Wooolfmesh Research / August 2026
We are simplifying Wooolfmesh into calm work studios. A Studio keeps the task, useful memory, sources, AI, and the current result in one place. The product should reduce navigation and mental overhead, not create another system to manage.
The previous direction could become a productivity control room: many pages, widgets, indicators, and modes. We now prefer a smaller mental model. Open a Studio, see where the work is, continue it, use AI or memory when useful, and leave with a clear next step.
Memory retrieval, task state, local search, ML ranking, agent runs, Obsidian context, and AI providers can become more capable without making the screen more complex. More intelligence should often produce less interface.
2026 direction
These are product hypotheses built from current tools and HCI research. They are not promises.
Use the same basic layout for research, planning, writing, analysis, or execution. Change the content, not the whole product grammar.
Why: stable geometry builds location memory. Adaptive interfaces can help, but research does not show that constantly changing menus are automatically better.
Test: keep navigation and primary actions fixed while adapting only the relevant content.
Adaptive UI study, 2026Show the shortest useful answer first. Reveal evidence, memory, assumptions, history, and agent detail when the user asks or when risk requires it.
Why: a 2026 experiment on generative-AI exploratory search found that progressive disclosure improved engagement and knowledge change while reducing unnecessary cognitive demand.
Test: answer → key evidence → details, instead of one large AI wall.
Kim & Ji, 2026A Studio should know its files, instructions, conversations, decisions, tasks, and selected memory. It should not silently mix every memory from every project.
Why: modern project workspaces increasingly use project-scoped context so long-running work can continue without repeating the background.
Test: make Studio context visible and editable as a small context packet.
OpenAI Projects, 2026The document, plan, checklist, analysis, task, or other result is the main surface. Chat is a tool beside the work, not the place where the work disappears.
Why: ChatGPT Work and other modern AI workspaces separate conversation from finished work and support longer task execution inside project context.
Test: allow the AI panel to collapse completely without hiding the current result.
ChatGPT Work, 2026Sources should belong to the current work, not live as a flat pile. A Studio can keep active sources, saved evidence, and output together.
Why: a CHI 2026 field study found frequent tab overload and showed value in supporting the structures that emerge around web work.
Test: a small Sources shelf grouped by current task, not another file manager.
CHI 2026: From Tabs to StructuresWhen a Studio opens, first show where work stopped and what the next concrete action is. Do not start with statistics about yesterday.
Why: unfinished work can leave attention residue, while externalizing next actions reduces the need to keep them active in memory.
Test: one compact resume line: last outcome, next action, blocker.
Leroy, 2009Do not show a permanent memory feed. Retrieve one to three memories only when they are likely to change the current decision or action.
Why: external memory is useful when it reduces mental load, but retrieval itself can become noise.
Test: a small “Relevant from memory” layer with dismiss, pin, and “why this?” controls.
Risko & Gilbert, 2016Scheduled or agentic work should appear inside the Studio it serves: last run, result, next run, history, and exceptions.
Why: current agent products are moving toward persistent, recurring work attached to project context rather than isolated one-shot chats.
Test: no global automation dashboard for normal use. A global view can exist only for troubleshooting.
ChatGPT Work release notesEvidence → hypothesis
We keep the evidence separate from the feature idea. Humans have already suffered enough from products that turn one paper into fourteen settings.
Finding: progressive disclosure in a 2026 generative-AI search experiment improved engagement and knowledge change and reduced extraneous cognitive demand.
For us: AI output should unfold in layers. Start with the decision or next step, then allow deeper evidence and reasoning.
International Journal of Human-Computer Studies, 2026Finding: a 2026 study compared many adaptive graphical menus with a static baseline. Results differed by design; there was no simple rule that adaptation itself improves the experience.
For us: adapt relevance, ranking, and suggestions. Keep core controls predictable.
Journal of Systems and Software, 2026Finding: CHI 2026 research with knowledge workers found persistent problems with excessive inactive tabs and argued for capturing the changing structures between pages.
For us: a Studio can hold sources as part of a task structure instead of reproducing a browser-like list.
CHI 2026Finding: a six-month randomized field experiment across more than 6,000 workers found that integrated AI users spent less time on email and completed documents somewhat faster, while meeting time did not significantly change.
For us: target preparation, retrieval, rewriting, follow-up, and context rebuilding before inventing new workflows.
Microsoft ResearchFinding: a CHI 2025 study of knowledge workers linked higher confidence in generative AI with less reported critical-thinking effort. Critical thinking also shifted toward verification, integration, and supervision.
For us: important AI suggestions need sources, assumptions, and clear approval points.
Lee et al., CHI 2025Finding: CHI 2025 research on daily-assistant agents found that both plan quality and user involvement matter. A plausible plan can still create misplaced trust.
For us: higher-impact actions should use Plan → Preview → Approve → Execute, while low-risk reversible actions can stay lighter.
Plan-Then-Execute, CHI 2025Finding: a 2026 study of productivity agents found lower workload with agentic support, while medium autonomy produced the best balance between productivity and preference for control in that setting.
For us: autonomy should depend on risk, reversibility, and user preference, not one global “agent mode” switch.
Geninatti Cossatin et al., 2026Finding: a 2026 study of AI-generated interface prototypes found positive pragmatic usability but weaker hedonic qualities such as originality and innovation.
For us: AI can help build Wooolfmesh, but it should not decide the visual language by averaging every SaaS dashboard it has seen.
Romero et al., 2026Visual direction
We want the interface to feel like a place to work, not a monitoring console.
#F4F2EC canvas and #FBFAF7 surfaces
reduce the “admin dashboard” feeling and give documents,
screenshots, charts, and code room to carry their own color.
Rule: large areas stay quiet.
#20201E is the main text color. We prefer
near-black to pure black for a softer long-session surface while
keeping strong contrast.
Rule: readability wins over decoration.
#4351C7 indigo is reserved for the main action,
current selection, and a small number of interactive signals.
Rule: if everything is accented, nothing is important.
#2B7555 means success or healthy completion.
#A83A40 means risk, destructive change, or failure.
Rule: never use semantic colors as decoration.
Google reports that strategic use of size, shape, color, and containment helped people find key UI elements faster in Material 3 Expressive studies.
Borrow: make one important action obvious instead of styling every card.
Google Design researchApple's current guidance says Liquid Glass works best for navigation and controls and should not fill the content layer.
Borrow: subtle depth can separate Studio controls from the artifact, but the content itself should remain stable and legible.
Apple HIGAnti-patterns
Some 2026 trends are useful only when used with restraint.
Personalization can rank content and suggest actions. It should not constantly move important controls or rename the user's workspace.
Do not open every task with three screens of generated explanation. Start with the smallest useful intervention.
Transparency, blur, glow, and motion can help separate control layers. Used across content, they become expensive visual noise.
Memory should help a decision or restart. A feed of everything remembered creates another inbox.
Animated “thinking” panels, token counters, and activity graphs are not outcomes. Show progress only when it helps the user steer, trust, or recover the work.
A single score invites optimization of the metric instead of the work. Prefer concrete signals such as completed outcome, restart friction, unresolved blocker, or useful memory reuse.
Research backlog
Small experiments can test the pivot before we build another empire of settings.
Close a Studio for 24 hours. Reopen it and measure how long it takes to identify the current state and next action.
Success signal: useful continuation without searching old chats or notes.
Layer 1: answer or next action. Layer 2: evidence and assumptions. Layer 3: full detail and history.
Success signal: less scrolling with no loss of confidence or task success.
Show at most three retrieved memories by default and explain why each one appeared.
Success signal: users reuse useful context without feeling they must review a memory inbox.
Run different work types through one Studio shell and change only the center artifact and contextual tools.
Success signal: users do not need new navigation knowledge for each workflow.
Save sources into the current Studio with a role such as evidence, reference, input, or result.
Success signal: fewer open tabs and faster return to useful sources.
Automatically allow low-risk reversible local actions. Preview important durable writes. Require explicit approval for destructive actions.
Success signal: fewer interruptions without loss of trust or recoverability.