Maxwell Seal
← Work

Structuring institutional knowledge for a support team

A knowledge base for RNLI's Supporter Experience agents — adopted by the team before it was ever mandated.

Role
Design and build, solo, supervisor-reviewed
Context
In-role initiative, RNLI CX team
Timeframe
Seeds late 2024, built spring 2026, live and iterating
Medium
SharePoint

A live knowledge base the email workstream now reaches for instead of asking a colleague or digging through stale Word documents — four months on, still iterating. It matters because team knowledge used to leave when people did.

Spine · state 1 — the answer path: 09:03, full inbox → confident answer in ~30s · Figma
The whole system in one path: a process question becomes a confident answer in about thirty seconds — structure doing what a training document can't.

It started as notes to my future self

Origin timeline — late 2024 · notes → SharePoint pages → Jack's shop hub → Clare taps Max, spring 2026 → live · Figma
Adoption ran ahead of the mandate: a colleague built his own hub from the example before the email workstream was ever assigned.

The constraints set the shape. SharePoint offers a fixed block palette, and a live page’s name can’t change without breaking every link into it — so names had to be right at launch. One author, a supervisor review gate, few hours once it shipped, and a reader who is an agent at 09:03 with a full inbox and a wide spread of experience across the team.

Names, ownership, and progressive disclosure

Three moves carry it. Ownership first: every page belongs to one hub, and if it’s linked from your hub, you own it — link, never duplicate, so a living system can’t fork into contradictions. Names are outcomes, not labels: filename, heading, and search description are written against the four things agents do — browse, bookmark, search, orient. And disclosure is laddered: a page opens with what’s needed first and holds the rest a layer down, clear of someone moving fast.

IA / hub-ownership map · ghost elements: sub-hubs never built (dashed), duplicates (struck), ~450 legacy templates outside the boundary, no role labels · Figma
If it's linked from your hub, you own it — link, never duplicate. The holes are shaped on purpose: no sub-hubs, no duplication, ~450 legacy templates left outside the curated boundary, no role labels in agent-facing copy.

A working mode, filed as a topic

Fast, informal replies lived under ‘templates’ as if they were a subject — but agents are inthat mode, they don’t browse to it, and Kat missed it unprompted in the first test. The fix became a rule, not a patch: that work gets its own tier, and a token marks an audience or working context, never a topic. A phone agent’s flag extended it team-wide. Heard in February, shipped in June — scarce hours, spent when they freed up.

Before · Feb — path breaks at the naming step
After · Jun — re-routed
February, the path breaks at the naming step; June, re-routed. The IA changed shape in response to what the tests showed.

Removing the correct answer

The path’s destination, up close. A technically-correct nuance about direct-debit cut-offs — true, useful to an expert — was cut, because an agent misreading it mid-run could cancel a supporter’s donation on the wrong day. Baseline confidence beats conditional complexity when the reader is time-pressed and the failure mode is someone’s money.

Page-anatomy zoom — the destination page, cut content ghosted (DD cut-off nuance) · Figma
The destination page with the cut content ghosted. What an agent doesn't see is a design decision.

Not welded to the tools that built it

Built with markdown, Git, and Claude Code for iteration speed. When the organisation standardised on Copilot, the workflow adapted and the pages didn’t notice.

What testing showed

2
Moderated tests
2
Agent types, sampled on purpose
n=2
Directional, counts not percentages
Two agents, two working styles, full transcripts. Named as n=2 and directional — the honesty is the credibility.

One browse-first agent, one search-first — sampled on purpose — with full transcripts and a self-built transcription pipeline (one clause, not a showcase). What pushed back matters as much as what confirmed: one finding became the beat above; another is still an open loop.

Adopted before it was mandated

Adoption preceded the mandate — the pattern replicated before the project was formal, and it’s still in use and still changing. Live metrics are thin by design: no baseline was captured before rollout, and that’s named, not hidden. The frame is a north star with proxies — questions resolved without interrupting a colleague, read through first-resort behaviour, time-to-answer, and where new starters go first.

What it taught

Structure teaches behaviour. The knowledge base works in the moments a page’s name, shape, and omissions make the right action obvious to someone with no time to hunt. Everything else serves that.