App page · split-view run surface

Revised /p/:slug. The chat-style single-column version hid the input once a run started — that broke trust. This brings input-left / output-right back, but with 2026 chrome: thin borders, quiet surfaces, typography-led output, one green accent. The input is always visible so you can see what you sent alongside what came back.

2026-04-22 v17 · supersedes chat-style draft source: /p/lead-scorer

← v17 index prior chat-style draft

State 1 · Idle — input focused, output shows the shape of what's coming back

floom.dev/p/lead-scorer

Inputs

action · score_batch

ICP description *

Used as the scoring rubric. Plain English — no schema.

Leads file *

leads-q2.csv

50 rows · 4 columns · 12 KB

Columns detected: company, domain, title, headcount.

Company size (headcount)

50 – 2,000

Rows outside this range get a 0 fit score.

Rubric preset

~1.2s · 5 free runs

Output

preview

Your result will look like this

Ranked table · one row per lead · fit score 0–10 + short reasoning.

5 columns · sortable

Company	Score

Hit Run to populate · first 5 runs are on us.

About this app Install in Claude Source

IDLE · input 3/4 · output preview 1/4 · ratio flips once you Run

State 2 · Running — input dims but stays legible, output streams row-by-row

floom.dev/p/lead-scorer

Inputs

read-only while running

ICP description

B2B SaaS, 200-2000 FTE, RevOps / founder / VP-level buyer, English-speaking markets.

Leads file

leads-q2.csv · 50 rows · 12 KB

Company size

50 – 2,000

Rubric preset

b2b-saas

Inputs locked until run finishes

Output

live · score_batch

row 19.2RampRevOps ICP · 850 FTE · strong domain signal

row 28.8LinearFounder-led · PLG fit

row 38.1FigmaVP Sales title · enterprise motion

row 46.4NotionWrong ICP tier · PMM contact

row 54.1Acme PlumbingOff-ICP · SMB services

row 68.7VercelFrontier devtools · series C

row 75.2SquarespaceWrong ICP · legacy stack

row 87.9AirtableOps-adjacent · 1.2k headcount

row 96.8WebflowBorderline · marketing-led

row 108.4RetoolInternal tools · 700 FTE

row 115.8Shopify merchantSMB · under threshold

row 12…Supabaseanalyzing domain + intent signals…

50 rows · est. 500ms remaining

About this app Install in Claude Source

input dimmed (opacity .72) · hover to focus · still legible · cancel lives in the header

State 3 · Complete — input stays put with "Edit & rerun", output fills with results

floom.dev/p/lead-scorer?run=r_a7f3

Inputs used

ICP description

B2B SaaS, 200-2000 FTE, RevOps / founder / VP-level buyer, English-speaking markets.

Leads file

leads-q2.csv · 50 rows · 12 KB

Company size

50 – 2,000

Rubric preset

b2b-saas

Sent 2 min ago

Output

done · 1.2s · 50 rows

Ranked results · sorted by fit score (desc) top 8 of 50

Company	Title	Headcount	Score	Reasoning
Ramp	Head of RevOps	850	9.2	RevOps ICP · strong domain signal
Linear	Founder	120	8.8	Founder-led · PLG fit
Vercel	VP Eng	540	8.7	Frontier devtools · series C
Retool	Head of GTM	700	8.4	Internal tools · enterprise motion
Figma	VP Sales	1,400	8.1	VP Sales title · enterprise motion
Airtable	Head of Ops	1,200	7.9	Ops-adjacent · mid-market
Webflow	Demand Gen Lead	480	6.8	Borderline · marketing-led
Notion	PMM	900	6.4	Wrong tier · skip

+ 42 more rows · view all · download CSV

share link floom.dev/r/r_a7f3

About this app Install in Claude Source

input stays put · "Edit & rerun" in situ · share link inline · no context shift

What changed vs the chat-style draft

Chat-style draft (rejected)

Single 720px column, inputs owned the stage until Run, then collapsed to a one-line bar
Post-run you could no longer see the ICP, the file name, the company-size range — just a summary label
To re-check what you submitted you had to click "Edit & rerun" and lose the output
Felt like ChatGPT, which is fine for conversation but wrong for "run a tool, check the result"

Split-view (this wireframe)

Always a split — never a full-page input or full-page output. The ratio flips with state.
Idle: input 3/4 (editable) + output 1/4 (shape preview of the table that's coming). Input is the focus.
Running + complete: input 1/4 (read-only recap) + output 3/4 (live stream or full result table). Output is the focus.
Input always visible — during run (dimmed, read-only) and after (read-only with "Edit & rerun")
Output running state streams row-by-row with a live pulse; you can scroll it while the input stays put
Output complete state fills with the real table; Run again + share link in the header
About / Install / Source demoted to quiet chips in a footer row — not primary nav

Design calls (please approve)

Ratio is driven by state, not fixed widths — idle uses 3fr 1fr (input owns the room), running + complete use 1fr 3fr (output owns it). Never full page for either column — you can always see what you sent and what came back.
Primary Run CTA lives in the app header, not inside the input panel. That way it's the same location idle / running (as "Cancel") / complete (as "Run again") — one button, three labels. Secondary Run button stays in the input-panel foot for idle only, because that's where the eye lands while filling the form.
Input dims to opacity .72 during run and becomes fully opaque on hover. Values stay legible even dimmed — this is the trust topic Federico flagged.
Idle output pane is a real shape preview — shows the column headers ("Company", "Score") and skeleton rows, not fake numbers. Tells you what the result will look like without pretending to be data.
<900px stacks vertically in all 3 states: input on top, output below. Same order, no disclosures hiding anything.
Score range: 0–10 with one decimal (9.2, 8.8, …) instead of 0–100. Easier to read at a glance, matches how humans rate things.
No emojis, single green accent, no colored left borders, no gradients. Design system same as landing + v17 Studio.

Contracts kept from prior drafts

/p/:slug?run=<id> hydrates into the Complete state (shared-run permalink).
data-testid="run-surface", run-surface-input, run-surface-output still apply — this is the same split-view the code already has, with a modernized skin.
BYOK modal, past-runs disclosure, CelebrationCard, RunCompleteCard semantics unchanged.
Multi-action tab strip (when an app has >1 action) would render above the split; out of scope for this wireframe.