Working with AI · A field guide

From prompt to output: what decides how good AI work gets

How AI systems like Claude actually work, and the seven things that set the quality of what you get back.

The labThe modelWorkspaceConversationPromptSkills & agentsYour review
The whole journey

Where you run it decides what the AI can use

① Workspace
Chat history
Memory
Standing instructions
Your files
Skills
Connectors
Plugins
Sub­agents
Acts on files
Runs without you
Where its setup lives
ChatClaude app on web, desktop, phone
Your account settings
Claude Projectchats sharing files & instructions
Account + project instructions & files
Coworkdesktop agent on a folder
Account + installed plugins + the folder
Claude Code · localon your computer
Global ~/.claude + the folder’s .claude
Claude Code · cloudon Anthropic's servers
The repo’s .claude, not your laptop
Inside other appsChrome, Excel, Slack…
That app’s integration
built inlimited or opt-innot availableColumns 1–4: what it can see · columns 5–10 (③ tools): what it can do
2
Your promptsame in every row (next slide)
→
4
ModelFast · Workhorse · Frontier
→
5
The AI worksreads, plans, acts, checks
→
6
Outputtext, file, action
→
7
You review↺ back to ②
② The prompt

What goes into a prompt depends on the job you're asking for

Every prompt mixes: Context+ Task+ Goal+ Constraints+ Output format … but each job leans on different ingredients.
Job
Example
Must include
Add this line
Ask
“What do the new EU pay-transparency rules require of us?”
Your situation, and what you'll use the answer for
Cite sources. Say if you're unsure.
Create
“A launch post for CFOs about automated approvals”
Audience, goal, tone (ideally an example), length
Ask me questions before you start.
Transform
“Turn this 20-page report into a 1-page exec summary”
The source material, the new reader, what must survive
Keep every number unchanged.
Analyse
“What's driving churn in this survey export?”
The data, the question, the decision it informs
Show the evidence for each finding.
Think
“Pressure-test our Q1 positioning”
Constraints, options already considered, how to judge
Give 3 options with trade-offs.
Do
“Research 5 competitors and save a comparison file”
Definition of done, where to save, what's off-limits
Plan first; wait for my OK.
The big picture

Output quality is a stack, and you control most of it

01
The lab
Which company built the AI. It sets the ceiling on capability, safety and data handling.
IT / procurement
02
The model
Which model in the family, and how much “thinking” it's allowed.
You
03
Workspace context
Standing instructions, reference files, memory: what the AI knows before you ask.
You / your team
04
Conversation context
The thread so far. It can build understanding, or pile up noise.
You
05
The prompt
What you ask, how precisely, and in what order you ask it.
You
06
Skills, tools & agents
Reusable playbooks, connections to company data, and helpers that can take actions.
You / your team
07
Your review
Checking, critiquing and steering. The step that turns a draft into finished work.
You
Under the hood

A language model does one thing: predict the next piece of text

1 · Text is split into tokens

Roughly ¾ of a word each. Models read, think and bill in tokens.

Our Q3 campaign targets finance leaders at mid-size firms who are frustrated by

3 · Repeat, token by token

Every word of the answer is produced this way. The model has no memory between chats and no view of your company except what's included in the text it's given.

2 · It scores every possible next token

…frustrated by ___  (illustrative probabilities)

manual
34%
slow
21%
expense
15%
spread
10%
their
7%
…50k more
13%
Everything you put in front of the model shifts these odds. That's the whole reason context and prompts matter.
The context window

Each turn, the model reads one big bundle, and your prompt is a small part of it

System
Tools & skills list
Workspace
Conversation history
Files read & tool results
You
Free space

Illustrative split, mid-way through a typical Claude Code session. Modern windows hold a few hundred thousand tokens or more: several books' worth.

System prompt. The app's built-in rules and personality.
Tools & skills. Short descriptions of what the AI can do.
Workspace. Your standing instructions and saved memory.
History. Every message so far, yours and the AI's.
Material. Documents it opened, searches, data it pulled.
Your prompt. Often under 5% of what's read.
When it fills up, older parts get summarised (“compacted”) and detail is lost. A long chat isn't automatically a better chat.
Managing what goes into this window is the single biggest skill in working with AI.
Layer 1 · The lab

A handful of labs build the frontier models

Anthropic

US
Claude · Claude Code
Strong at writing, long documents and agentic work. Focus on safety.
Closed

OpenAI

US
GPT · ChatGPT · Codex
Biggest consumer footprint, broad product range, images & voice.
ClosedSome open

Google DeepMind

US
Gemini
Built into Workspace; strong multimodal (video, images) and huge context.
ClosedGemma open

xAI

US
Grok
Integrated with X; fast-moving release pace.
Closed

Meta

US
Muse · Llama
Powers Meta AI across its apps; history of open models.
ClosedOpen

DeepSeek

China
DeepSeek V4
Very low cost; open weights you can run on your own servers.
Open weights

Alibaba

China
Qwen
Wide family of sizes; popular base for company-built models.
Open weights

Moonshot AI

China
Kimi
Long-context and agentic focus.
Open weights

Z.ai

China
GLM
Strong open models for coding and agents.
Open weights

Mistral

France
Mistral · Le Chat
European option; EU data-residency story.
ClosedOpen

Capability

Top labs leapfrog each other every few months. Pick a lab for its platform, not just for this month's lead.

Data & compliance

Where data is processed and kept, and whether it's used for training. In a corporation, this usually decides the lab for you.

Closed vs. open weights

Closed: used through the lab's app or API. Open: you can download the model and run it in-house, but you also run and secure it yourselves.

Layer 1 · How they compare

For writing and professional work, the latest scores

Creative writing: which output people prefer

Arena (LMArena): blind side-by-side votes from real users. Best model per lab · Elo · axis starts at 1380.

Professional deliverables (incl. marketing)

GDPval-AA (Artificial Analysis): real tasks from 44 occupations (docs, decks, spreadsheets), judged head to head. Best model per lab · Elo · axis starts at 1450.

Close races are ties. On writing, Gemini 4 Argon and Claude Opus 5.5 are within the margin of error.

Snapshots, not verdicts. Rankings shift every few weeks. Use them as a shortlist.

Your own test beats any chart. Run 5 of your real tasks through 2–3 models and compare.

Data as of 2 Oct 2026: arena.ai/leaderboard/text/creative-writing · artificialanalysis.ai/evaluations/gdpval-aa
Layer 2 · The model

Three model tiers: pick one for the job, then set how hard it thinks

Fast
Haiku 4.5
Capability
Speed
Cost

Quick, cheap, good enough for simple, high-volume jobs.

Tagging feedback, short rewrites, summarising emails, quick lookups
Other labs: “Flash”, “mini”, “Lite” models
Workhorse
Sonnet 5.5
Capability
Speed
Cost

Near-frontier quality for most tasks at a fraction of the price. The right default.

Drafting copy, documents, slides, spreadsheets, research summaries
Other labs: their standard model
Frontier
Opus 5.5 · Fable 5.1
Capability
Speed
Cost

Deepest judgement and nuance. Opus tops the writing and pro-work rankings; Fable is a premium tier for very long autonomous work.

Strategy, positioning, long-form content, ambiguous problems
Other labs: “Pro”, “Max”, flagship models

The second dial: effort / thinking

The same model can answer straight away or reason first. More effort means better answers to hard problems, but it's slower and costs more. Turn it up for analysis and strategy, down for quick edits.

Rule of thumb

Start on the workhorse. Move up when the output lacks judgement or nuance, and down when you're doing the same simple thing hundreds of times.

Claude model line-up as of October 2026. Dots show relative position within the family.
Layer 5 · The prompt

Anatomy of a strong prompt

You're a senior B2B content strategist.1 We're launching automated expense approvals for finance leaders at companies with 500–5,000 staff. Their pain: month-end close drags on because approvals sit in inboxes.2 Write a LinkedIn post announcing it.3 The goal is demo bookings, not likes.4 Keep it under 150 words, use one concrete number, avoid words like “revolutionary”, and end with a question.5 Here's a past post that performed well, for tone: <example>…</example>6 Give me 3 versions with different hooks (data-led, story-led, contrarian) in a table, with a one-line rationale each.7 Before writing, ask me up to 3 questions if anything important is missing.8
1
RoleSets the expertise and point of view.
2
Context & audienceThe part people skip most, and the one that matters most.
3
TaskOne clear verb and deliverable.
4
Goal / the “why”Lets it make smart trade-offs you didn't spell out.
5
ConstraintsLength, must-haves, words to avoid.
6
ExampleShows the tone faster than any adjective.
7
Output formatNumber of options, structure, layout.
8
ProcessHow to work: ask first, plan first, or check its own draft.
② The prompt · Write it once

Write the context once, and every prompt gets shorter

Without a workspace

Monday · LinkedIn post~130 words
Tuesday · trial email~130 words
Wednesday · video script~130 words

Role, audience, voice, rules, examples and format retyped every time, or forgotten and left to guess.

Consistent: every task, and every teammate using the folder, starts from the same brief.
Rule of thumb: said it twice? Move it to the file, or just ask “add this to CLAUDE.md”.
In the Claude app: a Project does the same with instructions and knowledge files.
→

With CLAUDE.md and reference files

CLAUDE.md read every session
## Audience
Finance leaders, 500–5,000 staff. Pain: slow month-end close.
## Voice
Senior B2B strategist. Direct, no hype. Never “revolutionary”.
## Defaults
Social posts <150 words, one concrete number.
## Delivery
3 variants in a table, save to drafts/.
## How to work
Ask up to 3 questions if the brief has gaps.
brand/voice-guide.mdtone, with good & bad examples
examples/top-posts.mdpast winners to imitate
research/personas.pdfwho the buyers are
product/facts.mdfeatures, pricing, claims we can make

Read when relevant; CLAUDE.md points to them.

Your prompts now carry only what's new · ~15 words

MonLinkedIn post announcing automated approvals. Goal: demo bookings.
TueEmail to trial users who haven't set up approvals yet.
WedTurn Monday's post into a 30-second video script.
Layer 5 · Word choice

Specific words change the output; vague words leave it to chance

Swap vague words for checkable ones

make it engaging→open with a surprising number about month-end close
keep it short→under 120 words, 3 short paragraphs
professional tone→like a CFO writing to a peer: direct, no hype
for our audience→for finance directors who already use an ERP
don't be wordy→one idea per sentence, plain verbs
make it better→make the benefit clear in the first line

Say what to do, not just what not to do. Naming a banned phrase can plant it.

Phrases that change how it works

Ask me questions before you startSurfaces the context you forgot to give
Think this through before answeringMore reasoning on hard or multi-step problems
Give me 3 options that differ in…Breadth instead of one safe answer
Be critical. What's weak here?Counters its natural agreeableness
Only use the attached documentsKeeps it grounded in your sources
If you're unsure, say soFewer confident-sounding guesses
Cite where each claim came fromMakes fact-checking fast
Here's why this matters: …Better judgement on things you didn't specify
Layer 5 · Best practices

Prompting habits that consistently pay off

01

Brief it like a smart new hire

Brilliant, but on day one. What would they need to know to do this well?

02

Explain the why

“Readers skim on mobile” beats “use short paragraphs”, and it carries over to choices you didn't spell out.

03

Show, don't describe

One or two real examples of “good” say more than a paragraph of adjectives.

04

Material first, question last

With long documents, paste the material at the top and put your ask at the end. Label each part clearly.

05

Define the output

Length, structure, number of options, table or prose. Don't make it guess the shape.

06

One job per message

Big tasks go better as a sequence of steps than one giant request.

07

Let it interview you

“Ask me questions one at a time until you have enough to write this.” Great for briefs and strategy.

08

Separate draft from critic

After a draft: “Now review it as a sceptical CFO.” Then: “Fix the top 3 issues.”

09

Ask it to improve your prompt

“Here's my prompt. What's ambiguous or missing?” Prompt-writing is a task it's good at.

10

Edit, don't argue

If a reply goes wrong, edit your earlier message and re-run instead of piling corrections on top.

11

Save what works

A prompt that worked becomes a template, then a project instruction, then a skill.

12

Spot-check against reality

Numbers, names, quotes and links are where errors hide. Verify those first.

Layer 5 · Sequencing

Good work comes from a sequence of prompts, not a single one

1

Brief

Give the situation, audience and goal. Hold back the writing.

Here's the launch, the audience and our goal. Don't write yet. What do you need to know?
2

Explore

Get options before committing. It's cheap to discard ideas here.

Suggest 4 different angles, with the trade-offs of each.
3

Draft

Commit to one direction with a clear choice.

Go with angle B. Draft it, keeping the constraints from earlier.
4

Critique

Switch the AI from writer to tough reviewer.

Review this as a sceptical CFO. List the 5 weakest points.
5

Revise

Make targeted fixes, not a full rewrite.

Fix points 1, 2 and 4. Leave the rest unchanged.
6

Package

Turn the result into the formats you'll actually use.

Final version, plus 3 subject lines and a 2-line summary for my manager.

Start a fresh chat when…

you switch topics, it keeps repeating the same mistake, or the thread has got very long.

…and carry the essentials over

“Summarise what we've decided and the constraints, as a brief I can paste into a new chat.”

Not every task needs all six

A quick rewrite is just step 3. A strategy doc deserves all six, possibly several times over.

Layer 4 · Conversation context

Everything in the thread stays in play, including the wrong turns

Draft a post for our new expense feature.
“Revolutionise your finance team!” …
No, way too hypey. Also it's for CFOs.
“CFOs: revolutionise your month-end!” …
Still hypey. And drop the exclamation marks.
“Month-end close shouldn't wait on an inbox.” …

The crossed-out turns are still read on every later turn. Their hype keeps pulling the output back.

When context helps

Decisions, preferences and corrections build up. By turn 10 it knows your audience, tone and constraints without being told again.

When context hurts

Rejected drafts, abandoned ideas and unrelated topics compete for attention. Very long threads get summarised automatically, and detail is lost.

How to keep it clean

Edit & re-run the message that went wrong, instead of correcting after it.

One topic per chat. A new task gets a new thread.

Summarise & restart once a thread has done its job.

Layer 3 · Workspace context

The workspace briefs the AI before you've typed a word

📁 q4-launch/
 ├─ CLAUDE.mdbrand voice, audience, rules · read every session
 ├─ brand/
 │  └─ voice-guide.mdreference · read when relevant
 ├─ research/
 │  ├─ personas.pdf
 │  └─ q3-results.csvdata it can analyse
 ├─ drafts/where it writes output
 └─ .claude/
     ├─ skills/linkedin-post/a reusable playbook
     └─ agents/fact-checker.mda helper agent

In Claude Code, the folder you open is the workspace. In the Claude app, Projects work the same way: instructions plus knowledge files.

Loaded every session

  • CLAUDE.md files: yours, the team's and the folder's standing instructions
  • Memory: facts it saved from earlier sessions
  • Names & one-line descriptions of available skills, agents and connectors

Loaded only when needed

  • Files it opens or searches
  • The full instructions of a skill that matches the task
  • Data fetched through connectors
Write the brief once, in CLAUDE.md, and every prompt in this folder starts from it.
Layer 6 · Capabilities

Tools, skills, connectors, plugins and hooks: what each one adds

Tools

Built-in actions: read and write files, search, browse the web, run commands.

read · edit · search · web fetch

Skills

A folder of instructions, templates and examples for one kind of task. Picked up automatically when the task matches.

brand-voice · board-deck · pdf

Connectors

Via MCP, a standard connection protocol. Lets the AI read from and act in other apps.

Drive · Slack · Jira · CRM

Plugins

One install that bundles skills, agents, connectors and commands for a role or team.

marketing · sales · design

Hooks

Automatic checks at fixed moments. They run every time, so the AI can't forget them.

spell-check after each edit · block deletes

Why many skills don't clutter the context: they load in stages

Always: name + one-line description (a few dozen tokens)
→
When the task matches: full instructions
→
If needed: templates, examples, scripts
③ Tools · Beyond text

Claude Code can drive other tools, and chain them into workflows

Claude Codeplans & orchestratesImage generationcampaign visualsVoice & videovoiceover, clipsDesign toolsFigma, CanvaAnalyticsGA4, AmplitudeCRM & emailHubSpot, SalesforceAd platformsGoogle, Meta, LinkedInDocs & chatDrive, Notion, SlackThe webbrowse, research, SEO

Three ways in: connectors (MCP), APIs called from a short script, and command-line tools. Each one needs your company's approval and its own account or key.

Campaign asset factory

From one brief: headline and copy variants, on-brand images in 4 ad sizes, all saved and named in a folder and sent to Slack for review.

Brief.md→Claude→Image API→Slack

Monday performance digest

Pulls last week's traffic and ad spend, spots what moved and why, makes charts and posts a one-page summary. Runs on a schedule.

GA4+Ads→Claude→Charts→Notion

Localise a campaign

Adapts the master campaign for 6 markets (not just translation), then generates a native-sounding voiceover for each video cut.

Master copy→Claude→Voice API→Drive

Competitor watch

Visits competitor pricing and feature pages weekly, compares them with last week's snapshot and flags any change worth reacting to.

Web→Claude→Diff file→Email

Personalised outreach

Takes a CRM segment, writes tailored email variants per persona and loads them back as drafts. A human approves before sending.

HubSpot→Claude→HubSpot drafts

Design hand-off

Reads a Figma frame, writes copy that fits the actual space, checks it against the brand guide and fills it back into the layout.

Figma→Claude+Brand guide→Figma
Layer 6 · Agents

An agent is a model working in a loop, with tools and a goal

Gather read files, search Plan decide next step Act use a tool Check did it work? repeat until the goal is met You can approve, interrupt or redirect at any point
Chat

You ask, it answers once. You do the legwork between turns.

Agent

You set a goal. It does the legwork, step by step, then reports back.

A real run in Claude Code

Goal: “Compare our pricing page with our 3 main competitors and draft positioning ideas.”

GatherReads CLAUDE.md and brand/positioning.md
ActFetches 3 competitor pricing pages from the web
PlanOne page failed to load, so it tries the cached version
ActWrites drafts/pricing-comparison.md with a table
CheckRe-reads the table and fixes a mismatched price tier
DoneSummarises 3 positioning angles and flags 2 assumptions to verify
Layer 6 · Agents, continued

Subagents: helpers with their own fresh context

You one request Main agent plans, delegates, combines its context stays light and focused Researcher reads 40 web pages Data analyst crunches the Q3 spreadsheet Fact-checker verifies every claim Short summaries back to main agent each has its own context window

Clean context

The heavy reading happens in the helper's window. Only the conclusion comes back, so the main thread stays sharp.

Parallel work

Several helpers can run at once: research, analysis and checking happen side by side.

Specialists

Each subagent can have its own instructions, tools and even a different model, like a cheap fast one for bulk reading.

Layer 7 · Your review

You stay accountable: verify, steer and protect

Verify

Check where errors actually happen

  • Numbers, dates, names, quotes
  • Links and sources: open them
  • Claims about your own company or product
  • Anything it couldn't actually have seen

AI can be fluent and wrong at the same time. Polished text isn't proof.

Steer

Control what agents may do

  • Ask first: approve each edit or action
  • Plan mode: it researches and proposes, you approve before it acts
  • Auto-accept: only for low-risk, reversible work
  • Interrupt and redirect at any time
Protect

Handle company data properly

  • Use the company-approved tool and plan (enterprise plans don't train on your data)
  • No customer personal data or secrets in personal accounts
  • Connectors inherit your access, so be deliberate
  • Label AI-assisted work where policy requires it
The AI drafts; you decide. Your name goes on the output, so your judgement has to go into it.
Troubleshooting

When output disappoints, find the layer that failed

Symptom
Usual cause
Fix
Generic, “AI-sounding” text
Prompt No audience, goal or example
Add the context and the why; paste one example of good work
Ignores earlier instructions
Conversation Thread too long or cluttered
Summarise and start fresh; put standing rules in project instructions
Wrong about our company
Workspace It was never told
Add reference files and a CLAUDE.md or project instructions
Confident but invented facts
Review Not grounded or checked
“Use only the attached sources and cite them”; verify the specifics
Shallow analysis
Model Too small or too little effort
Switch to a bigger model or turn up thinking
Same fix needed every time
Skills Know-how not captured
Turn the working prompt into a skill or template
Agent went off-track
Agent Vague goal, no checkpoints
Use plan mode, define “done”, and ask it to check its own work
The compounding loop

Don't just fix the output. Fix whatever produced it.

1 · Prompt ask for the thing 2 · Output draft arrives 3 · Spot the gap what's wrong, and why? 4 · Fix the system file, skill, instruction not the draft every lap starts better
If the output……fix it here
Sounds off-brand→Add a brand-voice reference file with good and bad examples
Gets facts about us wrong→Add a reference file: products, pricing, positioning
Forgets a house rule→Write the rule into CLAUDE.md or the project instructions
Has the wrong structure→Put a template inside the skill
Needs the same 5 corrections→Turn the corrections into a skill
Lacks live data→Add a connector to the source system
Breaks a must-never rule→Add a hook: an automatic check that always runs
Fix the draft by hand

Helps once. Same problem next time.

Fix the system

Helps every time after, and for your whole team.

Take this with you

Eight habits for better AI output, starting Monday

Model

Match the model to the stakes

Workhorse by default; flagship for judgement calls.

Workspace

Write the brief once

Standing instructions in a project or CLAUDE.md.

Conversation

One topic, one thread

Edit instead of arguing; restart with a summary.

Prompt

Context and the why

Audience, goal, example and format beat clever wording.

Sequence

Brief, explore, draft, critique

Several focused steps beat one mega-prompt.

Skills

Capture what works

Good prompts become templates, then skills.

Agents

Delegate goals, not keystrokes

Give an agent a clear “done” and checkpoints.

Review

Trust, then verify

Check the specifics. You own the output.

The model is the engine. The context you give it is the steering.
References

Sources & further reading

Benchmarks (data as of 2 Oct 2026)

Prompting

Claude Code

Questions?

Speaker notes

1 / 1 ← → to navigate · F fullscreen