Shared skill (agent-compounds). Symlinked into projects via deploy.sh — this is
the single source of truth; edit here, not in a consumer copy. Method-only and
portable (no project facts).

wiki — the synthesis layer

Purpose: write, update, and garden the org's synthesis pages — the "wiki" in the
Karpathy-pattern sense: prose that integrates many atomic facts/decisions into one
coherent, navigable narrative. Not a new source of truth; a derived, regenerable
cache over the substrate that already exists.
Constitution: ../context-engineering/SKILL.md (load it first — WRITE taxonomy,
homes, PLACEMENT/ALTITUDE all come from there; this skill is the synthesis type's
method).
Evidence base: references/research-basis.md — distills the field research + our
internal audit; read before seeding pages or designing a garden pass.
Status: live — schema + 7 canonical pages in <org>/wiki/ (ratified 2026-07-19
after adversarial citation review); provenance: <global>/plans/memory-wiki-upgrade.md Phase 3

---

When to Use This Skill

Intent Triggers:

• Drafting a new synthesis/concept/entity/topic/contradiction page
• Updating an existing wiki page with new source material
• Running a dedupe/merge/prune/staleness pass over wiki/
• The weekly distillation cadence (STRATEGY.md decisions log, NOW.md refresh)
• Deciding whether new material becomes a new page or an update to an existing one

When NOT to Use:

• Saving a single fact, rule, decision, or recipe → context-engineering routes it
• End-of-session capture → reflect
• Reading/browsing pages (no skill needed — index-first: open <domain>/wiki/index.md and follow links; qmd query when no page covers it; the hook auto-injects matching pages)

---

Page types

Type	It is…	Example
concept	an idea/pattern that recurs across facts/decisions	"the flywheel", "low time preference"
entity	a named thing with a lifecycle (product, agent, system)	"the example-app product thesis", "the agent org"
topic	a bounded subject area integrating several concepts/entities	"the content engine", "health protocol stack"
contradiction	two or more sourced claims that disagree, recorded not hidden	"one app's branch doctrine vs another's"

One canonical page per concept — kebab-case filename, <domain>/wiki/<slug>.md
(<org>/wiki/ is the first and only live domain; other domains get one only when
they earn it, per context-engineering's "no slot → flag a taxonomy bug, never invent a
folder" rule).

Frontmatter contract

---
title: <human-readable title>
type: concept | entity | topic | contradiction
sources: [<fact/decision/doc slugs this page distills>]
related: [<other wiki page slugs>]
created: <YYYY-MM-DD>
updated: <YYYY-MM-DD>
confidence: high | medium | low
trigger: <when should an agent load this page — the retrieval-trigger description>
status: draft | canonical
---

trigger: is not decorative — it's Devin's structural idea (see
references/research-basis.md): author when this should surface explicitly, don't
make the agent infer relevance post hoc. Write it the way you'd write a skill's
description: trigger clause. How it's consumed: the wiki collection is one of
the memory-retrieval hook's lobes, so the trigger text is part of the page's hybrid
search/embedding surface — a well-written trigger literally improves when the page
auto-injects. memory-lint enforces its presence.

Page structure: Compiled Truth + Timeline

Every page body has two layers below the frontmatter, split by a
--- <!-- timeline --> divider line:

• Compiled Truth (above the divider) — the current synthesis, freely rewritable,
  same regenerability mindset as the rest of this doctrine. This is what an injected
  page shows.
• Timeline (below the divider) — append-only, dated, one-paragraph entries, newest
  first. Never edited or deleted — a correction is a new dated entry, not a rewrite
  of an old one.

Garden/distill passes update Compiled Truth prose FROM the timeline (what actually
happened), never the reverse. Any material edit to a ratified page's Compiled Truth
requires appending a timeline entry recording what changed and why, same session.
A memory-lint script does not enforce this structure (schema/index/wikilink checks
only, if you have one — <your-deployment>/scripts/health/memory-lint.py) — this
checklist is the enforcement layer; see Common Mistakes below.

THE CITATION RULE

Every claim in a wiki page carries a [[wikilink]] to the fact, decision, or source
it comes from. Pages are DERIVED views, never a source of truth — the atomic facts in
the domain memory root (<memory-root>), the decisions in alignment/decisions/, and the raw sources in
references//knowledge/ remain canonical. A sentence with no citation trail is not
synthesis, it's a claim with nowhere to verify it — cut it or find its source first.

[[wikilinks]] resolve cross-home by design: a link may target another wiki page OR
a memory-fact slug in any substrate home (memory-lint's check_wikilinks enforces both).
Linking a fact by slug is the normal citation form — do not create stub wiki pages just
to satisfy a link.

This is the single most validated mechanism in the field evidence (GitHub Copilot's
production A/B test: citation-verified memory beat uncited memory on every measured
metric, p<0.00001 — see references/research-basis.md). It is also the guard against
hallucination laundering: an agent re-reading its own uncited synthesis with full
confidence is how a wiki quietly becomes fiction.

Create-vs-update judgment

Dedupe-over-append at the page level, same as facts. Before drafting a new page:
qmd search/grep the wiki/ dir for an existing page on the same concept. Found one
→ update it (new sources, revised claims, updated: bump) — never create a
near-duplicate "v2" page. Genuinely new concept, no existing page covers it → create.
One canonical page per concept is the invariant a garden pass exists to restore
when this judgment call was missed.

The authority chain

raw sources (references/, knowledge/, research reports)
        │ never edited by agents — read, cited, immutable
        ▼
canonical facts + decisions (the memory homes, alignment/decisions/)
        │ append-only claims — the ground truth
        ▼
synthesis pages (wiki/)
        │ derived narrative — regenerable, never precious

Agents may only write/edit the bottom layer. A wiki page claiming something not
traceable up this chain is a schema violation, not a stylistic nitpick.

Regenerability mindset

A wiki page is a rebuildable cache, not an asset to protect. If a page drifts,
duplicates, or goes stale, the fix is to regenerate it from its cited sources — never to
patch around a page's accumulated cruft to preserve its prose. The structure (the
taxonomy, the citation discipline) is the asset; any individual page is disposable.
Treat a "this page took effort to write, don't rewrite it" instinct as the anti-pattern
it is (references/research-basis.md: ouroboros compression, ingest-everything
automation).

The operator reviews before canonical

New pages and material edits land status: draft. A page only becomes status: canonical after the operator reviews it — flip the frontmatter field, nothing else changes.
Draft pages are still linkable and citable (a page in progress is still useful), but a
draft's claims carry lower trust than a canonical one until reviewed. Never self-flip
canonical — that's the self-certification anti-pattern
([[no-self-certifying-remedy]]).

Anti-patterns to refuse

Anti-pattern	Why it's refused
Mirroring operational state (status, logs, beads, health reports)	Pages point at live state, never copy it — copies go stale immediately (memory-wiki-upgrade decision log #4)
An uncited claim	Breaks THE CITATION RULE — find the source or cut the sentence
A near-duplicate page for the same concept	Update the existing page instead (create-vs-update judgment)
Scope creep ("all of my knowledge" pages)	Most-cited death in the field research — bounded ontology, one domain at a time
Precious/protected pages resistant to regeneration	Violates the regenerability mindset — pages are cache
Self-flipping status: canonical	The operator reviews every page before it's canonical
Building a browsing UI before the substrate is trustworthy	Obsidian is the browsing surface; no rendered site (not-build list)

---

Workflows

Workflow	Use when
workflows/seed-page.md	Drafting a brand-new synthesis page from cited sources
workflows/garden.md	Periodic dedupe/merge/prune + contradiction-page creation + staleness sweep
workflows/distill-weekly.md	Weekly review's distillation cadence — STRATEGY.md decisions log + NOW.md refresh
workflows/hallucination-audit.md	Monthly cadence — sample 5 random wiki claims, trace to cited sources, untraceable → proposal beads

Reference Documentation

File	Contents
references/research-basis.md	Distilled field research + internal audit evidence — read before seeding pages or designing a garden pass

---

Common Mistakes

Mistake	Fix
Writing a wiki page as if it were a new source of truth	It's a derived view — every claim traces up the authority chain
Skipping qmd search before drafting	Dedupe-over-append — update the existing page
Omitting trigger:	Write the retrieval-trigger clause, same discipline as a skill description
Flipping status: canonical yourself	The operator reviews first
Mirroring a beads/status dashboard into a page	Point at it, never copy it
Treating a stale page as precious	Regenerate from cited sources; pages are cache
Editing or deleting a Timeline entry	Timeline is append-only — correct by appending a new dated entry
Rewriting Compiled Truth without a Timeline entry	Every material edit needs a same-session timeline entry recording what/why