Three ways to close the gaps claude fills with guesses. From the video, made specific so you can run them today. Concept credit: a Claude Code team member's field guide on finding your unknowns.
When: before you start anything where you don't know the territory, a new part of a codebase, a skill you've never done (design, video, legal docs), a tool you just installed.
The prompt, fill the brackets:
I'm about to [what you're doing]. My experience with this: [honest level, e.g.
"never touched it" or "I know the basics"]. What I already know: [one or two
things]. Do a blindspot pass: help me find my unknown unknowns here, the
questions I don't know to ask, the mistakes people make, what good actually
looks like. Then tell me how to prompt you better for this task.
Filled example:
I'm about to add authentication to my app. My experience with this: I've only
followed tutorials. What I already know: I want email login and Google login.
Do a blindspot pass: help me find my unknown unknowns here, the questions I
don't know to ask, the mistakes people make, what good actually looks like.
Then tell me how to prompt you better for this task.
Why the context matters: the pass is only as good as what claude knows about YOUR starting point. Telling it your level changes which blindspots it surfaces.
When: right before claude builds anything real, a feature, a document, a design. Run it in the same session, then let it build.
The prompt, fill the brackets:
Before you build [the thing], debrief me first. Ask me questions one at a
time about anything ambiguous, prioritize the ones where my answer would
change the approach. Keep going until you have no important questions left,
then summarize everything I told you as the spec, and build from that spec.
Filled example:
Before you build the money map dashboard, debrief me first. Ask me questions
one at a time about anything ambiguous, prioritize the ones where my answer
would change the approach. Keep going until you have no important questions
left, then summarize everything I told you as the spec, and build from that
spec.
The two upgrades over just chatting: it ends (no endless questions), and the answers become a written spec claude builds from, so nothing you said gets lost.
Not a prompt, a setup. Create a folder claude reads at the start of every session, so it stops guessing at who you are:
- me.md — who you are, what you do, how you like to work, your voice
- projects.md — what you're building right now, current state, what's next
- rules.md — your hard preferences: stack, style, things claude should never do
Point claude at the folder at session start (or use a CLAUDE.md that loads automatically in Claude Code). Every answer you'd give in a debrief that will STILL be true next month belongs in these files.
Want a working version instead of building from scratch? I published a starter you can fork and make yours: https://github.com/kju4q/harness-starter
The models keep getting better. The gap is you now. Close it once.
Built by @kju4q, more at github.com/kju4q