Skip to main content
/synclair/knowledge is Synclair’s answer to the question “where does the AI get its context?” It’s the home for everything a person or agent needs to know about the product that isn’t in the code — the specs, the designs, the decisions — organized so it’s cheap to load and impossible to lose.

​
What it is

The place the project’s knowledge lives, in three layers that trade off freshness against readiness:

Sources

Links to the canonical PRDs, decks, and Figma files — wherever they actually live (Drive, Figma, GitHub).

Digests

The compressed, durable knowledge extracted from those sources and kept in the repo as skills — what agents read most of the time.

Summaries

Generated, audience-specific reads — onboarding briefs, diagrams, custom cuts — on the Summary tab.

​
What’s in it

  • Source-of-truth links (lib/system/knowledge/sources.ts) — each entry points at a canonical doc (a PRD in Drive, a Figma file, a deck) with enough metadata to find it and know what it covers. In existing-app mode these are harvested from the host repo’s READMEs, docs/, and ADRs.
  • Distilled digests — the durable essentials pulled out of those sources, stored as the product-spec skill (and per-area references/*.md). Each digest links back to the canonical source it came from.
  • Project summaries (data/knowledge/summaries/) — briefs and diagrams generated per audience, shown on the Summary tab.

​
How it works

The core rule is link, don’t copy — and it’s not a style preference, it’s what keeps the knowledge from rotting.
A pasted 40-page PRD goes stale the day someone edits the real Doc, and it blows an agent’s context window when read raw. So raw sources stay canonical in Drive / Figma / GitHub and are linked; the repo holds the distilled digest, not the dump.
Two mechanisms make rich context cheap:
1

Progressive disclosure

A digest’s one-line description is always in context; its body loads only when a task touches that area. A large knowledge base costs almost nothing until it’s relevant.
2

Digger agents

A retrieval agent (like prd-retriever or figma-frame-reader) reads the heavy source in its own throwaway context and hands back a tight brief — so the main conversation never loads the 40-page doc.
And a contract keeps it from decaying into a link dump: linking a source and distilling it are one act, not two. Adding a manifest entry includes writing its digest in the same pass — an entry with no digest is unfinished work, and the only exception is an explicit TODO: distill note when the source genuinely can’t be read yet. Write-back is the second half of the same flywheel: when a later dig surfaces something durable, the agent refines the existing digest — so the knowledge base converges on “what a builder routinely needs.” And because the sources keep moving where they live, npm run check:knowledge probes each linked source’s real last-modified against its distilledAt and flags digests whose upstream has moved — so a PRD edited in GitHub can’t leave its digest silently lying. It’s the knowledge-layer twin of the foundation’s call-home check.

​
How to act on it

Dig into the raw source with a digger (prd-retriever, figma-frame-reader) only when the digest is insufficient — then write the durable part back so the digest gets sharper.
Queue a summary from the Summary tab (or run the product-summary skill) — a brief, a diagram, or a custom cut for a specific audience. The button queues it; an agent drains the queue and writes it to data/knowledge/summaries/.
Point them at /synclair/knowledge — the Summary tab’s brief is the fastest way in, and the source links are there when they need the canonical detail.

Related: References

Knowledge is the inputs (specs, designs). References is the project’s findings — prior art and recommendations agents accumulate as they research.