Source-Grounded Summary & Synthesis Worksheet

Use this worksheet when you want AI to summarize notes, synthesize several sources, or turn messy material into a handoff-ready view. It keeps the source material visible so the output does not become detached from the evidence.

Summary Versus Synthesis

Choose the task before prompting.

• Summary: compress one source while preserving meaning and limits.
• Synthesis: combine sources to show agreements, differences, gaps, implications, and next questions.

Both may shorten. Neither should lose evidence, caveats, gaps, or uncertainty.

Source Inventory

List the material you are using.

Before prompting, remove material that is not approved for the tool or audience. Replace sensitive details with placeholders when they are not needed for the task.

Use source labels the reviewer can find without guessing.

Purpose And Audience

• Who will read or use the output?
• What decision, update, or next step should it support?
• Is the task a summary of one source or a synthesis of multiple sources?
• What would be risky if the output is wrong?
• What should the AI explicitly avoid doing?

Good:

Turn three status notes into a weekly rollup for a program lead. Preserve disagreements and label gaps.

Weak:

Tell me what is going on and what we should do.

Prompt Setup

Copy and adapt this block.

Task:
Summarize or synthesize the labeled source material below for [audience] so they can [intended use].

Context:
Use only the labeled sources. The sources may be incomplete.

Constraints:
- Do not invent missing facts, names, dates, owners, quotes, links, or numbers.
- Preserve disagreements, caveats, and uncertainty.
- Label unsupported interpretations as "Possible interpretation, not stated."
- Mark missing information as "Needs verification."

Output shape:
Use sections for [Key Points / Decisions / Action Items / Shared Themes / Differences / Gaps / Next Questions].
Include source labels for major claims.

Fidelity Ladder

Use this quick rubric when reviewing the output:

1. Grounded: important claims trace to labeled sources.
2. Compressed: the output is shorter, but it did not invent or erase meaning.
3. Drifted: the answer sounds smoother because caveats, disagreement, or limits disappeared.
4. Fabricated: adds names, dates, owners, numbers, quotes, causes, or recommendations that are not in the sources.

Aim for grounded and compressed. Revise drifted or fabricated outputs before handoff.

Summarization Pass

Use this when working from one source.

• Main points captured:
• Decisions stated in the source:
• Action items stated in the source:
• Risks or blockers stated in the source:
• Open questions:
• Details intentionally omitted:
• Items marked “Needs verification”:

Review question: did the summary preserve what the source says without adding a cleaner story?

Drift check:

Source says: “The launch date depends on final review.” Bad summary: “The launch is delayed.” Better summary: “Launch timing depends on final review; the source does not state a delay.”

Synthesis Pass

Use this when working from multiple sources.

Review question: did the synthesis preserve disagreements and gaps, or did it turn messy source material into false consensus?

False-consensus check:

Source A says onboarding steps are confusing. Source B says setup questions delay buying decisions. Source C lists three setup decisions but does not identify which one is hardest. Weak synthesis: “Setup needs to be simplified.” Better synthesis: “Sources agree setup clarity matters, but none identifies the highest-friction step. Next question: which setup step creates the most delay?”

Grounding Check

Mark every important claim.

If a claim has no source, rewrite as a question, label it as an interpretation, or remove it.

Worked row:

Omission And Hallucination Review

Check for:

• names, dates, owners, numbers, links, or quotes not present in the sources
• disagreements that disappeared from the summary
• caveats or uncertainty that became confident statements
• recommendations that were not supported by the sources
• action items that gained invented owners or due dates
• source material copied too broadly when a summary would be safer

Issues found:

• [List issues or write “none found.”]

Uncertainty Labels

Use these labels:

• Stated: directly present in a source.
• Possible interpretation, not stated: a reasonable inference that the source does not directly say.
• Needs verification: useful but not confirmed by the source packet.
• Sources disagree: keep both sides visible; do not average them into a cleaner claim.

Worked Mini-Example

Scenario: three training notes need to become one briefing for a team lead.

Handoff note:

I synthesized Sources A, B, and C. Setup clarity appears as a likely theme across A and B. Source C adds constraints but does not confirm the pain point. I removed the unsupported time-savings claim. Next owner should identify the highest-friction setup step.

Handoff Sign-Off

• Reviewer:
• Date:
• Intended audience:
• Task type: summary / synthesis / both
• Sources checked:
• Claims still uncertain:
• Disagreements preserved:
• Gaps or next questions:
• Sharing limits:
• Final decision: use / revise / verify further / escalate / do not use

The reviewer named above owns the output. AI drafted or transformed the material; a person is accountable for accuracy, omissions, labels, and sharing limits. An unsigned worksheet is unfinished.

Reference Links

These public references are useful starting points for deeper study. They are linked for attribution and further reading; the worksheet above is synthesized as original LIW training guidance.

• OpenAI: Prompt engineering
• Anthropic: Prompt engineering overview
• People + AI Guidebook