Daniel Cohen

How I work

Get your work challenged before it reaches people

Ask Claude to play the people who will read your doc, and see which questions each would raise first. Nine steps follow, each with what AI changes and a prompt you can copy.

Try these this week

Three habits you can copy today. Each is one prompt. Paste the prompt, fill the brackets and read what comes back with your own judgment.

  • 1 of 3

    Run a recipient panel on your next doc

    What it changes for you. Simulated readers raise likely objections before your real readers see the doc. You fix the ones two or more of them raise first. Your real readers still decide.

    Get the recipient-panel skill

    Paste only what your employer allows. Remove names and customer details first. The prompt below has Claude start one subagent per reader. A subagent is a separate AI helper Claude starts for one task.

    Prompt to copy
    Before I send this, run a recipient panel. The readers are: [role 1], [role 2], [role 3]. For each, spawn a subagent that reads it as that role and reports where they stop, their objections with the line each rests on, one fix each, and what would make them say yes. Then merge and tighten.

Document: [paste it]
    No subagents? One new chat per reader
    You are [role: for example the finance lead who approves the budget]. You must [what they decide or do with this doc]. You already know [their context]. You have [their patience, for example two minutes]. You say no when [their reasons].

Read the document below once, as this person. Then reply in under 300 words:
1. The line where you would stop reading, quoted.
2. Up to five objections, each with the line it rests on and one fix.
3. What would make you say yes.
4. One thing that works.

Document: [paste it]

    Open a new chat for each reader, then paste all the answers into one chat and ask it to merge the objections two or more readers share.

  • 2 of 3

    Ask for the strongest counter before you decide

    What it changes for you. You hear the best argument against your choice while changing it is still cheap, and you get a test that could settle it.

    Get the product-sparring skill

    Prompt to copy
    Challenge this decision as a peer product lead. Give your position, the strongest counterargument, what the plan leaves out and the cheapest test that could change my mind.

Decision: [what I plan to do]
Evidence: [what I have so far]

    The product-sparring skill adds the questions a peer product lead asks first.

  • 3 of 3

    Test a flow with personas from your exact segment first

    What it changes for you. You find likely confusing steps before you spend real users' time on them, so your interviews go to what only people can answer.

    See step 2, learn from users

    Prompt to copy
    Before I test this flow with people, walk it as personas from my exact segment. Segment: [who they are, the job they are trying to do, where and when they do it]. Build three personas plus one skeptic (four in total, one batch), each with a job to do, a context and a reason to quit. Spawn a subagent per persona. Each walks the flow step by step and reports where they hesitate, what they expected next and where they would quit. Merge by how many personas hit the same point, then list the questions only real people can answer.

Flow: [paste the flow or the screens]

    Persona answers are a rehearsal. Only real people's answers show demand.

    No subagents? A chat in the Claude app does not start them. Paste the single-reader brief on card 1 into a new chat once per persona, describe the persona in the role line and paste the flow where it asks for the document.

Where a second read changed the work

Three times a second read changed what I did: a panel of simulated readers, a persona walk and a check of AI output. Each limit sits next to the result.

  • Oct '26 · This site

    Install commands that would fail

    Simulated reader panels read this site as PMs on a fresh Mac. They flagged install commands that would fail, and I fixed them. Two simulated PM reviewers said they could not install without a terminal, so each skill now has a zip you upload in the Claude app.

    Limit. The reviewers were AI personas. A real person on a fresh Mac is the check that remains.

  • Oct '26 · Wellness Coach

    Entries that led nowhere

    Persona walks of the app turned the Start sheet's "not available yet" entries into timer chips. They also removed a Talk button for a voice feature that did not exist.

    Limit. As of Oct '26 the re-check is pending. The walks were AI personas, and the simulator has not confirmed the fixes yet.

  • Jul '26 · Sparktale

    AI conclusions that were wrong

    Three conclusions from AI agents were wrong. Each was caught only because it was checked. One was a report of 99 type errors that came from a stale build folder.

    Limit. A red team or a panel points to claims to check. Each correction still needs its source.

Simulated readers rehearse questions. Evidence of demand comes from real users.

How I use AI

I use AI to explore possible solutions, gather sources and challenge my assumptions. I check the evidence behind a claim before using it to decide. The decision stays mine.

Nine steps through product work

Enter at your current task. Return to an earlier question when new evidence changes the work. Use the guides in the documents and tools your team already has.

Each step says what AI changes for you, gives a prompt to copy, and shows what I check and what the work leaves behind.

  1. 1Frame the problemAI more sources, and a counter to your framing
  2. 2Learn from usersAI personas from your segment rehearse the flow
  3. 3Define successAI checks your metric against the question and the data
  4. 4Make a decisionAI the strongest counter, then a stakeholder panel
  5. 5Scope the workAI simulated engineering and design readers question your brief
  6. 6Support deliveryAI checks each change against the brief
  7. 7Check the workAI a red team that did not build it
  8. 8Prepare the releaseAI simulated users read the release note
  9. 9Measure and learnAI checks the inputs and argues with your reading of the result

Read the full nine steps

Principles behind the choices

  • 1 of 6

    I make time to choose the work

    I make time to decide which problems deserve attention and what to leave out.

    Where it showed up

    I built LandMyCV alone with Windsurf, Claude Code and Codex, so the open questions were about the offer. I chose monthly and quarterly plans for job searches that last a few months. I kept the analysis free because it convinces a visitor the product works. The DOCX and PDF download became the paid part.

  • 2 of 6

    I write down what would make me stop

    The rule names who counts, when the result is read and what passes. On the read date I compare the result with the rule before I decide whether to go on.

    Where it showed up

    I wrote the kill line (95% of the estimator's rows found, per drawing set) before building the redesign. No measured set reached it, and I stopped before the estimator-time test.

  • 3 of 6

    I work on what limits progress most

    I name the main limit on progress before adding more work. Before I spend to get more sales or users, I work out what one of them leaves after every cost.

    Where it showed up

    Reaching parents who would buy a personalised book limits Sparktale's growth, so I work on that first. I wrote a kill test before spending.

  • 4 of 6

    My starting list is what a general model already does

    Before I build on a model, I write down what a general model already does and which data or integration I have that it lacks. That second list is what a competitor would have to build or buy. If it is empty, I keep the idea small or drop it.

    Where it showed up

    A model can tailor a CV. What LandMyCV added was the workflow around it: a main CV, a tailored CV per role, an application pipeline and DOCX or PDF export. By Apr '26 general models tailored CVs well when given the right steps. I paused development.

  • 5 of 6

    I tie the work to a goal

    I write down how a proposed change serves the product goal before it enters the plan.

    Where it showed up

    SmartChef, my first app. The roadmap held import, meal planning, a shopping list and tags. Small features were fun to build and pulled me off the goal of all recipes in one place.

  • 6 of 6

    I look for why the user's task keeps failing

    What drives me is working out what is actually going on for the user: what they are trying to get done and why it keeps not working. I start with real interviews where I can get them. I write the finding down before any feature exists.

    Where it showed up

    My interviews with construction estimators put building a scope of work from drawings at 16 to 24 hours per project, depending on scope. That identified a task to investigate. How much time Parapet would save stayed untested, because the timed estimator test never ran.

I follow the work of Marty Cagan, Lenny Rachitsky, Aakash Gupta, Paweł Huryn, Claire Vo and others. Who they are and what I took