Describe your program and Sopact Sense draws the causal chain inline — inputs through impact — and grades every link by how much evidence backs it. For when you know what you do but have never shown why it works.
This guide is for teams ready to take a Theory of Change past a first diagram — to name the assumption under every causal link, attach one measurable indicator to each, rank the weak links by what collapses downstream, and wire the whole thing to live data so the picture updates as evidence arrives. Where the beginner guide draws one clean, graded diagram in an afternoon, this one turns that diagram into a working measurement system a funder can audit end to end.
Who this is for. Program evaluators, MEL and impact leads, fund and portfolio managers, and directors at established nonprofits, foundations, and accelerators — teams running more than one program, sitting on some outcome data already, or facing a funder who wants evidence rather than intentions. Building your first Theory of Change, or want the fast version? Start with How to build a Theory of Change.
The process. It builds on the same graded chain, then goes deeper. First, an assumption-and-indicator table — one row per arrow, each with the single indicator that would test it (who is measured, what changes, by when). Second, a weak-link diagnosis that ranks every amber and red by how much the chain depends on it, and separates a measurement gap from a design flaw. Third, an improvement plan — the highest-value fixes, each with a before/after rewrite and one data-collection step. Fourth, alignment to a shared framework (IRIS+ or the Five Dimensions) and a plan to track outcomes longitudinally on persistent participant IDs.
Time and outcome. Plan on a focused day or two, plus the cadence of real data collection. You end with an audit-ready Theory of Change: a graded diagram, an evidence trail behind every color, a prioritized measurement plan, and a way to keep it current as data arrives — not a poster on the wall, but the backbone of how you report impact.
Yes — and the version a funder trusts is built from evidence, not aspiration. A theory of change maps your program as a causal chain — inputs → activities → outputs → outcomes → impact — and the real work lives in the assumptions under the arrows, not the boxes. Most theories of change fail there: the outcomes sound right, but nobody can say what evidence holds each link up.
The advanced prompt below turns that principle into a complete audit. It reads only what your program states publicly, grades every node and arrow Green, Amber, or Red, and tells you exactly which fix to make this quarter. We ran it on The Lantern Network's public mentoring-program page — the sections that follow show the prompt for each part, then what came back.
Building a Theory of Change is foundational impact work—it clarifies how your program creates change. You have two paths:
Option 1: Use the Prompt Pack (Self-Service)
Download the prompt pack below and paste it into Claude, ChatGPT, or Sopact Sense. No setup needed. You get:
Option 2: Use the Impact Framework Builder Skill (Guided & Interactive)
If you have interview transcripts or prefer Claude to walk you through it interactively, use the impact-framework-builder skill. It will extract your theory from stakeholder conversations, check for logic gaps, and convert your framework into a survey.
Download the Theory of Change Prompt Pack
Works in Sopact Sense, Claude, or ChatGPT. Includes the master prompt, all four parts as standalone prompts, and a bonus page-tightening prompt.
Every grade depends on who is reading. A board sees "87% placed" as a headline; a renewing funder asks what evidence sits behind it. So the prompt's first instruction is a decision frame: before building anything, state in one line who would use this theory of change and for what decision. For Lantern, that came back as a corporate sponsor deciding whether to renew its grant — asking not "did placements happen?" but "does this mentorship dollar produce durable change in people?" Every judgment below is made from that reader's chair.
Here is the full prompt. Paste it whole, swap in your program name and source, and Sense produces all four parts in one pass:
Build a Theory of Change for [PROGRAM NAME] using only what the program publicly states at [SOURCE — URL or pasted program description]. Before building, state in one line who would use this theory of change and for what decision — then make every judgment from that reader's perspective. PART 1 — Render the full causal chain as an interactive left-to-right flow diagram: Inputs → Activities → Outputs → Short-Term Outcomes → Medium-Term Outcomes → Long-Term Outcomes → Impact. One node per element; write every outcome as a change in people, never as an activity. On each arrow, label the causal assumption that must hold for the left node to produce the right one. Color every node and arrow: GREEN = specific AND evidenced; AMBER = stated but vague, or specific but unevidenced; RED = missing, or exists only as [INFERRED]. Hovering any node shows the exact source language that supports it; hovering any arrow shows the assumption and its proposed indicator. Tag anything not explicitly stated as [INFERRED]. Include a legend, program name, source URL, and date. PART 2 — One row per arrow: Assumption | Grade | Evidence (quoted, paraphrased, or "none") | one measurable indicator that would test it, phrased so the program could actually collect it (who is measured, what changes, by when). PART 3 — List every AMBER and RED element, ranked by how much the causal chain depends on it. For each: why it is weak in one sentence, what claim collapses downstream if the assumption fails, and whether the fix is a program-design problem or a measurement gap. PART 4 — The top 3–5 fixes, in priority order for the decision named above. For each: current language (or "missing") → proposed rewrite, plus the single data collection step that would move it toward green. CLOSING SUMMARY — 3–4 sentences: overall strength of the causal logic, the single weakest link a skeptical funder would attack first, and the one action to take this quarter. RULES — Source fidelity is absolute: never invent program content; if the source does not say it, mark it RED or [INFERRED]. Every green grade must be traceable to specific source language. Diagram and text must agree.
Example source: https://www.lanternnetwork.org/mentoring-program. The rules at the end do the heavy lifting — no invented content, every green traceable to a quote, and the diagram and tables must agree. That last rule is what makes the output defensible: the visual is the claim, the tables are the evidence trail.
Render the full causal chain as a left-to-right flow diagram: Inputs → Activities → Outputs → Short-Term Outcomes → Medium-Term Outcomes → Long-Term Outcomes → Impact. One node per element; write every outcome as a change in people, never as an activity. On each arrow, label the causal assumption that must hold. Color every node and arrow green, amber, or red. Hovering any node shows the exact source language that supports it; hovering any arrow shows the assumption and its proposed indicator. Tag anything not explicitly stated as [INFERRED]. Include a legend, program name, source URL, and date.
Two disciplines make this diagram different from the boxes-and-arrows version on most websites. First, every outcome is written as a change in people — "mentees gain career clarity," never "we run workshops." Second, the arrows carry the argument: each one is labeled with the assumption that must hold for the left node to produce the right one, and the arrow gets its own grade. Hover any node and you see the exact source quote behind it; hover any arrow and you see the assumption plus the indicator that would test it.
The rubric is strict on purpose. Green means specific and evidenced — the program names it concretely and shows data ("87% secured internships, jobs, or promotions"). Amber means stated but vague, or specific but unevidenced — no number, no timeframe, or a claim with no data behind it. Red means missing entirely, or existing only as [INFERRED]. For Lantern, the pattern was immediate: the left half of the chain is concrete — 288 mentees served, 251 internships delivered, named sponsors — while everything past placement rests on three testimonials.
GRADE: green | 87% placed | 288 mentees, 251 internships — stated, specific; amber | confidence & clarity | claimed in three stories, never measured; red | mentoring dosage | no sessions, hours, or follow-up recorded
For every arrow, produce one row: Assumption | Grade | Evidence (quoted, paraphrased, or "none") | one measurable indicator that would test it, phrased so the program could actually collect it — who is measured, what changes, by when.
This table is the evidence trail behind the diagram — one row per arrow, so nothing in the visual floats free of a citation. The indicator column is the practical payoff: each one is phrased so the program could actually collect it, naming who is measured, what changes, and by when. For Lantern's weakest arrow — Outputs → Short-Term Outcomes — the evidence column reads "none: no dosage captured, no before/after measure of any soft outcome," and the indicator that would test it is "change in self-reported confidence and clarity, baseline versus program exit, same mentees tracked by ID." That's not a critique; it's a work order.
List every amber and red element, ranked by how much the causal chain depends on it. For each: (a) why it is weak in one sentence, (b) what claim collapses downstream if the assumption fails, (c) whether the fix is a program-design problem or a measurement gap — these require different responses.
The ranking is by dependence, not chain order — the question is which weakness takes the most down with it. Lantern's number one wasn't its red impact claims; it was Outputs → Short-Term Outcomes: no dosage is recorded and no before/after measure exists, so there is no evidence that being matched actually changes a mentee. If that assumption fails, confidence, clarity, and every long-term claim above it collapse — the whole ladder floats on the 87% placement number alone.
The design-versus-measurement tag matters just as much. A measurement gap means the program plausibly works but has never measured it — the fix is data collection. A design problem means the causal logic itself is unstated or untested — the fix is program thinking. Lantern's diagnosis came back mostly measurement gaps, which is the encouraging read: nothing needs redesigning, but almost everything needs measuring.
Give the top 3–5 fixes, in priority order for the decision named above. For each: show the current language (or "missing") → a proposed rewrite, plus the single data collection step that would move it toward green.
Each fix is a before-and-after pair, so the path from claim to evidence is concrete. Lantern's first fix: current language "TLN mentees consistently report increased self-confidence" — stated, never measured — rewritten as "X% of mentees show a measurable rise in job-search confidence and career clarity from intake to program exit." The single data step: a short baseline/endline scale with a persistent participant ID, so the same mentees are compared over time rather than two unrelated snapshots. Fixes two through five follow the same shape: capture mentoring dosage, follow placed mentees at 12 and 24 months, publish the denominator behind the 87%, and test which mechanism actually drives placement.
Close with 3–4 sentences: the overall strength of the causal logic, the single weakest link a skeptical funder would attack first, and the one action to take this quarter.
The summary is the executive read. Lantern's verdict: the chain is strong on its left half and unproven on its right — inputs, activities, and outputs are concrete and evidenced, but everything past placement rests on three testimonials, and the soft outcomes the program says drive its results are asserted, never measured. The one action this quarter: attach a persistent participant ID and a baseline/endline confidence-and-clarity measure to the next cohort, so next year's renewal case rests on tracked change in people rather than three stories.
Name the decision before you grade. The same chain earns different grades for different readers — a board celebrating placements and a funder testing durability are not asking the same question. If the grades feel too generous, the decision frame is usually too soft.
Point it at a URL and keep the receipts. Running the prompt against a public page is the honest test: it can only grade what the page says, which shows you your program exactly as a stranger reads it. Every green must trace to a quote — if you can't find the source language, the grade is wrong.
Separate design problems from measurement gaps. They demand different responses. Most ambers and reds in a real program are measurement gaps — the logic is fine, the data was never collected — and each one is fixable with a single indicator, not a program redesign.
Tighten your program page while you're here. Once the chain is graded, ask Sense to bring your public claims in line with your evidence:
Based on the grades above, suggest edits to my program page so its claims match the evidence. Flag every sentence that overstates what we can show, and rewrite it to be accurate and specific.
The same prompt works for a Logic Model, Logframe, or Results Framework — swap the framework name and keep the parts, rubric, and rules unchanged.
A theory of change is a causal map of how a program creates impact — inputs → activities → outputs → short, medium and long-term outcomes → impact — with the assumption behind each link made explicit. It explains not just what you do, but why you believe it works.
Describe your program in a few sentences (or paste its web page) and ask the AI to draw the causal chain, name the assumption under each arrow, flag the weak links, and grade every element green, amber or red by how much evidence supports it. In Sopact Sense this takes minutes and stays grounded only in what your program actually states — it marks anything inferred so it never invents outcomes.
Weak theories of change fail at the assumptions, not the outcomes. The weakest links are the ones stated as beliefs — "we believe mentoring builds confidence" — with no indicator, or long leaps to long-term impact with no follow-up data. Naming and measuring those assumptions is what makes the logic credible to a funder.
ChatGPT, Claude, and Gemini are fine for a quick test — but not for an answer you'll put in front of a funder or board. When it has to hold up, run it in Sopact Sense.
Try it in Sopact →
