When should Claude redesign the whole dashboard?
Answer for Dani · 21-11-2026 · read from the dev branch at commit 3728bae · nothing was changed in the repo
Checked by the mate (Opus 5.5), 21-11 afternoon: what changes
The approach on this page is right. Fix what each screen shows first, then the design agent does only the look. Two waves is also right, because otherwise the whole look waits on your 12 answers and on Max. But five things were missing or wrong:
- The tests lock the table layout. 24 places in 6 test files find a value by its position, for example "the 6th cell of the row". If the design agent turns a table into cards, those tests fail, and it may not touch tests. Prep job (an agent, $0): before the design session, give every shown value a field name (
data-field="reply_rate") and make the tests look for that name, not the position. - Real data never goes into a design tool. Higgsfield and similar tools are outside services and not Claude. The design agent works only on
synthetic.html(made-up people), never on the live dashboard. No screenshot of real data goes anywhere. - Nothing loads from outside. No Google Fonts, CDNs or tracking. Fonts, images and Blender renders are files in the repo, kept small (Cloudflare allows 25 MB per file). The page should stay fast.
- Wait for T4 first. T4 (#40, #42) changes the same list screens. #42 also adds something to the screen: a warning when a never-break rule breaks. This page said "no screen", which is wrong. Two agents editing the same files at once causes conflicts.
- A simpler contract. One file,
system/dashboard/DESIGN-BRIEF.md, with per screen the fields, the rules and the off-limits files. That replaces a JSON file per screen plus a new contract test. The field names from point 1, together with the existing tests, already enforce it. The scope check is onegit diffcommand.
Also corrected below:
- The live reader now has a working key. It is on, but reads nothing because of the cost bug, and a crew is fixing that now.
- #20 is closed.
- The collector was redeployed.
check:numberschecks the database, not the page. Page numbers are caught by the screen tests.- The test count is 907.
- The panel moves to wave A: its fields are final, and going live only switches the buttons on, so both button states are designed.
When, and how you'll do it
- Now (agents): T4 is built and merged. The prep job, point 1, plus
DESIGN-BRIEF.md, takes a few hours at $0. - Then the mate: makes a separate folder on its own branch,
design/dashboard, cut fromdev, and gives you its path. - You: add your design skills to that folder's
.claude/skills, nothing installed globally. Open Claude (Opus) in that folder and say "design only, follow DESIGN-BRIEF.md". It runsnpm run devand works on/synthetic.html. - Back to the mate:
- full tests, typecheck and lint;
- the scope check;
- screenshots in German and English (made-up data);
- a review;
- then the merge into
devand the dashboard deploy.
- Wave B later: the judge verdict, the Ideen screen and the money card with pause and limits. These are built straight in the new look, so a second big design session may not be needed.
The short answer (Sonnet's, with the corrections above)
- Your idea is right. Freeze what each screen shows first (numbers, maths, wording keys, rules). Then a design agent only builds the look and never decides what is shown.
- Do it in two waves, not one. Do not wait for the whole system to be finished, and do not redesign the unfinished parts.
- Wave A can start now. The screens whose content is final: login, numbers, cohorts, openers, people, follow-up list, break-off (core), the health cards. Plus the shared design system (colours, tables, rate cells, language switch).
- Wave B waits for four tickets: #38 judge verdict screen, #43 control room part 1 (pause and limits), #30 switch live (the panel) and Dani's yes on #8 (Umgesetzt). The coach screen waits for the coach-in-the-cloud decision.
- The tool that makes it safe: one content contract file per screen. The design agent may not change it. Tests plus
check:numbersmust stay green afterwards.
1. What the dashboard shows today, screen by screen
Source: system/dashboard/src/*.tsx and texts.ts on dev. “Final” = the fields and wording are settled; only rows of data change. “Changing” = a ticket that is not done will add or change fields.
| Screen (file) | What it shows | Data comes from | Ticket | Content |
|---|---|---|---|---|
Login and headerShell.tsx | Login form; header with the German/English button, log out; tab bar. Logged out, nothing else is drawn. | Supabase login; allowed_users (stored language) | #17 (logins), F7.4 | final (tab bar grows when new screens come) |
Overview: numbers by week and dayNumbers.tsx | One table per period kind. Per channel: messaged, replies, reply rate, bookings, booking rates. Every percent shows its period and its base group. | views numbers_by_week, numbers_by_day; counting_rules | #17 | final (open items are checks, not content) |
Booking rate by first-message weekCohorts.tsx | Of the people first messaged in a week, how many have an appointment. Young weeks are marked “so far”. | view cohort_numbers_by_week | W6 | final |
OpenersOpeners.tsx | Replies and bookings per opener variant, by week and day, totals, an “unrecognised openers” count and a note that rows are no fair comparison. | views opener_numbers_by_week/_by_day, opener_totals, unrecognised_openers | #19, #20 | final (#20 closed 21-11: before 11-08 is out of scope) |
Health cards and warningsCollector.tsx, Backup.tsx | Warning at the top of every screen when the collector stopped or the backup is over 48 hours old or unreadable; a card with the facts behind it. | collector_health, backup_health | #12 (collector), #13 (backup) | final (backup shows “none” until the second Supabase project exists) |
Money line and cardBudget.tsx | A line on every screen: what is left of the limit. Card: limit, spent, held by running calls, left, a bar, newest call, spend per part. | function claude_budget(); limit from the model table | W4 of 20-11, #44 (cap $60) | changing (#43 adds pause switch, daily limit, forecast) |
Test number cardShell.tsx | The first “tracer” number: a message count with its source and cut-off time. | view tracer_number | #10 | final but a leftover proof card. My suggestion: drop it, ask Dani. |
PeoplePeople.tsx | Search by person or by message text. One person: details, first and last contact, whole history, reader summary and “where the conversation stands”. Other language by button, one message at a time. | search_people, search_messages, person_overview, person_reader_forms, person_messages, message_translations; function translate | #18, #26 | final |
Break-off reportBreakoff.tsx | Where replied chats stop, step by step, with base and period; the causes with counts; examples that link to the chat; the health line for reasons; category list (new ones can be rejected once); reader accuracy. | breakoff_base/_steps/_causes/_examples, reader_review, reason_category_list, reason_health, reason_chats, reason_accuracy_latest | #27, W3, W5 | final fields; rows change when #37 turns on |
“Said yes, no appointment” blockSaidYes.tsx (inside break-off) | Counts per signal (email, clock time, call link, our “booked”), then the chats with a link. No message text. | said_yes_summary, said_yes_chats | W1, #40 | changing (#40 removes known clients: fewer rows, same fields) |
“Quiet after our message” blockQuietAfterOurs.tsx (inside break-off) | Counts per kind of our message and the top 20 normalised messages. Every count opens its chats. | quiet_summary, quiet_top_messages, quiet_message_chats | W3, #39 | final (#39 is speed only) |
Follow-up “Nachfassen”FollowUp.tsx | Who replied and went quiet, sorted by the reader's interest: days silent, summary, link to chat, open channel. CSV of exactly these rows. | view follow_up_candidates | #28 | final (#40 removes kiaus contacts: rows only) |
Panel “Wer braucht uns”Panel.tsx | Who needs a person, most urgent first, with the recap; taken-over chats with the 24-hour warning; switch buttons “Übernehmen” and “Zurückgeben”. | view panel_conversations; function takeover | #29, #30, #31 | final fields (mate: going live only turns the buttons on; design both states, enabled and disabled) |
CoachCoach.tsx | Weekly brief, briefs on request, chat answers; a box that queues a question (limit 10 open per login). Quotes as written. | table coach_outputs | #32, #41 | will change (coach moves to the cloud with visible thinking: control room plan) |
Ideen + UmgesetztIdeen.tsx, Umgesetzt.tsx | Ideas as a conversation between Dani, Max and the system: pick, change, merge, approve or send back; wording from the writer; copy-ready German wording, the “Umgesetzt” button, live-check state, days left of the judge's week. | tables ideas, idea_events, idea_drafts, umgesetzt | #33, #34, #8 | will change (#38 puts the judge verdict here; #8 waits on Dani's dated yes) |
GradingGrading.tsx | One chat at a time out of the 50, with the questionnaire; Dani and Max enter one agreed answer; progress; the key is frozen with a version. | grading_sets, answer_key_answers, crm_messages; freeze_answer_key | #22, #23 | final but a one-time tool. Lowest design priority. |
2. What is still coming to the dashboard
Sources: V1-TEAM-PLAN-20-11.md, NIGHT-RUN-RESULT-21-11.md, CONTROL-ROOM-PLAN-20-11.md, T6-JUDGE-DESIGN-21-11.md, V1-EVALS.md, V1-ENGINE.md. Night run of 21-11: T4, T7, T8, T9 did not start (#39, #40, #42, #17, #24, #26, #35, #43).
| Ticket | What it adds or changes on screen | State on 21-11 | Waits on |
|---|---|---|---|
| #38 Judge on the reply count | A verdict block per test: “{n} of {needed} chats, about {w} more weeks”; winner / not clear yet / tested, didn't work / lost on harm; buttons Extend and Revert; untouched steps shown beside; bookings shown beside (never deciding); wording-seen share; a note when the wording was never found; “retest allowed” (Dani only). Eight new text keys, German and English. | Design page merged. Not built. | Dani's 5 answers: 28-day before-window, 15-point minimum effect, never-found capped at “not clear yet”, untouched steps show-only or blocking, keep “booked” as a testable step |
| #43 Control room part 1 | Pause switch and daily dollar limit per paid part (reader, labeller, coach, writer); money forecast. Later parts: live map, activity feed, build-agent status, coach in the cloud with visible thinking, reader you can watch. | Not started. Only a plan. | Dani's 7 questions (who sees it; switches first; coach cost; thinking summary or full; build agents here or separate; how live; Telegram alerts) |
| #37 Live reader | No new fields. Lists and break-off fill within about a minute of a new message. Reader forms appear by themselves. | Key fixed and switched on at 10:09 on 21-11, but it reads nothing: one chat with 6,036 messages is over the $0.30 day on its own and blocks the rest. | Agents: a crew is fixing it now |
| #39 Per-chat facts table, faster lists #40 Said-yes list without known clients | Same fields, less waiting. Said-yes list and every lead list drop known clients and kiaus contacts (Max's friends). | Not started (no time or memory slot in the night run) | Agents. Nothing waits on a person. |
| #42 Never-break rules on the real tables | Adds a screen element (mate correction): a daily check, e.g. “nobody on said-yes has a real appointment”, and a broken rule shows on the dashboard (ticket text). Design the warning in wave A. | Not started | Agents |
| #30 / #31 Switch live and test with Max | The panel's “Übernehmen” and “Zurückgeben” become real buttons. A note about ai_off is on the panel. | Built and deployed; collector redeployed 21-11. Not live: no switch key set. | Max and Dani: key, tag, three passes. Still open for Max: does ai_off fire on every reply or only on TEST_DANI? |
| #8 / #34 Umgesetzt | Block is on the Ideen screen now. Live check finds the new wording in the AI's messages. Real Haiku test needs gate A34.live (under 1 cent). | Merged and deployed | Dani's dated yes closes #8; yes or no to gate A34.live |
| #13 Backup live | The backup card gets real values. | Code merged. The second Supabase project does not exist. | Dani creates it |
| #20 Openers | Openers screen: “unrecognised” line. W17-W20 added; since 11-08, 7.7% unrecognised, the rest explained. | Closed 21-11 | Nothing (Dani: before 11-08 is out of scope) |
| #32 / #41 Coach | First real brief stored (grade B). Coach chat answers only while Dani's PC runs the script; cloud version is part of the control room. | #41 closed; #32 open | Control-room decision |
| #36 First real cycle | No new screen. The cycle goes through Ideen, Umgesetzt and the judge verdict. | Not started | #38, switch test, Dani and Max |
3. The rules a redesign must keep
Every rule below is already written down and, where marked, already tested. The design agent gets this table with the contract.
| # | Rule | Source file | Tested by |
|---|---|---|---|
| R1 | Every number is a value from the database and is marked data-value. Nothing is counted or divided in the page. | V1-EVALS.md F7.1, L7.1; header of Numbers.tsx; Shell.tsx | dashboard.test.ts, the screen tests, check:numbers |
| R2 | Every percentage shows its period and its base group. | V1-EVALS.md F7.2; RateCell.tsx | F7.2 test reads the page |
| R3 | A rate whose base is under 30 people (MIN_BASE) is grey and says “too few”. Break-off has its own minimum per cause. | numbers-backend.ts, RateCell.tsx, _shared/breakoff.ts (MIN_CHATS_PER_CAUSE) | numbers-w6.test.ts, breakoff-screen.test.ts |
| R4 | A week whose people can still book is marked open: “so far”, not final. | Cohorts.tsx | numbers-w6.test.ts |
| R5 | German and English with one button. One text on screen at a time, never two side by side. Every fixed word comes from texts.ts. Max reads only German, Dani only English. | DECISIONS.md “German / English switch (Dani, 15-11-2026)”; texts.ts; BRIEF decision 9 (German words such as Angeschrieben, Antworten) | dashboard.test.ts (every text node outside data-value must come from texts.ts) |
| R6 | Customer messages are shown as written and marked data-value, so the language button never changes them. A translation is stored beside the original, never replaces it, and only Claude translates. | DECISIONS.md same row; V1-EVALS.md F26.1; header of Grading.tsx | people-screen.test.ts, grading-screen.test.ts |
| R7 | A shown quote is copied word for word. The reader's “stated reason” is found word for word in the chat (100%). A coach example is a real message it was shown. Quotes are never translated. | V1-EVALS.md A2, Q3, F32.2; Coach.tsx header | coach-screen.test.ts, coach-refusal.test.ts |
| R8 | Logins for Dani, Max and Adnan, each their own. Logged out, nothing is visible: the page carries no data and tells search engines to stay out. | DECISIONS.md Q7, Q10; V1-EVALS.md F7.4; Shell.tsx; public/_headers | check:dashboard, check:access |
| R9 | Every dashboard number must match an independent recount from the raw tables (code that shares nothing with the dashboard). Last run: 1,068 rows, 0 differences. | V1-TEAM-PLAN-20-11.md section 5; V1-EVALS.md F7.1, L7.1; system/scripts/check-numbers.ts | npm run check:numbers |
| R10 | The funnel stops at “booked”. No show-ups, closes, conversion or revenue. “Booked” means an appointment in the CRM. | BRIEF.md Locked decision 4; CONTROL-ROOM-PLAN-20-11.md (W1) | counting rules tests |
| R11 | Reasons the person stated are kept apart from reasons the AI guessed. The reader's own guess is shown, marked. | V1-ENGINE.md station 3; V1-EVALS.md Q3 | breakoff-screen.test.ts |
| R12 | No bare codes on screen. A code is shown as words; an unknown code is humanized, never raw. | Logs/Log-Folder-Self-Improving-System/20-11-2026-w3-dashboard-reasons-story-health-line-drill-down-no-bare-codes.md; codes.ts | reasons-dashboard.test.ts |
| R13 | A read that fails is shown as a warning or message, never as silence and never as an invented number. The failure of one block never fails the rest. | App.tsx; Collector.tsx; Budget.tsx; V1-EVALS.md L4.3 | claude-budget-screen.test.ts, collector-dashboard.test.ts, backup-dashboard.test.ts |
| R14 | The Anthropic budget stays in view on every screen: the limit and what is left, going down as it is spent. Limit is $60 now. | DECISIONS.md “Budget for Anthropic API use”; NIGHT-RUN-RESULT-21-11.md; Budget.tsx | claude-budget-screen.test.ts |
| R15 | Human in the loop. People decide what is tried. Every idea shows where it came from. The Umgesetzt click cannot be undone and the screen says so before it. The wording to copy is always German. | V1-ENGINE.md “Human in the loop”; Ideen.tsx, Umgesetzt.tsx headers | ideas-screen.test.ts, umgesetzt-screen.test.ts |
| R16 | The judge shows harm beside the result and never hides it. Bookings are shown beside the verdict. Never promise or claim anything we cannot deliver (applies to every text). | DECISIONS.md Q28; T6-JUDGE-DESIGN-21-11.md sections 4, 6-8 | judge tests (when #38 is built) |
| R17 | The dashboard writes only what each screen's backend already allows (ideas, events, Umgesetzt click, answer key, reject a category, a coach question, the switch). A redesign adds no write. Never an action on the CRM except the switch. | headers of Ideen.tsx, Panel.tsx, Grading.tsx, Coach.tsx; V1-ENGINE.md “Takeover switch” | check:access |
| R18 | Proof screenshots contain no names and no message text. | V1-EVALS.md H7.1 | made with the synthetic page synthetic.html |
Note: BRIEF.md decisions 5 and 8 (typed-in numbers, rates from the sheet) were replaced when we integrated with the CRM. The working files win where they differ.
4. The right moment, and how to do it
Two waves
| Wave | Screens | Starts when | Why |
|---|---|---|---|
| A: now | Design system (colours, type, table, rate cell, cards, tab bar, login, language button, grey “too few”, empty / loading / failed states) and the screens marked final: Login, Overview numbers, Cohorts, Openers, People, Follow-up, Break-off with its two blocks, Panel (both button states), Collector and Backup cards, money line, the #42 broken-rule warning. | Mate: after T4 (#39, #40, #42) is merged and the prep job (field names + DESIGN-BRIEF.md) is done. A few hours of agent work, $0. | Their fields are final. Open tickets there only change which rows are shown, not which fields (#37, #39, #40). |
| B: after four tickets | Ideen + Umgesetzt + judge verdict, money card with pause and limits. (Panel moved to wave A by the mate.) | #38 built, #43 part 1 built, #8 yes given | Each adds fields or buttons. Designing before would mean redoing it. |
| Later, separate | Coach in the cloud with visible thinking, live map, activity feed, build-agent status (control room parts 2 to 5). | Dani answers the 7 control-room questions | Their content is not even decided. Give them a “Control room” area, so wave A and B never wait for them. |
Tickets that must be done first, so the content is frozen
| Ticket | Needed for | Needed before | Who |
|---|---|---|---|
| #38 judge built (after Dani's 5 answers) | the verdict block, Extend / Revert, untouched steps, bookings beside | wave B (Ideen) | Dani answers, agents build |
| #43 control room part 1 | pause switch, daily limit, forecast on the money card | wave B (money card) | Dani answers 7 questions, agents build |
#30 (and #31 test), ai_off question | the panel's real buttons and what they say | wave B (panel) | Max and Dani |
| #8 yes | the Umgesetzt block as it stands | wave B (Ideen) | Dani |
| #39, #40, #42 (T4) | final shape of the lists, and the broken-rule warning | wave A (mate: they edit the same screen files, so the design waits for them to avoid conflicts) | agents (starting today) |
| #37 reading again (ceiling bug) | real forms arriving by themselves, to look at the redesigned lists with live rows | not needed to start; needed to review wave A on live rows | agents (crew running) |
| #36 first real cycle | nothing | not needed. V1 proven does not have to come before the look. | - |
The content contract: one file per screen mate: replaced by one DESIGN-BRIEF.md, see top
One file per screen, kept in the repo next to the screen, for example system/dashboard/contract/numbers.contract.json. The design agent receives these files and the rules table. It may not edit them. A person (Dani) signs each one with a date; that is the freeze.
Each file lists
- Screen id, ticket, who may see it (Dani, Max, Adnan)
- Status: frozen on a date, or open with its blocking ticket
- Fields: key, text key in
texts.ts(German and English), source (view or function and column), format (formatCount,formatPercent,formatDate,formatTime,formatUsd),data-valueyes or no - Order and grouping: what comes first, what the rows are (per channel, per week)
- Rules that apply: R-numbers from the table above, e.g. a percent field lists R2, R3
- States: loading, empty, failed, too few, so far
- Actions: the only writes the screen may do
Rules for the design agent
- May change: layout, colours, type, spacing, icons, charts that show the same values, animation, responsive behaviour, empty-state look.
- May not change: which fields exist, their order of meaning, their text keys and wording, their source, their format, any formula, any
data-value, any rule R1 to R18, any write. - May not touch:
texts.ts,*-backend.ts,supabase/,tests/, the contract files. A needed change is a request to Dani, not an edit. - New wording or a new field is a new ticket, not part of the redesign.
A sample contract (one field, short)
{
"screen": "numbers",
"ticket": 17,
"status": "frozen 2026-11-22 by Dani",
"see": ["dani", "max", "adnan"],
"rows": "one per period and channel, newest first",
"fields": [
{ "key": "reply_rate",
"label": "numbersReplyRate", // key in texts.ts, de + en
"source": "view numbers_by_week.reply_rate",
"format": "formatPercent",
"dataValue": true,
"rules": ["R1", "R2", "R3"],
"shows": ["period start", "base group", "count and total"],
"belowBase": { "min": 30, "style": "grey", "text": "numbersTooFew" } }
],
"states": ["loading", "failed", "empty"],
"writes": []
}
The label key name is an example. Real keys are the ones already in texts.ts.
How the redesign is checked afterwards
| Check | What it catches | How |
|---|---|---|
| Full test suite stays green (907 of 907 at the last full run) | A changed text, a lost data-value, a lost rule | npm test, npm run typecheck, npm run lint |
check:numbers stays at 0 differences | The database views against an independent recount (mate: it does not read the page; a wrong number on the page is caught by the screen tests) | run after deploy |
check:dashboard and check:access | Anything visible logged out; a new write | run after deploy |
| New, in the prep job: field names | A value lost or moved when the layout changes | every shown value gets data-field; the screen tests find values by name, not by cell position (24 places in 6 test files today) |
| New: scope check on the diff | The design agent touched a file it may not touch | git diff --name-only dev...design/dashboard may list only styles.css, screen .tsx files and design assets; texts.ts, backends, supabase/, tests/, DESIGN-BRIEF.md unchanged |
| Both languages, side by side | A layout that breaks with long German words | screenshots with synthetic.html (made-up data), via desktop-vision. No names or message text. |
| Pristine review, then Dani looks | The rest | no merge to dev without a clean run; no merge to main without Dani |
5. What I need from Dani
| Decision | My recommendation |
|---|---|
| Start wave A now? | Mate: after T4 and the prep job (both agents, $0, started today). Then you get the design folder path. |
| Drop the old “tracer” test card from the overview? | Yes. It was the proof of ticket #10 and says nothing a person needs. |
| Where do the control-room parts go? | A separate area, designed after the 7 questions are answered. |
| The 5 judge points (#38) | Answer them soon: they decide what the verdict screen has to show. They are the biggest blocker of wave B. Defaults in T6-JUDGE-DESIGN-21-11.md are 28 days, 15 points, cap on never-found, show only, keep “booked”. |
| #8 dated yes, gate A34.live | Quick answers that clear wave B. (The Anthropic key is done.) |
Checked and corrected by the mate (Opus 5.5) on 21-11-2026 afternoon against dev 3728bae, the live state and the test files. Sonnet's original: read from dev at 3728bae and the open GitHub issues on 21-11-2026. Not verified live: no database, CRM or browser was used. Ticket numbers are the GitHub numbers, as the screen files name them.