The labs are not building better chat apps. OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work" (OpenAI Charter). Anthropic's own Claude Sonnet 4.5 announcement says they have observed it "maintaining focus for more than 30 hours on complex, multi-step tasks" (Anthropic, 29 Sep 2025). And Demis Hassabis pitched DeepMind to investors on one line: "Step one: Solve intelligence. Step two: Use it to solve everything else." (Stanford GSB, 18 Jun 2026).
General is the word that matters, because general includes your job. The thing holding your output back is usually not the model's capability, it is that you re-explain your business from scratch in every new chat. So the fix is boring and it works: write your business down once, in one file, and give that file to the tool. Everything below is that setup.
This is the piece everyone skips, and it is the piece that does the work. A context file is a plain document describing your business the way you would brief a new hire on day one: what you sell, who buys it, how you sound, your real numbers, and what you are not allowed to say. The mechanism: Claude never starts from zero, so your prompts drop from a paragraph of setup to a single line of instruction, and the output stops being generic because it is no longer working from a generic business.
Here is the file. Everything in square brackets is yours to replace, and the headings are the part worth keeping as they are.
# Business context: [your business name]
## What we sell
Offer: [the main thing you sell, in one sentence]
Price points: [e.g. $497 one-off, or $2,000 a month retainer]
What is included: [deliverables, timeframe, what "done" looks like]
What we do not sell: [so Claude stops suggesting it]
## Who we sell to
Primary audience: [role, industry, company size, or life stage]
What they want: [the outcome, in their words not yours]
What they tried first: [and why it did not work]
Top three objections: [the exact phrasing you hear on calls]
## How we sound
Voice in three words: [e.g. warm, direct, specific]
We always: [e.g. lead with the outcome, use real numbers]
We never: [e.g. hype words, income promises, exclamation marks]
Two lines that sound exactly like us:
1. [paste a real line you wrote]
2. [paste another real line]
## Products and offers, in order
1. [the free thing] leads to [what it leads to]
2. [entry offer, price]
3. [main offer, price]
## Current numbers (update this monthly)
Audience or list size: [number, and where it comes from]
Best channel right now: [and the metric that proves it]
Conversion rate: [from what, to what]
Ad spend and cost per result: [if you run ads]
This month's one goal: [e.g. book 3 discovery calls a week]
## Constraints
Monthly budget ceiling: [number]
Tools we already pay for: [list them]
Claims we cannot make: [e.g. no income or medical claims]Six sections, and each one is doing a specific job:
If staring at a blank template is the thing that stops you, do not write it yourself. Open any Claude chat and paste this instead, and it will interview you and hand the finished file back.
I want you to build the context file that every future chat with you
will start from, and I would rather answer questions than write it myself.
Interview me about my business one question at a time, and do not move on
until my answer is specific enough to act on. Start with what I sell and at
what price, then who buys it and what they were doing before they found me,
then how I sound when I write, then my current numbers (audience size,
conversion rate, ad spend and cost per result if I run ads), and finally the
budget and the claims I am not allowed to make. If I give you a vague answer
like "small business owners" or "we sound professional", push back once and
ask me for the specific version with an example.
When you have everything, write the finished file as one markdown document
under the headings What we sell, Who we sell to, How we sound, Products and
offers, Current numbers, and Constraints, using my words rather than yours.
Finish with a short list of anything you still do not know about my business,
so I know exactly what to fill in later.Two ways to install it, and they do the same job. The free path keeps everything in the browser. The power path puts the file in a folder on your computer so Claude can read and write your real files. Start free, upgrade only when you feel the ceiling.
the free pathSame file, different home. Claude Code runs in your terminal and works on a folder, so it can read your actual exports, write files, and build the landing page as a real file instead of a block of text you copy out. Anthropic's own docs describe the file it reads as "instructions you write to give Claude persistent context", loaded at the start of every session (Claude Code memory docs).
curl -fsSL https://claude.ai/install.sh | bash. On Windows, open PowerShell and run irm https://claude.ai/install.ps1 | iex. Full options and troubleshooting: Anthropic's setup page.CLAUDE.md. That name is not decorative: it is the file Claude Code loads automatically at the start of every session in that folder.claude. It opens an interactive session and logs you in through the browser the first time./context and look under Memory files for your CLAUDE.md. If it is not listed, Claude cannot see it, and you are probably running from the wrong folder./init generates a starting CLAUDE.md from what is already in the folder, which is a fast way to begin if you have files there. Edit from there, since it can only describe what it can see.The mechanism: because Claude already holds your offer, your audience's objections and your voice rules, your brief shrinks to a single line and the output arrives as genuinely different angles you can test, rather than ten rewordings of one safe idea. This is also why the copy stops sounding like AI: it is working from your two real sample lines instead of an average of the internet.
Write me 10 ad variations for [the offer named in my context file], aimed at
[the audience segment], for a [cold traffic / retargeting] campaign on
[Meta / Google / TikTok].
Use the objections and the voice rules from my context file rather than a
generic ad formula, and make the variants genuinely different from each
other: at least three distinct angles (for example the problem angle, the
proof angle, and the contrarian angle), not ten rewordings of one idea. For
each variant give me the primary text, a headline, and one sentence on the
single reason you think this one might beat the others, so I know what I am
actually testing rather than just what I am posting.
Check every line against the constraints section of my context file, flag
anything that makes a claim I am not allowed to make, and rewrite it into
something I can say. Finish by telling me which one variant you would spend
the first $50 on, and what result would prove you wrong.The mechanism: you paste the numbers into the same conversation that wrote the copy, so Claude is grading its own work against benchmarks it already knows. That is the difference between "your CTR is 0.9%" and "the two winners both led with the objection, the loser led with the feature, here is what to write next". Same context, so the analysis connects to a decision.
Here are the results after 7 days of running the variants you wrote:
[paste your export, or type it out per ad: spend, impressions, clicks,
click-through rate, cost per result, conversions].
Compare them against each other and against the benchmark numbers in my
context file, and be honest about sample size: tell me plainly which
differences are real and which are still noise because the conversion counts
are too small to call. Then explain what the winners have in common at the
level of the angle and the promise rather than the exact wording, because
the angle is the part I can reuse and the wording is not.
Give me three next moves in priority order: what to kill, what to scale, and
what one new variant to write based on what the data actually showed. If
anything here contradicts my context file, an audience assumption, a price
point, or a conversion benchmark, say so and tell me exactly which line of
the file to update.The mechanism: the page inherits the same offer, objections and voice as the winning ad, so message match happens by default instead of by coordination. This is the step where the whole-workflow setup pays for itself, because the usual failure is a great ad pointing at a page written from a different brief three weeks earlier.
Build the landing page for the winning ad, as a single self-contained HTML
file I can open in a browser and hand to a developer.
Match the promise in the winning ad exactly, so nobody who clicks it feels
like they landed on a different offer, and handle the top three objections
from my context file in the order a real buyer meets them. Structure it as:
hero with the outcome, the proof, what is included, objection handling, a
simple pricing block, and one call to action repeated twice. Use my voice
rules, keep every claim inside my constraints section, and design it to be
read on a phone first.
Do not invent testimonials, logos, statistics or guarantees. Wherever social
proof would normally go, leave a clearly labelled placeholder, and list
underneath exactly what evidence I need to collect to fill each one.The mechanism: the report is written by the same thing that ran the campaign, so the numbers, the reasoning and the wording all come from one place. No re-explaining what happened, no reconciling a spreadsheet against your memory, and no writing in a stiff voice that is not yours because it is 6pm on a Friday.
Write the monthly report for [client or stakeholder name] covering
[date range], using everything in this conversation: the copy you wrote, the
results we analysed, and the page we built.
Write it in the voice from my context file, for a reader who is [their role,
and how technical they are], and lead with the outcome rather than the
activity. Structure it as: a one-line summary, what we did, what happened
with the real numbers, what we learned, and what we are doing next month.
Keep it under 400 words, and put anything that needs a decision from them in
a separate short list at the bottom so it does not get lost in the prose.
Never invent a figure. Every number has to come from data in this
conversation, and if something they will obviously ask about is missing,
list it under "needs confirmation" rather than estimating it. Finish with the
two sentences I could paste into an email as the covering note.Here is the part most people miss, and it is the actual argument for going deep on one general tool instead of collecting ten narrow ones. A narrow tool improves at its one job. A general tool with your context file improves at every job you have given it, on the same day, without you rebuilding anything.
Which is why the move is not "learn more AI tools". It is: write your business down once, put it somewhere a general tool can read it, and then give that tool more of your work.
/init and /context work. code.claude.com/docs