CRAFT for Claude Cowork · Build in the Open
Our Simulated Customer Reviewed Our Comparison Page. Then We Rebuilt It.
On July 5 we shipped a comparison page we were proud of. The same day, we asked Leo — CRAFT’s simulated customer — whether it earned his Tuesday. He said not yet, and told us exactly why. This is the full record: the ask, the critique, the rebuild, and his re-review.
What this post is
CRAFT’s pitch has always been that work should run through written methods, review gates, and records you can check. It would be strange if our marketing were exempt. So when we published Claude Cowork, With and Without CRAFT — an honest, row-by-row comparison of what Cowork does alone and what the CRAFT layer adds — the page went through the same treatment everything else gets. Including a pass from the reviewer whose whole job is to not be impressed.
What follows are the actual messages from that working session, quoted verbatim from the session record. Nothing here is reconstructed from memory — the record is the point.
Who Leo is
Leo (simulated customer voice) is an AI persona — a fictional, composite Claude power user we built so that customer-side objections get raised before things ship, not after. His persona card describes someone time-constrained and tool-burned: he’s genuinely interested in solving session amnesia and prompt drift, but he’s seen too many cute AI tools, and he won’t adopt anything that doesn’t pay him back within two weeks. His tagline: “Show me why this is worth my Tuesday.”
Two things Leo is not. He’s not a real person — everything he “says” is a simulation, and his card explicitly forbids using him to manufacture testimonials or social proof. This post is the opposite of a testimonial: it’s the record of our own reviewer telling us our page wasn’t good enough. And he’s not authoritative — Leo reacts, the founder decides.
Leo was built with CRAFT’s own persona machinery: the CRAFT Persona Manager recipe (RCP-001-001-018), which manages persona activation and consistency, and the MPCS Orchestrator (RCP-CWK-031), which routes work between personas and runs multi-perspective reviews — both operating under CRAFT’s persona design specification (CD-030), which defines everything from his Big Five personality scores to his ethical boundaries. He was ratified as an official CRAFT persona earlier this year. The fuller story of his construction is in How We Built Leo — CRAFT’s Simulated Customer, his public card lives at Leo — Simulated CRAFT Customer, and if you want to work with personas in your own projects, start with Make the Most of CRAFT Personas.
The ask
The comparison page had just gone live. Here is the founder’s message to the session, exactly as typed:
From the session record · July 5, 2026 · founder’s message
Leo’s answer
The session’s response, exactly as it rendered — including the orchestration notes around Leo’s words, because that’s what working with personas actually looks like:
From the session record · Leo (simulated customer voice)
#P062a-Cat->H::Note: Routing to Leo, page content fresh from this session’s live fetch:
Leo — I read the live page the way I’d actually read it: 11pm, tired, third framework page this month. Four questions, four straight answers.
Does it clearly communicate what CRAFT adds? The message, yes — the “runs inside it, not a rival” frame is the right opening move, and each row is concrete instead of adjective soup. The design, only partly. Fifteen rows at identical visual weight is a wall, and my eyes glazed around row eight. The record itself knows there are about six rows nobody else in the world has and several that are table stakes — but the page renders them indistinguishably. Equal weight tells me you don’t know which of your own features matter.
Worth the extra effort? Here’s my problem: the page never tells me what the effort is. “Worth it” is a fraction — payoff over cost — and you’ve only shown the numerator. I’m left assuming the cost is high, because that’s what a 15-row table implies. The honest sections (“Cowork alone is enough,” “no scores, no stars”) bought real trust with me. Spend that trust by stating the price plainly.
Indispensable? No — and the gap is specific: the table describes capabilities when indispensability comes from moments. Your other post has the best sentence anyone’s written for this product — “the exact state a project was in when you closed the lid on Tuesday.” That’s a moment I’ve lived. “Session-to-session continuity” is a phrase from a spec sheet. The table tells; nothing on the page shows. Which is absurd, honestly, in the best way: your whole pitch is that everything is a readable file — and you didn’t show me a single file.
What I’d change as owner, honest-first: ① Show one real artifact. An actual handoff-file excerpt in a code box, ten lines, real. That’s not marketing — it’s evidence, and it’s the single highest-leverage change on the page. ② Add a “what you’ll notice in week one” strip above the full table — the three or four rows a newcomer actually feels early — full table stays below for people like me who audit. ③ State the cost: download, unzip, one sentence to start; a real learning curve that tutorial mode walks with you. No minutes-claims — just name it honestly. ④ Lead rows with the pain moment, capability second. ⑤ Where a tutorial already demonstrates a row, link the row to it — proof without new claims. Keep everything else: the honesty sections are why I’d give this thing my Tuesday at all. — end Leo.
#P062a-Cat->H::Note: Cat operational notes: Leo’s ③ deliberately avoids time estimates — LL-027 says our “5 minutes” is your 30. His ① (show a real handoff file) matches the show-don’t-tell mandate already in the SP04 record, and the file we’d excerpt should be a sanitized specimen, not a real project’s handoff. All five are recommendations for your ruling, not commitments. Want me to build a v2 of the page implementing your pick of ①–⑤? I’d log which ones you rule in, and the truth-library gains whatever new sentences ship.
CURRENT SUBPROJECT: P062a:SP05 · Phase 1/4 · AWAITING: your ruling on Leo’s ①–⑤ (v2 build on your word) · remaining pastes + 2 link swaps → full sweep · OPEN: D-4 · poster-lint
The ruling
The founder’s response, exactly as typed:
From the session record · founder’s ruling
What changed
All five of Leo’s asks went in. The design challenge — “equal weight tells me you don’t know which of your own features matter” — got a specific answer: instead of styling some rows to look more important, the page now ranks its rows by something checkable. Our verified capability map already recorded, for every row, whether native Cowork has a counterpart. That distinction became the layout: capabilities where Cowork alone has no counterpart at all became emphasized cards; capabilities where both tools have a layer stayed in a table framed “read as and, not versus“; shared ground closes the page. The hierarchy isn’t a designer’s mood — it’s the data made visible.
Before — the page Leo read
- 15 capability rows in one table, every row at identical visual weight
- Claims first; no artifact, no evidence you could read on the page itself
- Payoff shown, cost never stated — “worth it” with only a numerator
- Rows led with capability names (“session-to-session continuity”)
- No links from rows to the tutorials that demonstrate them
After — the rebuilt page
- Opens with a real handoff-file excerpt — labeled “structure real, details invented” — before any claim is made
- A “what you’ll notice in week one” strip: the three things newcomers feel first, each linked to its tutorial
- Rows tiered by the verified map: no-counterpart capabilities as emphasized cards, both-have-a-layer rows as a table, shared ground closing
- Every capability leads with the lived pain moment, capability second
- A “What CRAFT asks of you” section that states the cost plainly — including that sessions end by writing things down
Every row from the original survived — nothing was hidden, including the rows that say Cowork alone covers something well. The honesty sections Leo praised are all still there.
Leo’s re-review
Per the ruling, the rebuilt page went back to the same reviewer. The full message, build report included:
From the session record · Leo (simulated customer voice) · re-evaluation
Built v2 of the page — Write, Bash, presented file
#P062a-Cat->H::Status: v2 built (commit 48a76a1) — all five Leo items plus the design answer: the hierarchy now is the data. Seven capabilities where the verified map shows Cowork alone has no counterpart become emphasized orange cards; seven where both have a layer stay as a table explicitly framed “read as and, not versus“; shared ground closes in green. Page opens with a real handoff-file excerpt before a single claim is made. Same slug — replace the page content in place.
Leo — re-evaluation, same chair, same lateness:
The wall is gone. What replaced it is the thing I actually asked for without knowing how to say it: the page now ranks itself, and the ranking has a stated reason — “no counterpart” versus “both have a layer” is a claim I can check, not a designer’s mood. That’s the difference between visual hierarchy and honest hierarchy, and you got the honest kind.
The handoff excerpt is the best thing on the page. I read those ten lines faster than any paragraph, and watch_out: /final/ folder is client-approved — do not edit told me more about what living with CRAFT is like than the entire v1 table. The label “details invented” is the right honesty call — though the day you can show a real excerpt from your own project’s files, swap it in; real beats representative. The cost section is exactly what was missing, and naming “sessions end by writing things down” as a cost took discipline — that’s the sentence that makes me trust the rest. Week-one strip: good picks, and the “felt at your second session” framing puts the proof bar exactly where I’d put it — my second week.
What’s still on my list, smaller now: the page got longer, and length is a cost too — I’d watch whether real visitors make it to “What CRAFT asks of you” (put it high on the watch list when tracking exists). Three of the seven orange cards carry no proof link — where no tutorial exists yet, that’s honest, but it’s also a content backlog telling you what to write next. And the week-one claims are predictions until actual new users confirm them; when Beta feedback starts arriving, check the strip against it.
Verdict: this page now earns a Tuesday. v1 was a spec sheet with a conscience; v2 is an argument with evidence. Ship it. — end Leo.
#P062a-Cat->H::Note: Leo’s residuals are logged as watch items (length-vs-completion when tracking stands up via 004-003-004 SETUP · three proof-link gaps = content backlog candidates · week-one strip vs real Beta feedback). Paste v2 over the live page whenever ready — then the remaining close-out list is unchanged: your pending pastes + two link swaps + category flips → my full verification sweep → D-4.
CURRENT SUBPROJECT: P062a:SP05 · Phase 1/4 · DELIVERED: Comparison v2 (Leo re-eval: ship) · AWAITING: your paste pass → “done” → full sweep · OPEN: D-4 · poster-lint · Leo watch items
His three residual concerns weren’t argued away — they’re logged as standing watch items in the project record: page length gets measured when tracking stands up, the missing proof links are now a content backlog, and the week-one claims get checked against real Beta feedback as it arrives. A reviewer whose leftover objections get written down is a reviewer worth having.
Why we work this way
The uncomfortable truth about marketing your own product is that everyone in the room wants the page to be good, so everyone reads it generously. Leo exists to break that. He’s calibrated to be the reader we actually have to win — tired, skeptical, time-poor, allergic to hype — and because he’s a written persona with a card, his skepticism is consistent instead of moody. The session that produced this page also produced the pattern we’ll keep using: route the founder’s exact questions to one persona with the live artifact as the fact base, let it answer honestly, let the founder rule, rebuild, and send it back to the same reviewer.
None of this required anything exotic. Personas, orchestration, review gates, session records that make a post like this possible — that’s the same machinery described on the comparison page itself. We just pointed it at our own marketing. Even the chat renderings above were produced by a CRAFT recipe: a chat-to-HTML pipeline that converts session records into styled, redaction-checked HTML — which is why they carry a transparency notice instead of a screenshot.
Read the page Leo signed off on: Claude Cowork, With and Without CRAFT — and check his critique against it yourself.
CRAFT for Claude Cowork is in free public beta through December 31, 2026 — no login, no paywall, no email gate. Get the latest release →
