Blueprint GTM · Field Notes · June 2026
How I actually talk to Claude
810 real prompts, audited against what the experts say
I'm Jordan Crawford. I run go-to-market for software companies, and I am not an engineer — I can't read code, and I do most of my typing by voice, typos and all. I also run my entire business through Claude Code — that's Claude installed on your own computer: same brain as the chat website, but it can read your files, browse the web, and actually do work for you instead of just talking about it. For this guide, I had Claude dig through every single instruction I typed into it over the last two and a half weeks, find the patterns, and then check those patterns against the official advice from Anthropic (the company that makes Claude) and what serious practitioners published in the last few months. Every quote below is word-for-word real, verified against the original session logs by a script. Client names are blacked out (█████); my typos are left in on purpose.
The numbers
810instructions I typed
(May 25 – Jun 10)
~21actions Claude took
per instruction
16projects, from client work
to finding a retirement home
106prompts on the
busiest single day
41%of my prompts
contain a question
1 in 3prompts are under
60 characters
Read the second stat and the last two together and you see the real shape of this work: it's not someone writing long, perfect "prompts." It's a conversation — short pushes, lots of questions, and the machine doing twenty things for every one thing I say. That's the single biggest mindset shift for a beginner: you're not writing essays for a robot, you're managing a very fast new employee.
The ten habits the data found
Claude clustered all 810 prompts and found I do the same ten things over and over. Each one below comes with real examples, the count of how often the pattern showed up, and — further down — whether the official guidance agrees with me.
1 Ask for the plan. Approve it in five words. ~27 prompts
On anything big, I don't ask Claude to do the work. I ask it to plan the work, and I make it interview me first. Then approving is nearly free.
"All I want you to do is come up with a gigantic plan and only execute the plan."retirement search · May 26
"Can you come up with a plan? Just ask me a shit ton of questions. This is a pretty complex thing."client work · May 28
"yes execute this plan"retirement search · May 26
The plan is where I add value — I know the customers, the data, the goal. The execution is where Claude adds value. Splitting them means I never have to watch the typing.
2 Say who's reading it and where it lands ~37 prompts
I almost never specify format the way engineers do. I name the human and the surface — a text message, a Slack block, a stage — and let that decide the length, tone, and layout.
"Oh, no, I'm gonna send this in a text message, so just like this should only be, I don't know, four sentences."data project · May 27
"Can you just describe in plain English what this thing can do? In nicely formatted blocks that I can put in Slack, not just a huge, dense paragraph in ninth-grade English."client work · Jun 9
"I'm on stage to make me look amazing. Please give me the summary of what we've done in the greatest possible way"Blueprint HQ · Jun 3
And the most important audience to name is yourself. When I don't understand my own project's output, I say so:
"Can you explain what we did and what we found in plain English at an 8th-grade level? Everything sounds like gibberish to me."client work · May 29
3 Point at the raw reality — don't summarize it yourself ~31 prompts
When Claude needs context, I don't retype what I remember. I point it at the actual source — my email, the call transcript, the files on disk — and let it read the original. My memory is lossy; the source isn't.
"I just forget his name. You gotta go check my email."Blueprint HQ · Jun 10
"There's a bunch of examples in the █████ playbook that should be on disc, so go pull examples of the three different types of images and what you can tell from space."Blueprint HQ · Jun 1
4 Make Claude write its own next prompt ~55 prompts — my signature move
This is the habit I'd teach first. Sessions — individual conversations — get long and confused, even good ones. So before a session dies, I make it write the opening brief for its replacement: what we did, where everything lives, what to avoid, what to do next. Then I paste that into a fresh session. The new Claude starts smarter than I could ever brief it myself.
"You keep messing up, so I'm just going to open a new agent. Just write me a brand new prompt for a new session and copy it on my clipboard."Blueprint HQ · Jun 1
"I just want you to give me a prompt for a new session that will avoid all the pitfalls that we've already found"client work · Jun 6
"If I want to have another agent verify all of this, can you just literally tell that agent where everything is and what we've done?"Blueprint HQ · Jun 9
5 Demand receipts before you believe anything ~53 prompts
Claude will cheerfully tell you the work is done. I don't accept the claim — I demand the artifact: print the actual email, show the real number, explain the squiggle. If you can't see it, it didn't happen.
"print the actual damn eamils so I can read them"client work · May 29
"Whhy do you say "~$254,000" why the ~ ?"Blueprint HQ · Jun 1
"Are the company names right? Are the people's names right? Did we join the information right? Are the emails right based on patterns? Did we leave room for anything on the table?"Blueprint HQ · Jun 9
Fittingly, this habit saved this very document: while building it, an early automated estimate put my prompt count near twenty thousand. The verified count was 810 — the estimate had counted the computer's own actions as mine. Interrogate every number.
6 Quote the bad line, point an arrow at it, state the rule ~55 prompts
My editing vocabulary is paste-and-arrow. I copy the exact sentence I hate, type <--, and give both the fix and the general rule behind it. One bad example becomes a permanent ban.
"Here's the thing nobody tells you <-- can you remove phrases like this? Here's the thing nobody tells you, and just be more direct."newsletter · May 26
"the pattern that I want to ban in all of my language is like here's what I will do, I will not do this, I will not do this. That's awful."newsletter · Jun 4
"This is much better, but it needs to read a little bit more naturally. Right now, it's too staccato, and it needs to be more curious"Blueprint HQ · May 27
Notice the second half of that habit: words like staccato, curious, narrative. I steer writing with texture words and let Claude figure out the mechanics — I never rewrite the sentence myself.
7 Turn every fix into a standing rule ~43 prompts
When something goes wrong, the fix is the small win. The big win is making sure the system remembers. Claude Code has a memory file that carries rules across sessions — every correction should end up there.
"why isn't it a draft on the substack itself? do that, and make sure you always do that"newsletter · May 26
"can you make sure to never make this problem again and save it to the way we do this"retirement search · May 26
"Can you just switch that driver too so that I don't have to do that the next time?"Blueprint HQ · Jun 8
8 Run a team, not a chat ~68 prompts — the habit I use most
Claude Code can spin up extra copies of itself — "subagents" — each with its own workspace and its own job. I direct them in plain English, like staffing a project: critics, researchers, fact-checkers, a rival to audit the work.
"Can you have three different narrative agents critique this?"Blueprint HQ · May 31
"send ten different agents with your PDF scale, Opus agents, to read the PDFs and individual pages to spot check and QA this"Blueprint HQ · Jun 8
"All right, I'm gonna have Codex check your work. Can you write a prompt for Codex, and then make sure it has the full █████ document and put it in my clipboard?"Blueprint HQ · May 27
That last one is its own sub-habit: I regularly make a competing AI cross-examine Claude's work, and vice versa. Two models that disagree are more useful than one that's confident. More on subagents below — this is the feature most beginners don't know exists.
9 Name the tool you want used ~36 prompts
Over time I've saved reusable "skills" — packaged instructions for publishing a page, reading a PDF, designing a layout. I invoke them by name like kitchen appliances. You accumulate these; they compound.
"Can't you just use your publish skill and then replace the existing page?"Blueprint HQ · Jun 1
"Yea use your skill creator skill so I can use it for other clients"client work · Jun 2
"No, Claude Code has a skill to do this. Use Claude Code's PDF Reading skill. That's what you need to do."Blueprint HQ · May 29
Notice that last one: when output disappoints, I often correct the tool choice, not the output. Wrong tool, wrong result — no amount of rewriting fixes it.
10 Fence off what it must NOT touch ~12 prompts
Modern Claude is eager. The most valuable sentence in many of my prompts is the negative one — don't get creative, don't re-diagnose, change nothing else. Saying what's out of bounds is half of saying what you want.
"My goal is to reverse engineer that exact message. Do not get creative on the message."Blueprint HQ · Jun 3
"Read that file first. Do NOT re-diagnose — it's confirmed from production logs"Blueprint HQ · Jun 4
"We shouldn't do the headless browser per state scraper. Don't do that."Blueprint HQ · Jun 9
What Anthropic and the experts say — and where we disagree
Claude researched the current official guidance (Anthropic's docs and best-practices pages, updated for the newest models) and recent writing from heavy practitioners. Here's the honest scorecard.
✓ The experts agree — plan first
Anthropic's official workflow for anything non-trivial is explore → plan → implement → verify, with a built-in "plan mode" that makes Claude propose before it acts. The docs even recommend my exact interview trick: for big features, have Claude question you first, write a spec, then execute it. One official caveat I'd echo: skip the planning ceremony when the change is small enough to describe in one sentence.
✓ The experts agree — fresh sessions beat long ones
My "write your own next prompt" habit turns out to be the textbook answer to the most important constraint in this whole field: Claude's working memory (its "context window") fills up, and quality drops as it fills. The official docs say to start fresh after a couple of failed corrections; one practitioner puts it better than I ever did: "AI context is like milk; it's best served fresh and condensed!" My handoff-brief habit is just a way of pouring a fresh glass without losing what the last glass knew.
✓ The experts agree — demand receipts
Anthropic's #1 tip for Claude Code is to give Claude a check it can run itself — a test, a build, a visible artifact — and require evidence of success, not claims of success. Their prompt-engineering overview opens by telling you to define success criteria and test against them — evidence, not vibes. My "print the actual damn eamils" is the non-engineer's version of the same doctrine.
✓ The experts agree — teams of agents, with two cautions
Official docs recommend subagents for exactly what I use them for: heavy research that would clog the main session, and fresh-eyes review of finished work. Two expert cautions worth keeping: Simon Willison notes you shouldn't fragment every task into specialists by default — one agent with room to think often beats five with amnesia — and Anthropic itself warns that adversarial reviewers always find something, so chasing every critique over-engineers the work. Use critics, then use judgment.
◐ Half agree — standing rules
Turning corrections into permanent memory is officially blessed — that's what the CLAUDE.md memory file is for. But the docs add a warning my data says I need: keep that file ruthlessly short, because a bloated rule file makes Claude start ignoring rules. If you adopt my "make sure that never happens again" habit, also adopt a monthly pruning of the rules it creates.
◐ Half agree — typos genuinely don't matter (but vagueness does)
The official "golden rule" is that if a colleague with no context would be confused by your prompt, Claude will be too. My corpus is a stress test of that rule: I voice-type everything, so Claude received "eamils," "comit and push thpse," even "pull from Maine" when I meant the main branch (a technical name it decoded purely from context) — and it sailed through nearly all of it. Spelling is not clarity. What did burn me, repeatedly, was vague pointing — "fix the grease thing that was noted here" works mid-session and fails completely in the next one. Type sloppy, point precisely.
✗ The experts say I'm wrong — shouting
When Claude misses, I correct in caps, sometimes with profanity ("Sorry NOT all the links JUST LITERLLY ALL THE INDIVIDUAL PAGES"). The current official guidance is blunt about this: modern Claude follows instructions literally, aggressive emphasis degrades rather than improves results, and you should say what you want, not what you're mad about. Looking at my own data, the experts are right: every time shouting "worked," it worked because the message also contained the quoted line and the corrected rule (habit #6). The volume was theater. The arrow was the fix.
✗ The experts say I'm wrong — bundling
I routinely stack four or five unrelated asks in one prompt. Every guide, official and independent, says scope one task at a time — and my own sessions show why: bundled items get silently dropped, and I only catch them because I babysit (see below). Beginners should not copy this. One ask, one prompt, until you've built the checking habit.
⚠ Do as I say, not as I do
The audit also caught me pasting live passwords and access keys directly into the chat — more than once, sometimes in the same breath as telling Claude "make sure NOT to commmit my keys." Every security guide, including my own written rules, says the same thing: keys belong in a locked settings file (called .env) that Claude can use without the key ever appearing in the conversation itself. This is the one habit in my corpus with no redeeming nuance. Don't inherit it.
Loops, in plain English
Here's the most embarrassing pattern in my data: roughly 60 of my 810 prompts were me poking a long-running job — "Still going?", "how are we doing gents?", "okay where are we? No i CAN wait I just like to know" (the data says I could not, in fact, wait). I checked: in two and a half weeks, I never once used the feature that exists to do this for me.
It's called /loop, and it's exactly what it sounds like: you give Claude an instruction and a rhythm, and it re-runs that instruction on schedule inside your session. "/loop 5m check on that big research job and tell me when it's done" means Claude checks every five minutes so you don't. Leave off the interval and Claude paces itself — checking quickly when things are moving, backing off when they're not, and stopping on its own when the job's done. You can stop a loop by pressing Escape while it waits for its next run, and loops automatically expire after seven days, so a forgotten one can't run forever.
My take for beginners: use loops to watch, not to act, for your first month. "Watch this job and summarize when it finishes" is a perfect loop. "Re-run the campaign every hour" is not — an unattended instruction that acts on the world repeats its mistakes on schedule. Watching loops replace my worst habit with zero risk; acting loops you graduate into once you trust your own instructions. There's also a bigger sibling called /schedule that runs jobs in the cloud on a calendar — every morning, every Friday — even when your computer is closed. I haven't used it yet either, which tells you something honest: you can get enormous value from these tools while still ignoring half their features.
Subagents — the interns with their own computers
This is the feature I lean on hardest (68 prompts) and the one I most want beginners to understand, because it's the part almost everyone misses: subagents can use tools.
The mental model: you're not talking to one assistant, you're talking to a manager who can hire. Each subagent it hires gets its own clean working memory and — this is the new part — its own set of hands: it can search the web, read files, run scripts, call the same paid data services you've connected. And you, or Claude, decide which hands each one gets. A research agent might get web search and nothing else; a writing critic gets read-only eyes and no ability to touch your files. The official docs call these tool allowlists, and they work like job descriptions: the intern who reads contracts doesn't get the company credit card.
Why this matters for sales and GTM work specifically: the expensive part of go-to-market is parallel grunt work — check a hundred companies, read 200 PDFs, verify every row of a list. That's exactly what a team of tool-using subagents does in minutes. From my own corpus: "send ten different agents... to read the PDFs and individual pages to spot check and QA this." Ten readers, simultaneously, each with the PDF-reading tool, each reporting back.
There's an even bigger version called workflows — Claude writes a small program that directs up to 16 agents at a time, hundreds over a run. You don't need to understand it to use it: describe how you'd split the work, say "use a workflow" (or the magic word "ultracode"), and Claude builds the whole assembly line itself.
My take: the entire skill, for a non-coder, is learning to think in roles instead of tasks. Don't ask "is this list right?" Ask "deploy three agents: one checks the emails against the company websites, one checks the names against LinkedIn, one tries to disprove the whole list." You already know how to do this — it's hiring. The expert caution stands, though: critics always find something. The agents bring evidence; you still make the call.
What you can actually build (no code knowledge required)
Everything below shipped from my Claude Code sessions — the same sessions these quotes came from. I didn't write a line of code for any of it.
Over 1,200 customer dossiersOne intelligence page per customer for a client — pulled from call transcripts, support tickets, and CRM history — browsable by the whole sales team.
An 80,000-row market mapEvery potential customer in a healthcare software vertical, enriched with contacts, reviews, and the software they currently run.
A Slack robot that reads sales callsClassifies every new deal by buyer type, posts the verdict to Slack, tags the rep, and writes the result into Salesforce. Runs without me.
~1,200 customers ranked by influenceA referral-program hit list: a client's customer base scored by how connected each owner is in their industry.
A retirement-home search enginePersonal project: agents scored places to live across 195 countries on climate, water, health care, and actual property listings.
Published web reportsInteractive pages like this one — written, designed, and published to a live URL by Claude, including the page you're reading right now.
The shopping list
Claude Code gets dramatically more useful when it can reach real data tools. Each of these is one account, one key dropped into a locked settings file, and Claude handles the rest. Prices checked against the vendors' own pricing pages on June 10, 2026 (sources in the footer) — treat them as approximate and re-check before you buy.
Start here (~$50–150/month total)
- Claude Code — claude.com — Pro $20/mo; Max from $100/mo
The workbench itself. Pro is plenty for a first month; heavy daily use (agent teams, long jobs) eventually wants Max.
- Exa — exa.ai — $7 per 1,000 searches; 1,000 free/mo
Semantic web search built for AI agents. Finds the weird PDFs, Reddit threads, and agency pages Google buries. Point Claude at it instead of its native search.
- Parallel — parallel.ai — $4–9 per 1,000 searches; 20,000 free
Exa's rival, and they disagree usefully — in our research for this very page, Exa found the analyst blogs while Parallel surfaced primary documents and community threads Exa missed entirely. Run both; it's pennies.
- Serper — serper.dev — 2,500 free queries, then ~$1 per 1,000
Plain Google results as a service, dirt cheap. The workhorse for "just Google this 5,000 times" jobs.
- MillionVerifier — millionverifier.com — $39 per 10,000 emails
Checks whether an email address actually exists before you send to it. Non-negotiable before any outbound list goes live.
When you scale
- Apify — apify.com — Starter $29/mo; Scale $199/mo
A marketplace of pre-built scrapers — small programs that copy data off websites: Google Maps listings, LinkedIn pages, review sites. My plan is the $199 Scale tier; start at $29.
- Firecrawl — firecrawl.dev — from $16/mo (5,000 pages)
Turns whole websites into clean text Claude can read. The bulk-reading tool when a list of 2,000 company sites needs reading.
- OpenWebNinja — openwebninja.com — free tier; Pro $25/mo per data feed
A bundle of niche data feeds — local business lookups, contact pages, social links. Cheap first stop before the pricier enrichers.
- Clay — clay.com — free tier; paid from ~$167/mo
The enrichment spreadsheet your team may already pay for. Claude can fill and read it; the credits model means costs need watching.
- DataForSEO — dataforseo.com — pay-as-you-go, $50 minimum
Search-ranking and keyword data for when your GTM motion includes content and SEO.
- Gamma — gamma.app — ~$15–20/user/mo (API requires the Pro plan)
Generates real slide decks from a prompt. Claude writes the narrative, Gamma makes it presentable.
One money note in plain English. If you're on the $20 or $100 plan, your Claude bill is flat — the meter only starts if you later connect tools that call Claude directly (that's "the API"), which bills by usage. Claude comes in sizes: Opus is the big expensive one, Sonnet the everyday one (about $3 in / $15 out per million words-worth of work), Haiku the cheap fast one ($1/$5). Big agent teams on the metered side can genuinely surprise you. Two boring fixes that work: ask Claude "estimate the cost before you run this" every time a job involves the word "every," and pilot before you commit — from my corpus: "can you run 500 domains and see how it performs." Test on 500, then run the 80,000. It's habit #5 — interrogate the number — applied to your own bill.
Your first week
- Day 1 — Install and introduce yourself. Go to claude.com, find Claude Code, and follow the install steps — they'll have you open the Terminal app and paste one line. That terminal window is just a chat box with no buttons; type into it like you'd type into ChatGPT. Then type this: "Save this to your memory file so you remember it every session: I'm [your name], I sell [your thing], and I am not an engineer and don't pretend to be one." Claude reads that file at the start of every conversation and calibrates everything to it.
- Day 2 — One real task, with a plan. Pick something you were going to do anyway — clean a list, summarize five calls, draft a one-pager. Ask for a plan and tell it to ask you questions first (habit #1). Approve. Watch it work.
- Day 3 — Demand receipts. Whatever it produced yesterday, interrogate it: show me the rows, print the actual text, where did this number come from? (habit #5). You'll catch something. Then make the fix permanent: "save a rule so this never happens again" (habit #7).
- Day 4 — The handoff trick. End the day with my signature move: "Write a complete prompt for a fresh session to continue this — everything it needs to know, every mistake to avoid — and put it on my clipboard." Tomorrow, start a new session and paste it (habit #4).
- Day 5 — Hire your first team. Take a list of 20 prospects and say: "Deploy three agents — one researches each company's website, one finds recent news, one plays skeptic and flags which of these are bad fits — then combine the findings into one table" (habit #8).
- Weekend — Set a watcher. Start one long job and put a loop on it: "/loop 10m — check progress and give me a one-line status." Close the laptop lid mentally. That's 60 of my prompts you'll never have to type.
The one-sentence version
Treat Claude like the most capable new hire you've ever had, on their first day: brief them like a person (audience, goal, boundaries), make them plan before they act, never accept "done" without seeing the work — and when in doubt, have them write the brief for their own replacement.