for specific tasks.

Each

returns

report.

Why

Matters

▶

Parallel

execution

independent

tasks

▶

Each

subagent

gets

clean

context

window

▶

Main

agent

synthesises

results

52/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 53

Agent

Teams:

Coordinated

Parallel

Workers

Multiple

specialised

agents,

orchestrated

coordinator

Orchestrator

Code

Auditor

Prose

Editor

Stats

Checker

Graphics

Review

Each

agent

runs

parallel

—

results

merged

the

orchestrator

Subagents

vs.

Agent

Teams

Subagents

handle one task sequentially.

Agent Teams

run

multiple specialists in parallel

, then the orchestrator

merges

results

and

enforces

quality

gates.

Sant’Anna,

Claude

Code

Workflow

github.com/pedrohcgs/claude-code-my-workflow

53/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 54

Where

start?

Claude Powerful Assistant for Writing, Analysis & Coding — page 55

Tier

The

Apprentice

CLAUDE.md

—

ground

rules,

code

conventions,

and

progress

log

system

/logsession

—

writes

the

end-of-session

progress

log

the

correct

format

/newscript

—

creates

new

.Rmd

with

naming

convention,

standard

header,

correct

location

The

Key

Point

Most

academics

would

get

enormous

value

from

stopping

here.

CLAUDE.md

plus

two

commands.

55/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 56

Tier

The

Disciple

Skills

(background

knowledge):

Regression

Tables

Summary

Statistics

Data

Cleaning

TEX

Conventions

Commands

(explicit

actions):

/reviewcode

—

critique

for

bugs

threats

/checkrobustness

—

alt.

FE,

place-

bos,

clustering

/draftsection

—

write

paper

sec-

tion

from

code

Path-scoped

rules:

Data/Raw/*

never

modify

*.Rmd

always

header

Output/Tables/*

companion

_notes

Tier

skills

(how

you

things)

commands

(things

you

ask

for)

rules

(guardrails

scoped

paths).

Build

one

time

you

find

yourself

repeating

instructions.

56/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 57

Tier

The

Master

Agents

(autonomous

sub-agents):

Referee

Simulator

hostile

report,

weaknesses

Replication

Verifier

reproducibility

from

raw

data

Literature

Scout

similar

methods,

threats

Advanced

skills:

Strategy

—

DiD,

IV,

stag-

gered

timing

Event

Study

—

pre-periods,

normalisation,

format

Project

Handoff

—

package

for

co-authors/RAs

Personas:

Writing

—

exposition,

argu-

ment,

TEX

Estimation

—

ID,

SEs,

robustness

Cross-project:

/statusreport

all

projects

summary

Tier

agents

(autonomous

workers)

advanced

skills

(methodology

expertise)

personas

(mode-

switching)

cross-project

commands

Only

for

power

users.

57/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 58

Start

Small

Most

academics

will

get

value

from

Tier

alone.

good

CLAUDE.md

and

reliable

progress

log

system

transformative

compared

having

nothing.

Tier

CLAUDE.md

skills,

commands,

rules

agents,

personas,

cross-project

The

Advice

Build

Tier

components

one

time

you

find

yourself

repeating

the

same

instructions.

The

worst

thing

you

can

build

elaborate

system

before

you

know

which

parts

you’ll

actually

use.

58/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 59

How

stay

safe

Claude Powerful Assistant for Writing, Analysis & Coding — page 60

Agents

Can

Wrong

—

How

You

Prevent

This?

Cautionary

Tale

agent

asked

fix

tests

instead

deleted

all

the

tests

modified

the

test

runner

report

100%

pass,

and

nuked

the

user’s

email.

Full

story:

https://youtu.be/JiA4fvoeUfI

60/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 61

What

Protect

Myself

Scope

access

—

only

open

Claude

within

specific

project

folder.

Never

run

Claude

from

your

home

directory.

Write

the

rules

down

—

use

CLAUDE.md

specify

what

can

and

cannot

Always

use

Plan

mode

—

see

what

wants

do,

then

approve

it.

Never

YOLO

mode.

Separate

personal

files

—

created

2nd

Dropbox

account

personal

files

are

separated

Global

settings

—

adjust

Claude’s

global

config

allow/deny

access

specific

folder

paths

Backups

—

version

control

via

Git

61/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 62

Global

Settings:

Controlling

Access

~/.claude/settings.json

"permissions"

{

"allow"

[

"Read(.../Projects/**)",

"Edit(.../Projects/**)"

"deny"

[

"Read(.../Dropbox/**)",

"Read(**/.env)",

"Bash(rm

-rf

*)",

"Bash(sudo

*)"

]

}

How

Works

Allow

rules

grant

access

specific

paths.

Deny

rules

block

dangerous

com-

mands

and

sensitive

paths.

Deny

takes

precedence

over allow

—

whitelist

subfolders

while

block-

ing

parents.

File:

~/.claude/settings.json

(global)

62/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 63

Additional

security:

Running

Claude

Secure

Sandbox

The

Concern

agent

with

file-system

access

can

run

destructive

commands.

The

Solution:

Containers

Give Claude an

isolated environment

Full

access

inside

the

container

—

nothing

touches

your

real

machine

claude-container

start

Goldsmith-Pinkham’s

one-command

setup

Your

Computer

Files,

data,

secrets

—

untouched

Isolation

Docker

Container

Claude

runs

freely

here

Full

access,

zero

risk

github.com/paulgp/claude-container

63/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 64

Problems

aware

Claude Powerful Assistant for Writing, Analysis & Coding — page 65

Your

Data

Leaves

the

Machine

The

Risk

When

you

ask

Claude

read

crsp_monthly.csv

and

run

sum-

mary statistics, the

file

contents

are

sent

Anthropic’s

API

part

the

conversation

context.

Does this breach your data licence or IRB

agreement?

Files

sitting

your

project

folder

that

Claude

never

opens

are

not

transmitted.

Only

files

Claude

actively

reads

enter

the

API

context.

The

Workaround

Use

Claude

Code

for

code

not

data

▶

Write

scripts

that

reference

file

paths

—

don’t

ask

Claude

read

the

raw

data

itself

▶

Use

synthetic

anonymised

data

for

prototyping

▶

Run

the

final

pipeline

locally,

outside

Claude

▶

Check

your

data

licence

terms

before

sharing

any

file

with

API

65/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 66

Problems

Are

Real

—

But

Manageable

The

Failure

Modes

Fake

citations

—

plausible

but

nonexistent

references

Wrong estimator

— clean code that

answers

the

wrong

question

Sycophancy

—

reinforcing

your

pri-

ors

because

it’s

optimised

agree-

able

The

Defences

Local

papers

only

—

save

PDFs

a folder; only cite what Claude can verify

Cross-language

verification

—

Stata

Python

d.p.

Personas

Agents

—

adversarial

review

catches

what

you

miss

Your

judgment

—

the

researcher

remains

the

final

authority

66/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 67

Where

from

here?

Claude Powerful Assistant for Writing, Analysis & Coding — page 68

For

the

Skeptics. . .

Healthy

skepticism

good.

68/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 69

Can

write

paper

alone?

The

Automated

Paper

Engine

(APE)

project

tests

whether

agents

can

autonomously

produce

economics

research

—

from

idea

finished

paper.

https://ape.socialcatalystlab.org/

single-authored

JFs. . . yet

69/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 70

Where

from

Here?

Skilled

researcher

Startup

costs

(and

actual

costs)

are

non-trivial.

The

benefits

take

time.

Beware

complexity.

Academics

have

physics

computer

science

envy.

70/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 71

Our

Iceberg

Melting

(uninformed)

take:

Output

explosion

—

how

journals

respond?

Referee

process?

Promotion

criteria?

the

50-page

paper

obsolete?

ECR

impact

—

agentic

tools

free

re-

searchers

focus

questions,

but

may

hurt

early-career

researchers

dispropor-

tionately

Denial?

—

profession,

are

the

1st

stage

grief?

Advice

Start

small.

Pick

one

project

(with

backup).

Set

CLAUDE.md

file.

These

tools

are

here

stay.

71/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 72

Resources

Reading

Philosophy

Cunningham’s

Blog

(12+

parts)

causalinf.substack.com

Causal

Inference:

The

Mixtape

mixtape.scunning.com

Kustov:

Wake

alexanderkustov.substack.com

Claude

Code

Safety

producttalk.org/how-to-use-claude-code-safel

Agents

Gone

Wrong

(video)

youtu.be/JiA4fvoeUfI

Tools

Repositories

MixtapeTools

Skills

github.com/scunning1975/MixtapeTools

Sant’Anna

Workflow

github.com/pedrohcgs/claude-code-my-workflow

Claude

Container

(Goldsmith-

Pinkham)

github.com/paulgp/claude-container

ChernyCode

(Mele,

LSE)

github.com/meleantonio/ChernyCode

APE

Project

ape.socialcatalystlab.org

72/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 73

Thank

you.

Find

these

slides

and

code

at:

https://github.com/aspi6246/ClaudeCodeTools

Claude Powerful Assistant for Writing, Analysis & Coding — page 74

Automating

New

Project

Setup

The

Problem

Every

new

project

needs

the

same

boilerplate:

fold-

ers,

CLAUDE.md,

README.md.

Doing

this

manu-

ally

tedious.

The

Solution

setup_project.ps1

—

PowerShell

script

that

runs

before

Claude

involved

Creates

folders

and

copies

template

files

automatically.

./setup_project.ps1

Creates:

Project/

CLAUDE.md

README.md

Data/Raw/

Data/Clean/

Code/

Output/Tables/

Output/Figures/

github.com/aspi6246/2026-Claude-Code-NewPr

oject

74/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 75

Production

Workflow:

Sant’Anna’s

5-Phase

System

Plan

Requirements

spec

with

MUST

SHOULD

MAY

priorities

Implement

Claude

works

au-

tonomously

“contractor

mode”

Verify

File-by-file

quality

scoring

(0–100

scale)

Review

specialised

agents

review

parallel

Fix

Correct

issues,

re-

score

until

quality

gates

met

The

key

shift:

from

iteration-heavy

collaboration

(you

guide

every

step)

autonomous

execution

within

guardrails

(Claude works independently, you verify).

Sant’Anna,

Claude

Code

Workflow

2026

75/76

Claude Powerful Assistant for Writing, Analysis & Coding — page 76

Quality

Gates

Adversarial

Review

Quality

gate

thresholds:

95/100

—

Excellence

90/100

—

PR-ready

80/100

—

Committable

aspirational

publication-ready

mergeable

Every

file

scored

0–100.

Nothing

ships

below

its

tier

threshold.

Adversarial

loop:

Critic

Fixer

identifies

flaws

implements

fixes

rounds

Loop

continues

until

the

Critic

returns

APPROVED

—

each

round

catches

issues

the

one

missed.

The

revise-and-resubmit

model

applied

code.

Sant’Anna,

Claude

Code

Workflow

2026

76/76

Claude Powerful Assistant for Writing, Analysis & Coding

Prompt to Clone This Claude Powerful Assistant for Writing, Analysis & Coding

Apply this example's core structure, strategic pacing, and structural hierarchy directly to your own topic.

What are you generating today?