AI for Everyone study guide · Lesson 6 of 14
Writing and Documents
Separate Drafting from Editing · Summarize Long Material · Transform Tone and Format · Ask Questions of Documents
This is the lesson's study guide from the FlashFeed app, plus its key facts and a self-test.
What you'll be able to do
- Use AI to draft, revise, summarize, transform, and inspect documents while preserving meaning and provenance.
Separate Drafting from Editing
Drafting starts from a goal and source material; editing starts from existing language. Name which job you want. “Rewrite this” may invite unnecessary invention, while “flag unclear sentences without changing the text” produces a diagnostic pass you can inspect first.
For substantive revision, state what must be preserved: facts, voice, quotations, dates, commitments, or legal terms. Ask for a change log when silent alterations would matter. Keep the original beside the revision so fluency does not hide lost meaning.
Three ideas to retain:
- Drafting and Editing Are Different Jobs: Drafting creates a first version from a brief. Editing diagnoses or changes existing language according to stated goals.
- Preservation Rules Prevent Drift: Identify facts, quotations, commitments, voice, and other elements that must survive a rewrite.
- A Change Log Makes Revision Visible: For consequential documents, a short list of material edits helps the author inspect what changed and why.
Key term: Developmental edit means “changes structure, argument, or coverage.”
Summarize Long Material
A summary is a selection, not a smaller copy. Specify who needs it, which questions it should answer, what must be retained, and the desired length. Distinguish an executive summary, abstract, meeting recap, chronology, and extractive quotation list.
Long inputs create special risks: decisive details can be missed, especially when buried among many documents. Chunk with overlap when needed, summarize sections first, and build a final synthesis from those traceable notes. Require page, section, timestamp, or passage references for material claims.
Three ideas to retain:
- A Summary Has a Purpose: Audience and decision need determine which details are central, which are supporting, and which can be omitted.
- Long Context Is Not Perfect Recall: Research finds that model performance can vary with where relevant information appears, including degradation for material in the middle.
- Traceability Improves Review: Section, page, timestamp, or passage links let a reader compare summary claims with the original material.
Key term: Executive summary means “decision-focused overview for a busy reader.”
Transform Tone and Format
AI is useful for turning notes into an email, prose into a table, or a technical explanation into plain language. A transformation should state the new audience and format plus the semantic invariants—the facts, qualifications, quantities, and commitments that cannot change.
Tone is relational, not decorative. “Warm” for a condolence note differs from “warm” for a sales message. Describe the relationship and purpose, then inspect whether the revision added promises, removed caveats, or sharpened uncertain language into certainty.
Three ideas to retain:
- Transformation Changes Representation: A format or tone change should preserve the underlying facts and intended meaning unless the brief explicitly requests substantive revision.
- Audience Defines Plain Language: Useful simplification depends on what the reader already knows and what action they need to take.
- Fluency Can Hide Semantic Drift: A smoother rewrite may drop a caveat, change a quantity, or turn a possibility into a promise.
Key term: Semantic invariant means “meaning that must remain unchanged.”
Ask Questions of Documents
Document question answering works best when the answer must come from a defined packet. Tell the system to cite the exact passage, say when the packet does not answer the question, and keep external knowledge separate. For extraction, define a schema and what to do with missing fields.
A missing mention is not proof that an event never occurred. “Not found in these documents” is narrower and more honest than “false.” Sample the extraction against originals, especially dates, names, negation, units, and table cells.
Three ideas to retain:
- Bound the Evidence Set: A document Q&A task should identify which files count as evidence and whether outside sources are allowed.
- Not Found Is Not False: Failure to locate a claim in a packet shows only that it was not found there; it does not establish that the claim is untrue everywhere.
- Schemas Make Extraction Testable: Named fields, types, and missing-value rules turn loose reading into output that can be sampled and compared with originals.
Key term: Evidence packet means “the defined documents allowed to support an answer.”
Sources
Key terms
- Revision brief
- Instructions describing what to change, what to preserve, and how the result will be judged.
- Abstractive summary
- A condensed restatement written in new language rather than copied passages.
- Tone
- The attitude and relationship conveyed by choices of wording, rhythm, and emphasis.
- Document grounding
- Constraining an answer to information found in a supplied set of documents.
Key facts
Preservation Rules Prevent Drift
Identify facts, quotations, commitments, voice, and other elements that must survive a rewrite.
A Change Log Makes Revision Visible
For consequential documents, a short list of material edits helps the author inspect what changed and why.

Long Context Is Not Perfect Recall
Research finds that model performance can vary with where relevant information appears, including degradation for material in the middle.
Traceability Improves Review
Section, page, timestamp, or passage links let a reader compare summary claims with the original material.

Audience Defines Plain Language
Useful simplification depends on what the reader already knows and what action they need to take.

Fluency Can Hide Semantic Drift
A smoother rewrite may drop a caveat, change a quantity, or turn a possibility into a promise.
Not Found Is Not False
Failure to locate a claim in a packet shows only that it was not found there; it does not establish that the claim is untrue everywhere.

Schemas Make Extraction Testable
Named fields, types, and missing-value rules turn loose reading into output that can be sampled and compared with originals.
Matched pairs
Each item on the left goes with the one on the right.
- Developmental edit
- Changes structure, argument, or coverage
- Copy edit
- Improves clarity, consistency, and correctness
- Change log
- A concise record of material revisions
- Executive summary
- Decision-focused overview for a busy reader
- Extractive summary
- A selection of passages taken from the source
- Semantic invariant
- Meaning that must remain unchanged
- Format transformation
- Recasting content into a new structure
- Plain-language rewrite
- A version adapted for readers without specialist knowledge
- Evidence packet
- The defined documents allowed to support an answer
- Missing value rule
- How absent or unknown fields should be represented
- Source citation
- A pointer from an answer to supporting evidence
Self-test
Answer each question from memory, then check your work against the answer key.
1. What should a rewrite prompt protect?
- AMeaning that must not change
- BOnly the word count
- CA random persona
- DEvery original typo
2. Best first step for sensitive prose?
- ARemove the original
- BReplace it without review
- CHide all changes
- DAsk for diagnosis before rewriting
3. What should a summary prompt specify?
- AA fictional credential
- BAudience and coverage
- CNo source boundaries
- DOnly “make shorter”
4. How can a long-document summary be checked?
- ARemove section labels
- BHide omitted material
- CLink claims to source locations
- DTrust confident wording
5. What must a format conversion preserve?
- AMaterial facts and qualifications
- BA random metaphor
- CThe model's first guess
- DEvery line break
Go deeper
- WatchThe Buried Detail Bury a decisive detail among many documents. Ask an artificial intelligence model for a summary. That detail can be missed.
- StudyAI for Everyone in the FlashFeed app
Answer key
- 1. A Meaning that must not changePreservation rules keep fluent revision from changing substance.
- 2. D Ask for diagnosis before rewritingA diagnostic pass lets the author choose which edits to accept.
- 3. B Audience and coverageSelection depends on who needs what from the material.
- 4. C Link claims to source locationsTraceable references enable comparison with the original.
- 5. A Material facts and qualificationsRepresentation can change while substantive meaning remains fixed.
Sources
- Google Cloud — Prompt design strategies — cloud.google.com/vertex-ai/generative-ai/docs/learn/prompts/prompt-design-strategies
- Liu et al. — Lost in the Middle — aclanthology.org/2024.tacl-1.9/
- NIST — Generative AI Profile — www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence
Image credits
- Long Context Is Not Perfect Recall: AI-generated (gpt-image-1.5)
- Audience Defines Plain Language: AI-generated (gpt-image-1.5)
- Fluency Can Hide Semantic Drift: AI-generated (gpt-image-1.5)
- Schemas Make Extraction Testable: AI-generated (gpt-image-1.5)