TeamShared
Start free

Mode B — the UI review skill that keeps shrinking PR findings

Most teams paste a UI checklist into a chat once and watch it go stale. The next agent never saw the last pass. The same hierarchy bugs show up on the next PR. Humans re-explain the bar every time.

Mode B is TeamShared’s answer inside skill teamshared.ui-design. It evolved organically from a review playbook into a recursive coding-review loop: score the PR, write a findings table, fix, re-review — until Critical and High findings are gone.

Chad’s shorthand for that loop is Plan B. The canonical name in the catalog is Mode B on teamshared.ui-design.

The problem this week

Frontend PR review with agents usually fails in three places:

  1. The design bar lives outside the agent’s working context.
  2. A first pass leaves Highs open; a second agent starts cold and re-discovers half of them.
  3. Nothing durable records how the team closes a class of UI findings, so the org does not get sharper between PRs.

You get comment volume without a shrinking open set.

What Mode B is

Skill teamshared.ui-design has two modes. Mode A creates or redesigns. Mode B is PR / design review.

Mode B’s loop:

  1. Scope the UI path from title, body, and diff — skip if there is no visual surface.
  2. Run a 0.5s focal-point test.
  3. Score against three pillars: Frictionless (task, ≤3 steps, one primary, clear exit), Craft (tokens, hierarchy, motion, editorial density), Trust (actionable errors, inviting empty states, labeled AI content).
  4. Walk the category pass: hierarchy → spacing → type → colour → states → motion → responsive/touch → a11y → flow → interaction continuity.
  5. When layout or canvas matters, drive a preview — don’t guess full-bleed bugs from CSS alone.
  6. Output one findings table:
Severity Before After Why

Severity: Critical = broken use / a11y; High = hierarchy / spacing / primary damage; Medium = polish; Low = nit. End with Top 3 fixes. Comment on the PR. Never merge.

Then the implementer clears the table. The next Mode B pass only has to prove what is still open. That recursion — review → fix → re-review — is how issues discovered on reviews keep getting reduced instead of accumulating as noise.

Standing UI PR review routines in TeamShared run this skill so the loop is not optional folklore. It is the house process.

Why memory injection matters

TeamShared injects memory and skills into agent prompts (SessionStart / soul + playbook + skill). The next seat does not need a human to re-paste the UI review playbook. Mode B travels with the org.

So org memory and procedural intelligence evolve organically: a playbook becomes a skill; the skill lands in every relevant PR wake; each cleared High teaches the next pass what still matters. Skill text can still improve through memory_skill_feedback proposed rewrites (human accept → version bump) — evolution with a gate, not silent mutation.

Who this is for

Teams shipping console or product UI with agents on the PR path — especially anyone tired of re-explaining “one primary,” full-bleed vs document measure, or token discipline on every review.

Try it

  1. Connect TeamShared: teamshared.com → Connect / docs install.
  2. Fetch skill teamshared.ui-design (memory_skill_get).
  3. On the next frontend PR, run Mode B: findings table → Top 3 fixes → fix → re-review until Critical/High are clear.

Plugin install (interim marketplace path): add https://github.com/teamshared-ai/teamshared-plugin, then /add-plugin teamshared, then Tools & MCP → Connect.

Ship notes (for the curious)

All posts