Wiki Qa

Wiki Q&A — graphify graph-based retrieval

Answer the team's wiki/planning questions without guessing, grounded in the actual documents. The key is to not read everything: use the graphify graph first to narrow down "where to read" — token-efficient, and scales as the wiki grows.

Principles (invariant)

• No answer without evidence. Quote only what actually exists in raw/wiki. No

guessing, no invention — if it isn't there, answer "not there yet / undecided".

• Cite sources. State which files the answer came from

(e.g. (sources: wiki/concepts/billing-model.md, raw/PRD.md)).

• This channel = this project's lore vault only. Never pull in concepts from other

projects.

• related/[[ ]] are graphify-derived views — never edit them directly (read only).

Retrieval procedure

Working directory is the project lore vault root (where graphify-out/, wiki/, raw/ are visible).

1. Narrow the concepts — query graphify first:

• Free-form query: graphify query "" --graph graphify-out/graph.json

→ returns the start node plus a BFS-connected subgraph of related nodes, edges, and source file paths. Query with the question's key concept term as it is worded in the wiki (e.g. "billing model", "payment scope").

• Detail on one concept: graphify explain "" → that node's neighbors

and relations (depends_on/references etc.).

• Relation/ripple between two concepts: graphify path "" "" → shortest path.
• If nothing matches (label mismatch, synonym, typo): look up the related

concept name in wiki/index.md (the concept catalog, one-line summaries) and re-query with that name, or fall back to search_files (grep). Never force a guessed answer.

1. Read the evidence (only what was narrowed): read_file only the

wiki/concepts/.md files that query/explain pointed to. If decision history or original wording is needed, also the relevant parts of raw/PRD.md and raw/decisions/*.md. Do not read whole files the graph did not point to.

1. Answer: answer concisely from what you read and attach the sources. Add

one line of graph-connected concepts (e.g. Related: payment scope, free-plan policy) — surfacing connections even when unasked is graphify's value.

Fallbacks / limits

• If graph.json is missing (not yet built): answer via wiki/index.md + direct

read_file, and if the wiki looks like it needs refinement, point to refine-wiki.

• graphify query matches on node names, so it is weak on synonyms and spelling

→ on a miss, fall back to index.md/search_files; if still nothing, honestly answer "no evidence in the wiki".

• If the question is a change request rather than a lookup (planning edit,

decision, backlog): do not answer it — hand off to the relevant skill (resolve-conflicts/backlog-update etc.) and its approval gate. Q&A is read-only.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: loremaster-ai
• Source: loremaster-ai/loremaster
• License: MIT

Install and usage instructions live in the source repository linked above.