Concept Mapping

Quick facts

Mechanism: Builds a concept map - a non-hierarchical network of concept nodes joined by directed, labeled linking phrases so each node-link-node reads as an explicit proposition, with cross-links across clusters - then surfaces gaps, missing links, and questionable propositions.

Produces: concept-map

Evidence: M/P · Domain: Synthesis

Use when a domain has many interrelated concepts and the goal is to externalize and inspect how they relate, forcing every relationship to be named rather than left as a vague association.

When a domain is described in prose or sketched as a diagram, the relationships between its concepts stay vague: a line is drawn between two boxes, or two ideas are called “related”, and how they actually relate is never specified. Concept mapping refuses the unlabeled link. It builds a non-hierarchical semantic network in which every connection is a directed, labeled linking phrase, so each node-link-node triple reads as an explicit proposition (“free tier - increases - signup volume”), and clusters are joined by cross-links that connect concepts across different parts of the map. The load-bearing move is forcing every relationship to be named, which externalizes how the domain interrelates and makes gaps, missing links, and false propositions visible. The output is a concept map plus a list of surfaced gaps. It externalizes and inspects how concepts relate; it does not claim to improve learning, retention, or decisions.

When to Use

• A domain has many interrelated concepts and the goal is to make how they relate explicit and inspectable.
• You suspect hidden gaps or misconceptions in how a space is understood and want them surfaced as checkable propositions.
• Integration across sub-areas matters, so the cross-links between clusters carry the value (for example linking a pricing concept to a support-cost concept).

When NOT to Use

• To evaluate whether one argument or recommendation is sound - use argument-mapping. Both produce “maps” and this is the easiest confusion: argument mapping has a claim, reasons, co-premises, and objections and judges soundness; a concept map is a network of propositions and judges nothing.
• To decompose one big question top-down into MECE parts - use issue-tree. A concept map is a non-hierarchical network, not a decomposition tree, and does not aim for mutual exclusivity.
• To cluster many raw notes bottom-up with no named relationships - use affinity-mapping (the KJ method). Affinity mapping groups items into themes and deliberately does not name the relationship between them; if you only need themes, not propositions, use it.
• To move a problem down fixed event / pattern / structure / mental-model levels - use iceberg-model. The iceberg has prescribed causal levels; a concept map has none.
• If you drop the labeled-link / proposition constraint, you are doing free-association mind-mapping with unlabeled branches (the Buzan method), which this library excludes (X-tier; Farrand 2002). The named-relationship discipline is the skill; without it this collapses into the excluded method.

Instructions

When asked to map how the concepts in a domain relate, follow these steps:

1. List the concepts. Pull out the key concept terms in the domain (nouns / noun phrases). Aim for a focused set, not everything.
2. Connect with labeled, directed links. For each genuine relationship, draw a directed link and name it with a linking phrase (a verb or short phrase: “causes”, “is a type of”, “constrains”, “increases”, “depends on”, “trades off against”). Never leave a link unlabeled. Each link must make source - link - target read as a true sentence.
3. Read every link back as a proposition. Write the node-link-node triple out as a sentence. If the sentence is vague, false, or one you cannot defend, fix the label or cut the link. This is where misconceptions surface.
4. Add cross-links across clusters. Find concepts in different clusters that should be connected and add labeled cross-links between them. These integrative links are the highest-value part of the map.
5. Surface gaps and missing links. Flag concepts with few or no links (under-connected), pairs of concepts that obviously relate but were never linked (missing links), and any propositions that look questionable.
6. Emit the concept map per references/TEMPLATE.md: the labeled-proposition network, the cross-links, and the explicit gaps / missing-links / questionable-propositions list.

Output Format

Use the template in references/TEMPLATE.md. The deliverable is the network of labeled propositions plus cross-links and the surfaced-gaps list, not prose.

Quality Checklist

Before finalizing, verify:

• Every link is a named, directed linking phrase; there are no unlabeled associations.
• Each node-link-node triple reads as a complete, defensible proposition (a sentence).
• At least one cross-link connects concepts across different clusters.
• Gaps (under-connected concepts), missing links, and questionable propositions are listed explicitly.
• The structure is a non-hierarchical network, not an argument tree or a MECE decomposition.
• No overclaim: it externalizes and inspects how concepts relate; it does not claim to improve learning, retention, or decisions (see evidence/dossier.md).
• The output is the concept-map artifact, not prose.

Evidence

Tier M/P, with a deliberate scope caveat. Concept mapping has a large human meta-analytic base (Nesbit & Adesope 2006: 55 studies, n=5,818; Schroeder et al. 2018: 142 effect sizes, n=11,814, overall g=0.58, constructing g=0.72 > studying g=0.43) - but those studies measure human knowledge retention, a memory-encoding outcome that does not transfer to an AI agent. That is why this skill is M/P and not S even though its base is larger than the S-graded argument-mapping: tier is set by whether the measured outcome transfers (reasoning quality does; retention does not), not by sample size. The transferable, practitioner-grade claim (Novak & Canas 2008; Davies 2011) is narrow: externalizing how concepts interrelate and forcing every relationship to be named surfaces gaps during construction. Evidence is transferred from human studies, not AI-validated. Full grading: evidence/dossier.md.

Examples

See references/EXAMPLE.md for a completed concept map.

Deep dive: worked example

A full worked run (the shared Northwind scenario)

Concept Map - Worked Example

A completed run of think-concept-mapping, on the shared Northwind scenario. This is the quality bar a generated map should meet.

Northwind is a B2B SaaS weighing a self-serve free-tier launch. The team has a pile of beliefs about how the free tier connects to growth, cost, and sales, but nobody has laid out how those concepts actually relate. This map externalizes the relationships and tests each one as a proposition.

---

Focus question (optional)

• How does a self-serve free tier relate to Northwind’s growth, costs, and sales motion?

Concepts

• Cluster A (Acquisition / growth): free tier, signup volume, activation, paid conversion, customer base, word-of-mouth
• Cluster B (Cost / capacity): support load, infrastructure cost, gross margin, free-user count
• Cluster C (Sales motion): self-serve, sales-assisted motion, enterprise deals, sales team capacity

Propositions (labeled links)

Source concept	Linking phrase (named)	Target concept	Reads as a sentence
free tier	increases	signup volume	The free tier increases signup volume.
signup volume	feeds	activation	Signup volume feeds activation.
activation	drives	paid conversion	Activation drives paid conversion.
paid conversion	grows	customer base	Paid conversion grows the customer base.
free tier	grows	free-user count	The free tier grows the free-user count.
free-user count	increases	support load	A larger free-user count increases support load.
free-user count	increases	infrastructure cost	More free users increase infrastructure cost.
support load	erodes	gross margin	Support load erodes gross margin.
free tier	enables	self-serve	The free tier enables a self-serve motion.
self-serve	reduces reliance on	sales-assisted motion	Self-serve reduces reliance on the sales-assisted motion.
customer base	generates	word-of-mouth	The customer base generates word-of-mouth.
word-of-mouth	increases	signup volume	Word-of-mouth increases signup volume (a loop back into Cluster A).

Cross-links (across clusters)

Source (cluster)	Linking phrase	Target (cluster)	Why this cross-link matters
free-user count (B)	dilutes	paid conversion (A)	Names a tension the rosy growth chain hid: a flood of free users can lower the conversion rate even as raw signups rise.
support load (B)	competes for	sales team capacity (C)	If support is partly staffed by the same people, free-tier load steals capacity from enterprise deals - links cost to the sales motion.
self-serve (C)	weakens	enterprise deals (C->A)	A strong self-serve path can cannibalize higher-ACV sales-assisted enterprise deals, not just add to them.

Surfaced gaps, missing links, and questionable propositions

• Under-connected concepts (gaps): “gross margin” connects only inward (support load erodes it) and nothing downstream - the map never links margin to a decision or to runway, so the cost side is described but its consequence is not. “enterprise deals” is barely connected, which means the map under-represents the high-value sales motion the free tier might threaten.
• Missing links: there is no link from “paid conversion” or “customer base” back to “gross margin” or revenue - the map traces how the free tier creates cost but never closes the loop on how it creates value, so any read of net effect is currently unsupported. Add a “paid conversion -> increases -> revenue -> supports -> gross margin” path before concluding.
• Questionable propositions: “free tier - increases - signup volume” is safe, but the implied chain “free tier therefore grows customer base” rests on “activation - drives - paid conversion”, and the cross-link “free-user count - dilutes - paid conversion” directly contests it. That contested junction (does the free tier net-grow paid customers, or just inflate free signups and conversion drag?) is the real open question, and it was invisible until every link had to be named.

---

Note: the value is not a prettier diagram. Forcing each line to be a named proposition turned a vague “the free tier will grow us” belief into a network with a contested junction (dilution vs conversion) and an unclosed value loop (cost is mapped, value is not). Those two findings are the gap list, and they hand the team a sharper next question than the prose ever did. The map judges nothing about whether to launch - that decision work routes to a decision skill; this only makes the relationships inspectable.

Grounding: the full evidence dossier

What the research does and does not show, with graded sources

Evidence Dossier: Concept Mapping

The single source of truth for the concept-mapping skill. The SKILL.md, the sidecar, and the eval cases all derive from this file. If a claim is not here, it does not belong in the skill.

Skill	thinking-framework-skills.concept-mapping (installable name think-concept-mapping)
Family	synthesis
Evidence tier	M/P (moderate-to-practitioner for the transferable claim; see the scope note below)
Confidence	High that the construction move works and that a large human meta-analytic base exists; the catch is that the base measures the wrong outcome for transfer
Status	draft (authored 2026-06-01 from the discovery corpus)

---

1. The mechanism (what actually does the work)

Concept mapping builds a non-hierarchical semantic network: concept nodes joined by directed, labeled linking phrases, so that each node - link - node triple reads as an explicit proposition (for example “free tier - increases - signup volume”, “support load - constrains - margin”). Clusters of related concepts are then joined by cross-links that connect concepts across different clusters, which is where the integrative insight lives.

The load-bearing move is forcing every relationship to be named. An ordinary diagram lets you draw a line between “free tier” and “conversion” and move on; concept mapping refuses the unlabeled line and demands a verb: does the free tier increase, cannibalize, delay, or gate conversion? Naming the relationship is what externalizes how a domain actually interrelates, and it is what surfaces gaps (a concept with no links), missing links (two concepts that obviously relate but were never connected), and misconceptions (a labeled proposition that, once written as a sentence, is plainly false or unsupported).

The work is done by the labeled-proposition constraint, not by drawing a picture. Drop the constraint and the technique collapses into free-association mind-mapping with unlabeled branches (the excluded Buzan method, below).

2. Lineage

• Novak & Canas (2008), “The Theory Underlying Concept Maps and How to Construct Them” (IHMC Technical Report). The canonical definition: concepts in labeled nodes, connected by labeled linking words/phrases to form propositions; cross-links across map regions as the marker of integrative understanding. This is the source of the labeled-link / proposition constraint that this skill implements.
• Davies (2011), “Concept mapping, mind mapping and argument mapping: what are the differences and do they matter?” Higher Education 62(3). Establishes concept mapping, mind mapping, and argument mapping as three distinct techniques with different structures and purposes - the basis for the hard walls in “When NOT to Use”.
• Nesbit & Adesope (2006), “Learning With Concept and Knowledge Maps: A Meta-Analysis.” Review of Educational Research 76(3). 55 studies, 67 effect sizes, n = 5,818.
• Schroeder, Nesbit, Anguiano & Adesope (2018), “Studying and Constructing Concept Maps: a Meta-Analysis.” Educational Psychology Review 30. 142 effect sizes, n = 11,814; overall g = 0.58, with constructing maps (g = 0.72) outperforming studying pre-made maps (g = 0.43).

No trademark on “concept mapping”; it is a generic descriptive term, named descriptively here, lineage cited.

3. What the evidence shows, and what it does NOT show

This is the crux of the honest grade, so it gets stated plainly.

What the large meta-analyses actually measure: Nesbit & Adesope (2006) and Schroeder et al. (2018) pool studies of human knowledge retention and learning - how much a person remembers or understands after building or studying a map, versus reading text or other study conditions. The reported numbers (Schroeder overall g = 0.58; constructing g = 0.72 > studying g = 0.43) are memory-encoding / learning-transfer outcomes in human students.

Why that does NOT transfer to an AI agent: an AI agent does not have the human memory-encoding bottleneck these studies measure. “Building a map helps a student remember and learn the material better” is precisely the outcome that does not carry over to a model that already has the material in context. So the large effect sizes - although real and well-powered - are evidence for the wrong outcome for an AI use case. They cannot be borrowed to claim the skill makes an agent reason or decide better.

The honest comparison (why this is M/P and not S): the shipped argument-mapping skill is graded S, even though its meta-analytic base (van Gelder and successors) is smaller than concept mapping’s. The reason is outcome, not sample size: argument mapping’s studies measure reasoning quality, which is the outcome that does transfer to an agent. Concept mapping’s much larger base measures knowledge retention, which does not. Tier here is set by whether the measured outcome transfers, not by how big the meta-analysis is. A bigger study of the wrong outcome does not earn a higher grade. This is the discipline the library exists to enforce, so the grade is M/P, not S.

The transferable claim (what this skill may honestly claim): the construction move itself - externalizing how concepts interrelate and forcing every relationship to be named - has practitioner and construct-validity support (Novak & Canas; Davies) independent of the retention outcome. The act of naming each link, demanding cross-links, and reading each triple as a proposition is a structural discipline that surfaces gaps, missing links, and questionable propositions during construction. That gap-surfacing is inspectable and does not depend on a memory mechanism, so it transfers. The claim is scoped to exactly that and no further.

NOT shown / explicitly disclaimed:

• No claim that running this skill improves an agent’s learning, retention, decisions, or downstream answer quality. The retention meta-analyses do not support that for an agent and are not cited for it.
• No borrowing of g = 0.58 / 0.72 / 0.43 as evidence of agent benefit. Those numbers are reported here only to describe what the human literature found, and to make the transfer failure explicit.
• No claim of an effect size for the gap-surfacing benefit; that benefit is practitioner-grade and directional, not quantified.

4. Transferred-evidence flag (required honesty)

All cited evidence is from human learners, and the strongest, largest results measure human knowledge retention - a memory-encoding outcome that does not apply to an AI agent. Evidence is transferred from human studies, not AI-validated, and for this skill the transfer is partial by design: the retention finding does NOT transfer, and only the construction-discipline / gap-surfacing aspect is claimed to. The AI value that survives the transfer test: a model asked to relate a set of concepts will happily emit unlabeled adjacency (“X and Y are related”) or a fluent paragraph that hides which relationships it never actually specified. Forcing every link to be a named proposition, and requiring cross-links, is a direct counter to that vagueness and yields a durable, inspectable artifact in which a false or missing proposition is visible.

5. When it works / when it fails (drives “When NOT to Use” and eval anti-cases)

Works best when:

• A domain has many interrelated concepts and the goal is to externalize and inspect how they relate, not to answer one question.
• You suspect hidden gaps or misconceptions in how a space is understood, and want them surfaced as explicit, checkable propositions.
• Integration across sub-areas matters, so cross-links between clusters carry the value (for example connecting a pricing concept to a support-cost concept).

Fails or misleads when (poor-fit / anti-patterns):

• Evaluating whether one argument or recommendation is sound - that is argument-mapping. Both produce “maps”, so this is the easiest confusion: argument mapping has claims, reasons, co-premises, and objections and judges soundness; concept mapping is a semantic network of propositions and judges nothing. Route soundness work to argument-mapping.
• Top-down decomposition of one question into MECE parts - that is issue-tree. A concept map is a non-hierarchical network, not a decomposition tree, and does not aim for mutual exclusivity.
• Bottom-up clustering of many raw notes with no named relationships - that is affinity-mapping (the KJ method). Affinity mapping groups items into themes; it deliberately does not name the relationship between items. If you only need themes, not propositions, use affinity mapping.
• Fixed event / pattern / structure / mental-model causal levels - that is iceberg-model. The iceberg has prescribed levels; a concept map has none.
• Dropping the labeled-link / proposition constraint. Unlabeled branches radiating from a center is free-association mind-mapping (the Buzan method), which is excluded from this library (X-tier; Farrand, Hussain & Hennessy 2002 found minimal and possibly motivation-offsetting benefit). Without the named-relationship/proposition discipline this skill becomes that excluded method. The constraint is the skill.

6. Output artifact

A concept map: a non-hierarchical network of concept nodes connected by directed, labeled linking phrases, where each node - link - node triple is written as a readable proposition; cross-links that connect concepts across different clusters; and an explicit list of surfaced gaps (under-connected concepts), missing links (relationships that should exist but were not drawn), and questionable propositions (named links that, written out, look false or unsupported). The deliverable is the network of propositions plus that gap/issue list, not prose.

7. Sources

1. Novak, J. D., & Canas, A. J. (2008). The Theory Underlying Concept Maps and How to Construct Them. IHMC Technical Report (Florida Institute for Human and Machine Cognition). - definition: labeled nodes, labeled linking phrases forming propositions, cross-links as the integrative marker.
2. Davies, M. (2011). “Concept mapping, mind mapping and argument mapping: what are the differences and do they matter?” Higher Education, 62(3), 279-301. - the three techniques are distinct.
3. Nesbit, J. C., & Adesope, O. O. (2006). “Learning With Concept and Knowledge Maps: A Meta-Analysis.” Review of Educational Research, 76(3), 413-448. - 55 studies, 67 effect sizes, n = 5,818; outcome = human learning/retention.
4. Schroeder, N. L., Nesbit, J. C., Anguiano, C. J., & Adesope, O. O. (2018). “Studying and Constructing Concept Maps: a Meta-Analysis.” Educational Psychology Review, 30, 431-455. - 142 effect sizes, n = 11,814; overall g = 0.58; constructing g = 0.72 > studying g = 0.43; outcome = human learning/retention.
5. Farrand, P., Hussain, F., & Hennessy, E. (2002). “The efficacy of the mind map study technique.” Medical Education, 36(5), 426-431. - cited only to mark the excluded unlabeled mind-mapping (Buzan) method as X-tier.

Verification status: The Novak & Canas definition and the Davies distinction are well-attested and safe to cite. The Nesbit & Adesope and Schroeder et al. study/effect-size figures (counts, n, g = 0.58 / 0.72 / 0.43) are reported as the published meta-analytic values; confirm the exact figures against the primary articles before any public-facing quantified claim, and note that they are cited here as human-retention findings that do NOT transfer to an AI agent, never as evidence of agent benefit. The core caveat - large base, wrong outcome for transfer, therefore M/P not S - is the load-bearing honesty of this dossier.