AIppocampus UX

Overview

Use this skill to make AIppocampus usable by a foreground agent on behalf of a human. The bar is not generic visual polish: AIppocampus UX should be clear like Linear, fast like Raycast, and integrable/recoverable like Stripe while preserving source-backed boundaries.

This is a product/agent-use skill, not a design-director critique. Optimize for the next agent being able to decide, pull the right source route, act, recover, and leave feedback without reading audit-scale internals.

For broad reviews, issue triage, or new foreground contracts, read references/agent-facing-ux-charter.md. For a tiny copy, CLI, or card fix, use the workflow below directly.

Workflow

1. Reopen the task source.

Use the current issue, PR, comments, docs, code, or CLI output before judging the surface. If older AIppocampus context may affect the work, use the normal AIppocampus orientation/recall route first. Treat route packets as navigation until source is opened.

1. Classify the surface.

Pick one primary class before recommending fields or tests:

• foreground_agent_action: compact route, action card, hook hint, MCP/CLI

result, or recovery card used by an active agent.

• operator_debug: diagnostics, red lines, policy ledgers, run metadata, or

private local detail for maintainers.

• public_demo: public-safe docs, examples, screenshots, or first-run proof.
• source_court: deepen/source-open output used for exact, public, stale,

sensitive, disputed, or high-risk claims.

• benchmark_report: evaluation, readiness, or evidence reports.
• setup_onboarding: install, update, hook, provider, sync, or readiness

flows.

1. Map the agent journey.

Use the Access / Context / Tools / Orchestration map:

• Access: can the agent tell whether AIppocampus is installed, current, and

callable in this host?

• Context: does the agent receive a small useful continuity lead, not

audit-scale JSON?

• Tools: are recall, deepen, explain, feedback, setup, and recovery paths

discoverable with clear scopes?

• Orchestration: can a fresh agent complete recall -> deepen/source -> act

-> feedback without inventing broad manual search first?

1. Review against the four product standards.

• Clear: one primary next action, one handle vocabulary, action card before

audit payload, no visually interchangeable display ids and callable handles.

• Fast: first useful packet appears early, foreground fields stay small,

copy-paste commands exist where useful, no broad search when a route already exists.

• Integrable: each tool/card answers What / When / After, schemas are

task-first, host boundaries are explicit, public demos use stable commands.

• Recoverable: missing args, stale handles, blocked source, missing setup,

no active contract, or permission/config problems return one executable recovery action or chooser card.

1. Preserve source-backed usefulness.

Route usefulness is not a claim shortcut. A route can be actionable before it is claim-ready; exact/public/stale/sensitive/high-risk claims still need deepen/source reopen. If a safe useful route exists but the surface demotes it to vague scent or silence, count that as UX debt rather than safety success.

1. Recommend the smallest durable slice.

Prefer one repair that improves the agent journey across sibling surfaces. If the smell is recurring across CLI/MCP/cards, point to a shared contract or helper. If it is isolated, keep the patch local. Do not introduce a broad schema, scoring layer, or governance abstraction unless it removes repeated product debt.

Output Shape

Start with a direct verdict, then the smallest useful next action.

Use this structure for reviews or plans:

• Verdict: ship, fix_one_slice, needs_contract, or blocked.
• Surface class: one of the classes above.
• Primary next action: one command, patch target, issue update, or source route.
• Findings: 3-5 items max, ordered by user impact, with file/line/issue refs

when available.

• Fix slice: exact files/docs/tests to change; name non-goals.
• Verification: behavior checks first, then JSON/schema/docs checks.
• Unknowns: mark with [⚠️] and say how to confirm.

For implementation, use normal codebase patterns. Add comments only where a fallback, gate, threshold, field projection, or recovery boundary prevents a future agent from making an attractive but unsafe simplification.

Do Not

• Do not run a generic visual critique unless there is an actual UI visual.
• Do not use Nielsen-style scoring as the main answer for CLI/MCP/agent cards.
• Do not turn AIppocampus UX into marketing copy.
• Do not require foreground agents to understand full provenance, benchmark

ledgers, or internal policy machinery before acting.

• Do not weaken source reopen, privacy, high-risk, or local-path boundaries to

make the foreground feel simpler.