Causal Loop Diagrams

Quick facts

Mechanism: Builds a signed causal loop diagram by closing the feedback loops in a situation, labeling each loop reinforcing (R) or balancing (B) with its link polarities, and reading likely dynamics (spiral, goal-seeking, or oscillation) off which loop dominates.

Produces: signed-causal-loop-diagram

Evidence: M/P · Domain: Systems and Consequences

Use when a situation feeds back on itself - growth that funds more growth, a fix that recreates its own problem, capacity that relieves then re-attracts demand - and you need to see why it keeps accelerating, stalling, or overshooting. Not for a single accumulation, a one-directional consequence tree, or a genuinely linear chain.

Picture it

Every causal loop diagram is built from two kinds of closed loop. A reinforcing loop (R) feeds on itself and accelerates; a balancing loop (B) pushes back toward a goal or limit.

graph TD
  subgraph Rloop["Reinforcing (R): spirals up or down"]
    direction LR
    a1["Users"] --> a2["Word of mouth"]
    a2 --> a1
  end
  subgraph Bloop["Balancing (B): seeks a goal / hits a limit"]
    direction LR
    b1["Users"] --> b2["Server load"]
    b2 --> b3["Slower app"]
    b3 --> b1
  end
  Rloop ~~~ Bloop
  classDef r fill:#fde7e7,stroke:#dc2626,color:#7f1d1d
  classDef b fill:#e6f0fe,stroke:#1967d2,color:#10316b
  class a1,a2 r
  class b1,b2,b3 b

Sign each link, close the loop, and label it R or B. Which loop dominates tells you whether the system spirals, settles toward a goal, or oscillates.

People narrate systems as one-directional chains and silently drop the loop-back. “More users, so more revenue” omits “…which funds acquisition, which brings more users” - the cycle that actually drives the behavior. This skill performs one distinct move: close the feedback loops and sign them. Trace each cycle back to its start so it closes, give every link a polarity (does a rise in A raise (+) or lower (-) B), and label the whole loop reinforcing (R) when the signs multiply to net-positive (it amplifies: a vicious or virtuous spiral) or balancing (B) when they multiply to net-negative (it counteracts: goal-seeking, or oscillation when delayed). Then read likely behavior off the structure: which loop dominates, and therefore whether the system spirals, seeks a goal, or oscillates. The output is a signed causal loop diagram framed as a structured argument about dynamics - not a prediction. It corrects a specific, well-evidenced failure (people misperceive feedback); it does not claim to predict the system or to teach systems thinking wholesale.

When to Use

• A variable plausibly feeds back on itself through a cycle (growth funds growth; a fix recreates its problem; relief of a constraint re-attracts the load).
• The puzzle is why does this keep accelerating / stalling / overshooting and undershooting - behavior that a linear story cannot explain.
• You want an inspectable, signed structure (R/B loops with polarities) before reasoning about leverage or intervention.

When NOT to Use

• A single accumulation, no loop (one stock, net flow, no cycle): use think-stocks-and-flows-reasoning. That skill reasons about one quantity from its net flow; it does not close or sign a loop.
• You only need to name that feedback exists as one structural layer among events, patterns, and structure: use think-iceberg-model. It names feedback as a structure item but does not close, sign, or diagram loops.
• Forward, one-directional consequences that fan out and do not loop back: use think-futures-wheel. It is an acyclic consequence tree by construction - no loop, no polarity.
• The structure is genuinely open-loop / linear. If the chain does not actually feed back, forcing a loop manufactures false feedback. Say “no closed loop found - this is a linear chain” and stop; do not invent a loop to fill the diagram.
• Teaching general systems thinking, hunting leverage points, or wholesale systems mapping - out of scope (separate catalog rows). This skill does one move: close and sign loops, then read dominance.

Instructions

When asked to map why a situation keeps accelerating, stalling, or oscillating, follow these steps:

1. List the candidate variables. The quantities and conditions in play (users, marketing spend, capacity, defect rate, morale). Use variables that can rise or fall, not events.
2. Find the loops by closing them. For each chain, follow it forward and check whether it returns to a variable it started from. If it does, you have a closed loop. If it does not, mark that chain open (linear) and set it aside - do not force it closed.
3. Sign each link. For every arrow A -> B, assign a polarity: + if a rise in A pushes B up (and a fall pushes it down, same direction), - if a rise in A pushes B down (opposite direction).
4. Label each loop R or B. Multiply the link signs around the loop. Even number of negatives (net +) = reinforcing (R): it amplifies (spiral). Odd number of negatives (net -) = balancing (B): it counteracts (goal-seeking; oscillation if there is a delay). Name and number each loop (R1, B1, …).
5. Note delays. Mark any link where the effect arrives late; delays in balancing loops are what turn goal-seeking into oscillation/overshoot.
6. Read the behavior off the structure. State which loop currently dominates and the resulting dynamic: reinforcing dominance = a vicious or virtuous spiral; balancing dominance = goal-seeking; a delayed balancing loop = oscillation. Frame this as an argument (“if R1 dominates, expect…”), explicitly not a prediction.
7. Record the open parts honestly. Note where no loop closed, so the diagram does not overstate how much of the situation is actually feedback-driven.
8. Emit the signed causal loop diagram per references/TEMPLATE.md.

Output Format

Use the template in references/TEMPLATE.md. The deliverable is the signed loop diagram - the R/B loop inventory with link polarities and the behavior read-out framed as an argument - not prose.

Quality Checklist

Before finalizing, verify:

• Every loop is genuinely closed (the chain returns to a variable it started from); open chains are marked linear, not forced into loops.
• Every link has a polarity (+/-), and each loop’s R/B label follows from the product of its signs.
• Delays are marked where they exist (they drive oscillation/overshoot).
• The behavior read-out names the dominant loop and the resulting dynamic (spiral / goal-seeking / oscillation).
• The read-out is framed as a structured argument, not a prediction or forecast.
• No loop was manufactured to fill the diagram; the linear parts are recorded honestly.
• No overclaim: it externalizes and signs structure; it does not predict the system or teach systems thinking wholesale (see evidence/dossier.md).
• The output is the signed causal loop diagram artifact, not prose.

Evidence

Tier M/P, transferred-evidence. The strongly evidenced fact is the failure this skill targets: people systematically misperceive feedback and accumulation (Sterman 1989, Management Science; Sweeney & Sterman 2000, System Dynamics Review). That base is shared with think-stocks-and-flows-reasoning and does not by itself prove that drawing a causal loop diagram fixes it. The CLD-specific evidence is moderate and conditional: a 2025 quasi-experimental study (ScienceDirect S2451958825000284) finds a conditional effect, and Schaffernicht (2010, Systems Research and Behavioral Science) documents CLD reliability problems (subjectivity, non-reproducibility) - cited here against inflation. All evidence is human-subject, not AI-agent-validated. The transferable claim is scoped to externalizing loop structure and signing polarity, not to predicting system behavior. No effect size is quoted because none has been verified against the source. Full grading: evidence/dossier.md.

Examples

See references/EXAMPLE.md for a completed signed causal loop diagram.

Deep dive: worked example

A full worked run (the shared Northwind scenario)

Signed Causal Loop Diagram - Worked Example

A completed run of think-causal-loop-diagrams, on the shared Northwind scenario. This is the quality bar a generated diagram should meet.

Northwind is a B2B SaaS weighing a self-serve free-tier launch. The free tier is meant to drive a viral, self-funding growth engine. This skill closes and signs the loops to see whether that engine is reinforcing, what would balance it, and which loop is likely to dominate - framed as an argument, not a forecast.

---

Variables in play

• Free-tier signups, Active free users, Word-of-mouth referrals, Paid conversions, Revenue, Support load, Support quality, Free-user experience, Server/cost strain.

Closed loops (signed)

Loop	Link path with polarities	Net sign	Label (R/B)	Delay?	What this loop does
R1 (growth engine)	Active free users -(+)-> Word-of-mouth referrals -(+)-> Free-tier signups -(+)-> Active free users	+ (two positives)	R (reinforcing)	yes (referrals build slowly)	amplifies: a virtuous spiral of self-funding growth IF it engages
R2 (revenue reinvest)	Paid conversions -(+)-> Revenue -(+)-> Marketing spend -(+)-> Free-tier signups -(+)-> Active free users -(+)-> Paid conversions	+ (all positive)	R (reinforcing)	yes	amplifies: revenue funds more acquisition that funds more revenue
B1 (support strain)	Active free users -(+)-> Support load -(-)-> Support quality -(+)-> Free-user experience -(+)-> Word-of-mouth referrals -(+)-> Active free users	- (one negative)	B (balancing)	yes (quality erodes before churn shows)	counteracts: rising free users degrade support, which throttles referrals - caps the engine
B2 (cost ceiling)	Active free users -(+)-> Server/cost strain -(-)-> Free-user experience -(+)-> Free-tier signups -(+)-> Active free users	- (one negative)	B (balancing)	yes	counteracts: free users you do not monetize raise cost strain that degrades the product, limiting growth

Open / linear parts (no loop closed - recorded honestly)

• “Launch announcement -> initial signup spike”: a one-time, linear kickoff. It seeds the loops but does not itself feed back, so it is not a loop. (Mapping its downstream consequences acyclically would be think-futures-wheel, not this skill.)

Behavior read-out (an argument, not a prediction)

• Dominant loop right now: none yet - at launch the system sits before any loop has gained. The structure argues that R1/R2 dominate early (low users, slack support and cost headroom, so the balancing loops are weak), then B1 and B2 strengthen as free users pile up and begin to dominate.
• Resulting dynamic: overshoot-then-stall, not a clean spiral. Reinforcing growth runs first; because the balancing loops act with a delay (support quality and cost strain erode before the slowdown is visible), the likely shape is overshoot - rapid early growth, then a stall or dip as B1/B2 bite. This is the classic “limits to growth” structure, argued from the loop signs, not predicted as a number.
• What would flip dominance: raising Paid conversions (so R2’s revenue actually funds the support and infrastructure that weaken B1/B2) keeps the reinforcing loops on top longer. If conversion stays low, the free tier loads B1 and B2 (cost and support) without feeding R2, and the balancing loops dominate sooner - growth that pays for nothing.
• Honest scope: a structured argument about likely dynamics from the loop structure, not a forecast of signup numbers or dates. Another modeler might sign or include loops differently (CLD reliability is a known limit); the value is the explicit, inspectable structure, not a prediction.

---

Note: the load-bearing move is closing the loops and signing them. The naive story - “free tier -> viral growth -> revenue” - is a one-directional chain (an R-loop narrated without its loop-back, and with the two balancing loops dropped entirely). Signing the loops surfaces that the same free users who drive R1 also feed B1 (support) and B2 (cost), and that the delay makes overshoot, not a smooth spiral, the structure’s argument. It also reframes the real lever - paid conversion - as the thing that keeps the reinforcing loops dominant, which routes the decision back to monetization design.

Grounding: the full evidence dossier

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

Evidence Dossier: Causal Loop Diagrams

Single source of truth for the causal-loop-diagrams skill. The SKILL.md, sidecar, and evals derive from this. A moderate-evidence, transferred-evidence skill: the failure it targets is strongly evidenced; that causal loop diagrams (CLDs) specifically fix it is only moderately and conditionally evidenced. Grade honestly; do not borrow the strong number for the weaker claim.

Skill	thinking-framework-skills.causal-loop-diagrams (installable name think-causal-loop-diagrams)
Family	systems-and-consequences
Evidence tier	M/P (transferred-evidence; conditional)
Confidence	High that people misperceive feedback; moderate-and-conditional that CLDs improve it; the externalization move is mechanically sound
Status	draft (authored 2026-06-01 from the discovery corpus)

1. The mechanism (what actually does the work)

The distinct cognitive move is closing feedback loops and signing them. You notice that a variable feeds back on itself through a cycle (sales fund marketing, marketing drives sales), trace the cycle back to its start to close it, assign a polarity to each link (does a rise in A raise or lower B: + or -), and then label the whole loop by the product of its link signs:

• A loop with an even number of negative links (net positive) is reinforcing (R): it amplifies, producing a vicious or virtuous spiral (exponential-looking growth or collapse).
• A loop with an odd number of negative links (net negative) is balancing (B): it counteracts, producing goal-seeking behavior toward a target, or - when delayed - oscillation.

The skill then reads likely behavior off the structure: which loop currently dominates, and therefore whether the system spirals, seeks a goal, or oscillates. The load-bearing work is the loop closure plus R/B polarity assignment: making the cycle explicit and signed, rather than narrating a one-directional chain. No shipped skill in this library performs loop closure or polarity signing; that is the gap this skill fills.

This is a structured argument about structure, not a prediction. A signed CLD says “if this loop dominates, expect a spiral” - it does not forecast a number or a date.

2. Lineage

• System dynamics: Jay Forrester (origin); John Sterman, Business Dynamics (2000), which formalizes CLD notation (R/B loops, link polarity); Donella Meadows, Thinking in Systems (2008), which frames reinforcing and balancing feedback in plain language.
• CLDs are a qualitative companion to stock-and-flow models (the quantitative side of the same discipline). This skill is the loop side; think-stocks-and-flows-reasoning is the single-accumulation side.

No trademark. Named descriptively (the field’s own generic term).

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

This skill banks two separate evidence pools. Keep them separate - the honesty of the grade depends on not merging them.

Pool A - the failure (strong, S-tier, but it is the SAME base stocks-and-flows banks). People systematically misperceive feedback and accumulation. Sterman (1989), “Misperceptions of Feedback in Dynamic Decision Making,” Management Science, shows subjects manage a dynamic system (the Beer Distribution Game) far worse than the structure allows, because they ignore feedback and time delays. Sweeney & Sterman (2000), “Bathtub Dynamics,” System Dynamics Review, shows even highly educated subjects (MIT graduate students) fail simple feedback/accumulation tasks. This pool is robust. But it only establishes that the error is real. It does NOT establish that drawing a CLD fixes it. Borrowing this S-tier strength for the CLD-effectiveness claim would be a laundered statistic; this dossier refuses that.

Pool B - CLDs specifically (moderate and conditional). A 2025 quasi-experimental study, “Influence of Causal Loop Diagrams on Systems Thinking” (ScienceDirect, article S2451958825000284), finds a conditional effect of CLDs on systems-thinking performance: a benefit under some conditions, not a clean uniform improvement. Treat this as moderate, conditional support for the externalization-and-signing move, not as proof that CLDs reliably improve reasoning for everyone.

Counter-evidence (cited deliberately, against inflation). Schaffernicht (2010), “Causal Loop Diagrams: An Analysis of the Reliability of an Inference Tool,” Systems Research and Behavioral Science, critiques CLD reliability: different modelers draw different loops and polarities from the same situation (subjectivity), and CLDs are not reproducible. This is a real limit on the predictive use of CLDs, and a reason to scope the claim to externalizing and signing structure rather than to predicting behavior.

Honest net framing. The strongly evidenced fact is the failure (Pool A), which is shared with stocks-and-flows and does not by itself credit CLDs. The CLD-specific evidence (Pool B) is moderate and conditional, and there is named counter-evidence (Schaffernicht) on reliability. So the defensible claim is: making feedback loops explicit and signing their polarity externalizes structure that humans (and narrating models) routinely miss; the resulting signed diagram is an inspectable argument about likely dynamics. The claim is NOT that the diagram predicts what the system will do, nor that CLDs reliably improve systems thinking for all users.

4. Transferred-evidence flag

All cited evidence is human-subject. None of it validates causal-loop diagramming for an AI agent. A language model narrating a situation slips into the same trap the human studies document: it describes a one-directional chain (“more users, so more revenue”) and silently drops the loop-back (“…which funds more acquisition, which brings more users”). Forcing explicit loop closure and R/B signing is a direct structural counter to that, and the signed CLD is inspectable. But the transferable claim is scoped: it covers externalizing loop structure and signing polarity, not predicting system behavior. The behavior read-out is a labeled hypothesis (“if R1 dominates, expect a spiral”), never a forecast.

5. When it works / when it fails

Works best when: the situation plausibly contains feedback - a variable that, through a cycle, affects itself (growth that funds more growth; a fix that quietly creates the problem it fixes; capacity that relieves then re-attracts demand). The question is why does this keep accelerating / stalling / overshooting and undershooting.

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

• The structure is genuinely open-loop / linear. If the chain does not actually feed back, forcing a loop manufactures false feedback. An honest output here is “no closed loop found; this is a linear chain - use a different tool.” Do not invent a loop to fill the diagram.
• A single accumulation with no loop - that is think-stocks-and-flows-reasoning (one stock, net flow, no cycle, no polarity sign).
• You only need to name that feedback exists as one structural layer - that is think-iceberg-model, which names feedback as a structure item but does not close, sign, or diagram loops.
• Forward, one-directional consequences (a consequence tree that fans out and does not loop back) - that is think-futures-wheel, which is acyclic by construction.
• Teaching general systems thinking, finding leverage points, or wholesale systems mapping - out of scope here; those are separate catalog rows. This skill does one move: close and sign loops, then read dominance.

6. Output artifact

A signed causal loop diagram: an inventory of the closed feedback loops in the situation, each labeled reinforcing (R) or balancing (B) with its link polarities shown, plus a behavior read-out stating which loop currently dominates and the resulting dynamic (spiral, goal-seeking, or oscillation). The read-out is framed as a structured argument (“if R1 dominates, expect a virtuous spiral”), explicitly not a prediction. The artifact also records, honestly, where no loop closed (the linear parts) rather than forcing loops onto them.

7. Sources

1. Sterman, J. (1989). “Misperceptions of Feedback in Dynamic Decision Making.” Management Science. (Pool A - the failure.)
2. Sweeney, L. B., & Sterman, J. (2000). “Bathtub Dynamics: Initial Results of a Systems Thinking Inventory.” System Dynamics Review. (Pool A - the failure, in educated subjects.)
3. “Influence of Causal Loop Diagrams on Systems Thinking” (2025). ScienceDirect, article S2451958825000284. (Pool B - CLD-specific, conditional effect.)
4. Schaffernicht, M. (2010). “Causal Loop Diagrams: An Analysis of the Reliability of an Inference Tool.” Systems Research and Behavioral Science. (Counter-evidence - subjectivity and non-reproducibility; cited against inflation.)
5. Sterman, J. (2000). Business Dynamics; Meadows, D. (2008). Thinking in Systems. (Lineage and CLD notation.)

Verification status: Pool A (Sterman 1989; Sweeney & Sterman 2000) is well-attested and is the same misperception base the stocks-and-flows dossier banks - it does NOT by itself prove CLDs work. The 2025 CLD study (S2451958825000284) reports a conditional effect; confirm the exact conditions and any effect size from the source before quoting a number - none is quoted here. Schaffernicht (2010) is cited deliberately as a reliability caution. No effect size is stated in this dossier because none has been verified against the source; do not add one without checking. The honest scope - “externalize and sign loop structure,” not “predict behavior” - is the core caveat.