This method explains how the public Today Queue demo was produced from a cluttered private inbox/workspace surface without publishing private email, LinkedIn, calendar, or resume content.
Use it when an agent is asked to reimagine a crowded command center, CRM, inbox, recruiter tracker, or personal operating workspace.
The executable agent playbook lives at:
.claude/skills/minimal-inbox-redesign/SKILL.md
Claude Code can load it as a skill. Codex, Cursor, Windsurf, and other coding agents can read the same file as a portable methodology.
private screenshots / live UI
-> screenshot and workflow diagnosis
-> public pattern research
-> information architecture collapse
-> synthetic public demo
-> browser and privacy verification
-> README artifact trail
The goal is not a prettier dashboard. The goal is to reduce the number of things the user must interpret before acting.
Research interaction patterns, not visual decoration.
Useful reference buckets:
| Reference Type | Pattern To Extract | How It Applies |
|---|---|---|
| Task managers | Today, Upcoming, Inbox | Separate immediate action from staging |
| Email clients | Fast triage, split inbox, keyboard flow | Make reply/book/archive obvious |
| Linear-style tools | Command menu, fast lists, low-friction status | Prefer one searchable queue over page sprawl |
| Notion-style workspaces | Quiet hierarchy and editable structure | Let details unfold instead of showing every block |
| Agent skill repos | Repeatable procedures and verification loops | Make the redesign reproducible by other agents |
Capture each reference as:
Reference -> Pattern observed -> How it applies here -> What we will not copy
Most overbuilt personal workspaces can be reduced to four states:
Now = respond, book, decide today
Prep = research before acting
Batch = handle with a repeated lightweight action
Later = defer, archive, or low-fit
For this repo, the older workspace had many similar surfaces: Today, Inbox, Briefing, Career, Admin, Progress, Automations, Analytics, and Settings. The redesign collapsed the active work into one Today Queue plus supporting context.
The agent panel should be useful without making the page feel like a generic chat app.
For desktop agent-native workspaces, use a full right-side rail. For tablet/mobile, collapse it into a drawer.
The design philosophy is:
chat first
context compact
actions nearby
proof on demand
permission always visible
This means the right rail should not expose every internal artifact as separate always-open cards. The first view should show the selected object, a few short action commands, a usable conversation thread, and the composer. Longer readiness, tool trace, and missing evidence can be expanded when the user wants to audit the recommendation.
selected queue row
-> active pinned context
-> public research proof
-> agent suggestion
-> explicit human action
For a premium desktop rail, structure it as:
header: active object + local-first status
compact object bar: selected object + score + lane + proof state
mode state: mode / model / tool activity chips
run state: compact steps / tools / latency / cost / writes
quick actions: draft / explain / verify
chat thread: primary answer, follow-up messages, source chips, message actions, collapsible reasoning
inspect-run drawer: step plan / tool calls / model route / telemetry / write gate
inspect-context drawer: readiness / proof / trace / missing evidence / boundary
sticky composer: suggested prompt chips + short prompt input
The inspiration is a blended pattern:
The rail should answer:
It should also infer simple prompt intent. If the user types “draft”, answer with a draft. If they type “book”, “schedule”, or “calendar”, answer with a booking plan. If they type “research”, “background”, “verify”, or “source”, answer with the verification checklist.
The drawer should not auto-send email, auto-book calendar events, or expose raw private text in a public demo.
The redesigned rail borrows the NodeBench chat pattern: every useful answer has a run shape, not just a final message.
plan -> tool_start -> tool_done -> answer -> ledger
The visible rail should show the run in one line:
Run ready / 5 steps / 5 tools / 2.4s / $0.004 / 0 writes
The Parity Studio chat reference keeps this subtle: tool calls are small inline chips, tool completions are faint single-line rows, and the composer remains the main affordance. Avoid a wall of telemetry cards unless the user explicitly opens the run ledger.
Keep the first screen chat-first: the selected context should be one compact object bar, not a summary card. Move the recommendation and next action into the assistant response.
The detailed drawer should show:
For the Gmail/job workflow, the canonical public-safe tool trace is:
| Tool | Public-Safe Purpose |
|---|---|
gmail.thread.extract |
Extract sender, subject, snippet, and labels while keeping private bodies local |
pinnedContext.match |
Match against local resume/story-bank summaries |
nodebench.context.pack |
Fetch reusable public company/person/role context |
calendar.availability.suggest |
Suggest a time block without writing to calendar |
gmail.draft.create |
Show draft capability but keep execution approval-gated |
This is intentionally not an always-open telemetry wall. The run strip gives trust at a glance; the drawer gives auditability when the user asks, “what did the agent actually do?”
Do not publish:
.env valuesPublish only:
Before claiming the artifact is public-ready:
The Today Queue demo applies the method as:
Many private inflows
-> daily sweep snapshots
-> ranked career/action queue
-> Now / Prep / Batch / Later filters
-> pinned-context fit reasons
-> selected detail card
-> right-side agent rail
-> synthetic public artifact
The public demo is intentionally smaller than the private workspace because the demo should prove the decision model, not recreate the user’s entire inbox.