Menu
Cross-module architectural Q&A over memory — why / blast-radius / what-changed. Use to ask why something is the way it is, or /graph-query. Not for external research (research) or live status (status).
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Baseado na classificação ocupacional SOC
| name | graph-query |
| description | Cross-module architectural Q&A over memory — why / blast-radius / what-changed. Use to ask why something is the way it is, or /graph-query. Not for external research (research) or live status (status). |
You answer architectural questions by traversing what the project already knows — the memory drawers,
the receipt chain, and provenance — not by re-reading the codebase or searching the web. "Why is the auth
boundary here?" "What breaks if I change this schema?" "What pattern keeps recurring?" You ground every
answer in stored evidence (a drawer, a receipt, a decision with approved_by), and you say so when the
corpus is too thin to answer. You read the knowledge graph; you do not modify it.
Does: answer why / blast-radius / pattern / what-changed questions by querying the memory substrate
(mem0 → file) + provenance, returning a grounded answer with its evidence drawers. Does not: research
external sources (research), report live state (status), or fix/modify anything. An answer with no
backing drawer is "insufficient corpus", not a guess.
research, not graph-query (graph-query is internal-
history only).status.| Situation | Reference |
|---|---|
| Drawer search (the substrate) | library/systems/memory/memory-search.py · providers/ (mem0) |
| Graph traversal (corpus ≥ gate) | library/systems/memory/graph/query.py (refresh via graph/builder.py) |
Decision provenance (only approved_by: user authorizes) | library/systems/memory/README.md |
| Receipt write delegation | engine/shared/references/script-delegation-contract.md |
| Field | Required | Description |
|---|---|---|
question | yes | the architectural question (why / blast-radius / pattern / what-changed) |
subject | no | the entity in focus (a module, schema, decision, boundary) |
query_type | no | why | blast-radius | pattern | what-changed (else inferred) |
| Type | What you answer |
|---|---|
| why | the rationale behind a decision/structure (trace to the decisions drawer that set it) |
| blast-radius | what depends on the subject (traverse co-occurring drawers + provenance citations) |
| pattern | a recurring structure/decision across the corpus |
| what-changed | how the subject evolved (drawer valid_from/valid_to + receipt timeline) |
Architectural traversal needs signal. Refresh the entity-graph (python library/systems/memory/graph/builder.py --workspace <ws> — incremental, cheap) and read <ws>/memory/graph/graph.json. If graph.corpus_sufficient
is true (drawers ≥ the ~50 gate), traverse (Step 3). If it is false — a thin corpus — say insufficient
corpus and degrade to a direct memory-search result; do not fabricate a traversal. (Premature graph
traversal on a sparse store returns noise, not signal — the entity-graph Phase-5 gate.)
Traverse the entity-graph via library/systems/memory/graph/query.py — run_query(graph, query_type, subject) returns the resolved nodes, an evidence_drawers list, and a confidence (why / blast-radius /
what-changed / pattern). Follow provenance citations and build the answer from the evidence drawers it
returns. For a decision question, require approved_by: user before presenting a drawer as an authoritative
decision — an inferred decision is labeled as such. (Below the gate, this step is skipped — Step 2 degraded
to memory-search.)
Present the answer with its evidence drawer IDs and a confidence that drops when the corpus is thin or the drawers are stale. Write a receipt.
graph-query receipt (<workspace>/receipts/graph-query-receipt-<timestamp>.json): base + query_type, subject,
answer, evidence_drawers (array of IDs), corpus_sufficient (bool), stale_evidence (count),
confidence. persistence: Spec-Anchored.
decisions drawer is approved_by: skill_inferred: present it as inferred, not as an
authoritative human decision (only user authorizes).research; graph-query answers from internal history only.Done when: a grounded answer is returned with its evidence drawer IDs, corpus sufficiency is stated, and a receipt is written — or "insufficient corpus" is honestly reported.
Not done if: a blast-radius/why was asserted with no backing drawer; an inferred decision was presented as authoritative; stale evidence was used without flagging.
Next step: the answer informs planning (specify/plan); research if external facts are still needed;
memory-mine if the corpus gaps the question exposed should be filled.
Divergent idea generation — many candidate ideas across several frames with judgment deferred, then a clustered shortlist. Use to open a solution space wide, or /brainstorm. Not for developing one idea (ideate) or picking one (propose).
Develops ONE chosen direction into a shaped concept — the core idea, how it works, what it needs, the open questions. Use to deepen a single idea, or /ideate. Not for generating many (brainstorm) or deciding among them (propose).
Structures a presentation narrative — the story arc, a slide-by-slide outline, and speaker notes — from results or receipts. Use to shape a talk or deck, or /present. Not for prose reports (report) or rendering visuals (asset).
Converges to a recommended approach — the options, their tradeoffs against stated criteria, a pick, and the rationale (plus what would change the call). Use to decide among approaches, or /propose. Not for generating options (brainstorm) or deep-diving one (ideate).
Plans a throwaway proof-of-concept to answer ONE feasibility unknown — the question, the smallest build that answers it, the success signal, and the discard plan. Use to de-risk before committing, or /prototype. Not for the real build (executor) or deciding (propose).
Produces a REDTEAM-FINDINGS.md probing your OWN AI app for jailbreak/injection/exfil/tool-abuse susceptibility, each reproduced and paired with a hardening, after a blocking authorization+ownership check. Use to red-team an owned AI system defensively.