analyze-network
/analyze-network
Run a rigorous structural analysis of the loaded Gephi graph, compare centralities, characterize communities, and report what the data does and does not support. Use for comprehensive network analysis or interpretation rather than one claim or a styling-only task.
- Author
- mattartzanthro ↗
- Repository
- gephi-ai ↗
- Licence
- Apache-2.0
- Last repository change
- 12.10.2026
Assessment and security
- Files in this package
- 1
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 12.10.2026, without references to retired models or APIs.
Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.
What it does and when to use it
Run a rigorous structural analysis of the loaded Gephi graph, compare centralities, characterize communities, and report what the data does and does not support. Use for comprehensive network analysis or interpretation rather than one claim or a styling-only task.
Original author description
Run a rigorous structural analysis of the loaded Gephi graph, compare centralities, characterize communities, and report what the data does and does not support. Use for comprehensive network analysis or interpretation rather than one claim or a styling-only task.
How it works
- Rules this workflow must keep
- Steps
- Reading pass (after the numbers)
- Boundaries
SKILL.md preview
---
name: analyze-network
description: Run a rigorous structural analysis of the loaded Gephi graph, compare centralities, characterize communities, and report what the data does and does not support. Use for comprehensive network analysis or interpretation rather than one claim or a styling-only task.
---
# Comprehensive Network Analysis
Run a full structural analysis of the current graph and present a detailed report of its properties.
Follow the `gephi` skill's rules throughout; this workflow repeats only the ones
it is most likely to break. Read `../gephi/references/statistics-guide.md`,
`../gephi/references/reading-network-maps.md`, and
`../gephi/references/claim-verification.md` as needed. These references are
authoritative for interpretation; do not substitute a remembered rule of thumb.
**Tell the user what you are doing at each step.** Narrate briefly before each tool call (for example, "Computing modularity...", "Running centrality analysis...").
## Rules this workflow must keep
- **Asking.** When a step says to ask, ask once. If the person cannot answer, is away, or asked for a finished product, use the default named in that step and remove or overwrite nothing. At the end, give one short line for each choice that changed the result, under "Choices I made"; leave the section out when every step used its obvious default. If there is no input to work on, stop and say what is needed.
- **Session start.** Start with `gephi_health_check`. Then check which workspace is open (`gephi_list_workspaces`) and whether a filter is active (`filter_active` in replies): a filter from an earlier conversation stays on, and exports and checks then see only what it shows.
- **Ranking.** To rank nodes on a metric, call `gephi_query_nodes` with `sort_by` set to the metric, `columns` set to the few columns you need, and `limit` about 10: nodes are ordered before paging, so the first page is the top of the whole graph. Without `columns`, every attribute of every node comes back and a long listing can overflow.
- **Stability.** Run `gephi_community_stability` before naming, captioning, or colouring by groups, and say how stable they are.
- **No scale-free label.** Never call a heavy-tailed degree distribution "scale-free" or a "power law": those fits are near-indistinguishable from log-normal in practice and smuggle in a universal-law claim (Jacomy 2020; Broido and Clauset 2019). Describe hub dominance as a property of this network.
## Steps
1. **Session start**: Call `gephi_health_check`. If it fails, tell the user to start Gephi and stop. Then call `gephi_list_workspaces` and note `filter_active` in the replies. If a filter is on, say so before any number is reported, and ask whether to analyze the filtered view or the whole graph (`gephi_reset_filters` shows every node again). Default: leave the filter on, analyze what is visible, and say in the report that a filter was on.
2. **The intake question** (skip if already answered in this conversation):
ask in one sentence what the nodes and ties are and what they want to
learn. Use their vocabulary in the whole report, and treat their
expectations as hypotheses the analysis will confirm or contradict.
Default: use the graph's own column names as the vocabulary and report
the structure without prior hypotheses.
3. **Profile first**: Call `gephi_profile_graph` (one call: size, density,
degree distribution, components, isolates, weights, modularity,
clustering, flags). Open the report with a plain-language first reading
that combines their description with these numbers, and let the profile
decide which deeper analyses are worth running rather than running
everything.
4. **Degree distribution**: Read it from the profile (minimum, maximum, average, and whether it is heavy-tailed, meaning a few high-degree hubs, or even). Do not query every node for it. Call `gephi_compute_degree` so the `degree` column exists for ranking in step 9.
5. **Community structure**: Call `gephi_compute_modularity` with resolution 1.0, then `gephi_community_stability` to check that the communities hold up. The modularity score alone never shows strong communities.
6. **Path analysis**: Call `gephi_compute_avg_path_length` to get average path length, diameter, and radius. When the user asks about two particular nodes, `gephi_find_shortest_path` gives the path between them and how many equally short paths exist.
7. **Clustering**: Call `gephi_compute_clustering_coefficient` to measure local cohesion (how often a node's neighbours are tied to each other).
8. **Centrality**: Call `gephi_compute_betweenness` and `gephi_compute_pagerank`.
9. **Key nodes**: Rank the top nodes on `degree`, `betweenesscentrality`, and `pageranks` with the ranking rule above: one `gephi_query_nodes` call per metric, `column` and `min` set, `limit` 20 or less.
10. **Report**: Present a structured summary:
### Network Overview
- Nodes, edges, density, average degree, graph type
### Connectivity
- Number of components, size of giant component
### Community Structure
- Number of communities, `mean_stability` in plain words, the stable cores, and the modularity score with that context
### Small-World Properties
- Average path length, clustering coefficient, comparison with random network expectations
### Key Nodes
- Top 5 by degree, betweenness, and PageRank
### Structural character
- Describe the network's structure in this-network terms: hub dominance (a few
nodes concentrate ties) or even degree; clustering and path length relative to
size (small-world-*like*, stated as a comparison, not a label); fragmentation.
No universal-law labels such as "scale-free" (see the rules above).
### Choices I made (only if a choice changed the result)
- One short line per choice that changed the result.
## Reading pass (after the numbers)
Once statistics are computed and a layout exists, walk the guided reading
process from `../gephi/references/reading-network-maps.md`: state the reading rules
(axes are arbitrary, only distances matter), identify main clusters with
temporary letter names, point out the structural holes and what they imply,
overlay one attribute at a time and compare its distribution to the structure,
inventory the special nodes (bridges via betweenness, within-cluster hubs via
degree, off-color outliers), and only then name clusters in the person's own
vocabulary, after the stability check. Present insights as hypotheses or
findings and say which.
Before delivering YOUR reading of a fresh map (and again after an attribute
overlay changes it), ask one concrete question first ("where does your eye go
first?" or "which groups look connected to you?"), then give your reading and
compare the two aloud. Their unprimed look is evidence your fluency would
otherwise erase; the differences between the readings are often the finding.
Skip this on task turns, and drop it for the session if they wave it off.
Default: give your reading directly.
## Boundaries
- Interpret; do not recolor, relayout, filter, or edit the graph.
- Compare metrics rather than interpreting one in isolation.
- Treat every structural assertion as a claim to check, and pair patterns with
a rival explanation when the graph cannot distinguish causes.
- Never call a heavy-tailed network scale-free or claim a power law.
Versions
The database has no other skills named analyze-network.
Database snapshot: 13.11.2026 · package version ec4b0b455071 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 90explore-network
Explore the graph already open in Gephi through intake, profiling, goal-directed analysis, styling, layout, and an overview. Use when the user has a graph loaded and wants an initial guided exploration without importing a file. - 90verify-claim
Independently verify one plain-language structural claim about the loaded Gephi graph and report confirmed, refuted, or cannot tell with measurements and live-graph receipts. Use for claims about centrality, connectivity, comparison, grouping, or robustness. - 100networkx
Create, analyze, and visualize complex networks and graphs in Python with NetworkX. Use when working with network/graph data structures, computing graph algorithms (shortest paths, centrality, clustering), detecting communities, generating synthetic networks (random, scale-free,… - 100graph-the-network
Build an entity-relationship link-analysis graph of an investigation — nodes, typed edges carrying source and confidence, aliases, and temporal validity — to expose shared infrastructure, bridging nodes and the real principal behind a frontman. Covers Maltego, Neo4j and Cypher,…