# Logframe — LLM Prompt Pack (free bonus)

Use these with ChatGPT, Claude or Gemini, alongside `logframe-template.xlsx`. They follow the Verified harness method: **generate → verify → log**. Four prompts, one framework reference, one checklist.

> Rule zero: the AI drafts, a human signs off. If you cannot trace a number to a source, it does not go to the donor.

---

## Prompt 1 — The brief (do not skip)

```
You are an M&E design assistant. Before writing anything, interview me.
Ask the questions below ONE AT A TIME and do not generate the logframe until all are answered.
If I say "I don't know", record TBD — never invent a value.

1. Programme name and sector
2. Country and context
3. Donor and required format (FCDO / EU / UN / World Bank / foundation / other)
4. The problem statement in one sentence
5. Target group and how many people
6. Duration (start–end)
7. Budget scale
8. Main activities you already know
9. Intended outcomes (what changes for people)
10. Long-term impact (why it matters)
11. Baseline data that exists today (and its year/source)
12. Existing monitoring systems and data sources

At the end, summarise the brief back to me in a numbered list and ask me to confirm before continuing.
```

## Prompt 2 — Generate the logframe

```
Using ONLY the confirmed brief above, produce a logframe as a markdown table with these columns:
Level | Result statement | Indicator | Indicator definition / unit | Baseline | Target |
Disaggregation | Means of verification | Frequency | Responsible | Assumptions & risks

Rules:
1. Exactly 1 impact; 2–4 outcomes; 1–3 outputs per outcome.
2. Every indicator must be SMART: specific, measurable, achievable, relevant, time-bound.
3. Baselines: use the brief. If absent, write "TBD" and add a row in the notes suggesting a baseline study. NEVER invent a baseline.
4. "Means of verification" must name a specific system or document (e.g. EMIS records, school register, payment records) — never "reports".
5. List 3–5 assumptions per outcome level: conditions outside the project's control that must hold.
6. Disaggregate by sex, age, disability and geography where relevant.
7. Label anything that is your inference with [INFERENCE] and anything not supported by the brief with [NOT IN SOURCE].
8. Do not dress correlation as causation. Use "contributed to" language unless the brief includes a comparison design.
9. Finish with a section called "What I could not verify from the brief" listing every gap.
10. After the markdown table, output the same logframe as CSV (comma-separated, quoted) so I can paste it into Google Sheets.
```

## Prompt 3 — Verify it like a hostile donor reviewer (the checking step)

```
Act as a hostile donor reviewer with a checklist. Audit the logframe above and output a table:
Finding | Severity (fail / warn / pass) | Evidence from the logframe | Minimum fix

Check all of the following:
1. NUMBERS: every baseline and target must trace to the brief. Mark untraceable numbers [NOT IN SOURCE].
2. SOURCES: does every indicator have a named means of verification that could actually produce that data?
3. INDICATORS: are they SMART? Flag vague verbs, missing units, missing timeframes.
4. LEVELS: flag any activity written as an output, or output written as an outcome.
5. CAUSATION: flag any claim that the project caused a change without a comparison or contribution argument.
6. CONSENSUS: flag phrases like "stakeholders agreed" without showing the distribution of views.
7. OMISSION: what important result, risk or negative finding does the draft leave out? List at least three candidate omissions.
8. ASSUMPTIONS: are the riskiest assumptions stated, or only the safe ones?
9. DISAGGREGATION: who could be erased by these indicators — women, youth, people with disabilities, remote areas?
10. DONOR FIT: does the structure match the required donor format stated in the brief?

End with: "Claims extracted: n · Traceable: n · Labelled inference: n · Omission audit: pass/fail · Reviewer: [me] · Date: [date]".
```

## Prompt 4 — Convert between instruments

```
Convert the logframe above into:
(a) a Theory of Change narrative with if–then chains and the assumptions between each link, and
(b) an indicator tracking table with columns: Indicator | Baseline | Target | Actual | Attainment % | Data source | Frequency | Responsible.
Keep every indicator unchanged — do not add or silently rename indicators.
(Reverse also works: paste a Theory of Change and ask for a logframe using the rules in Prompt 2.)
```

---

## Framework reference

**Results chain:** Inputs → Activities → Outputs → Outcomes → Impact. Outputs are what you deliver; outcomes are what changes; impact is why it matters.

**Logframe vs Theory of Change:** the ToC explains *why* change is expected (causal pathways and assumptions); the logframe is the *accountability table* (results, indicators, sources, targets) that donors monitor. FCDO commonly expects both, with annual milestones.

**Donor format notes (post-USAID):**
- **FCDO:** logframe + annual milestones + explicit ToC + value-for-money narrative.
- **EU / ECHO:** intervention logic + baseline/target indicators, OECD-DAC criteria addressed.
- **UN & World Bank:** agency results frameworks + indicator reference sheets; results and indicators must reconcile.
- **Foundations & climate funds:** leaner narrative, stricter outcome/attribution language.
- **US federal (post-USAID):** assistance runs through State, PEPFAR and MCC — confirm the current vehicle and template before promising a format.
- Always check the donor's current template at submission time — formats change.

**OECD-DAC criteria to reflect in the design:** relevance, coherence, effectiveness, efficiency, impact, sustainability.

## The 10-point logframe quality checklist

Score one point each — below 10/10, it does not leave your desk:
1. One impact, 2–4 outcomes, outputs only under outcomes.
2. Every indicator is SMART with a unit and timeframe.
3. Every baseline is sourced, or explicitly TBD with a plan.
4. Every indicator has a named means of verification.
5. Effects are disaggregated by sex, age, disability and geography where relevant.
6. Frequency and responsible role are filled in for every indicator.
7. 3–5 riskiest assumptions are stated, not just the safe ones.
8. No correlation dressed as causation.
9. Omission audit run: negative or missing findings surfaced.
10. Structure matches the donor's current format.

---

*Generated by EvaluationAI. Free to use and share with attribution: "Logframe template by EvaluationAI (evaluationai.com)". The full system — skill files, golden sets, verification checklist, provenance ledger and donor-format variants — is in the Verified Harness Pack.*