Draft in bounded passes

Generated prose often arrives with smooth transitions and confident rhythm. That surface can make a first pass feel final. Treat it as editable material. A draft earns trust only after its claims, logic, fit, and language have been reviewed.

Avoid asking for an entire important document in one turn. A long generation hides where decisions entered, makes omissions harder to locate, and can drift from the brief. Work in bounded passes:

1. select one outline section;
2. provide its promise, evidence, audience need, and word budget;
3. state preservation and invention boundaries;
4. generate one or two alternatives;
5. compare against the brief;
6. edit by hand before moving on.

This sequence keeps authorship visible. You choose claims and evidence; the model proposes expression.

For each section, assemble a small packet:

SECTION JOB
Explain why a decision record should capture rejected options.

READER CONTEXT
Managers think recording the final choice is sufficient.

APPROVED MATERIAL
- Rejected options prevent future teams from repeating old analysis.
- Example: [verified internal example].

REQUIRED QUALIFICATION
Not every minor choice needs a formal record.

BOUNDARIES
No invented statistics, quotations, policies, or outcomes.
Mark missing support as [SOURCE NEEDED].

OUTPUT
Two versions, 140–170 words: direct and example-led.
After each, list every factual claim.

The packet narrows the model’s freedom where accuracy matters while leaving room for rhetorical options. If you need a quotation, supply the exact quotation and source. Never ask the model to reconstruct wording from memory.

Begin with a plain version that makes reasoning explicit. Ask for topic sentences, evidence placement, and qualifications before vivid openings, metaphors, or stylistic flourish. Decoration can make a weak claim feel persuasive without making it sound.

A useful paragraph usually contains a controlled sequence:

• a claim or purpose;
• explanation;
• evidence or concrete example;
• qualification where needed;
• a link to the next idea.

Not every paragraph requires all five, but checking these roles exposes unsupported assertions and filler. Ask the model to label roles during review, then remove labels from the prose.

When the evidence is thin, keep the language proportionate. “Our sample suggests” differs from “research proves.” “May reduce” differs from “will eliminate.” Models often strengthen weak modal language because certainty sounds cleaner. State the required level of certainty in the packet and inspect it afterward.

Two drafts are useful only if they test meaningful choices. Request a direct opening and a scenario opening, or a concise explanation and an analogy-supported explanation. Do not generate ten versions and choose the most polished by instinct.

Compare alternatives using criteria from the brief:

• Does the reader learn the central point early?
• Is every claim supported by approved material?
• Does the version respect the word budget?
• Is uncertainty preserved?
• Does it perform the section promise?
• Does it create the intended transition?

Often the best result combines a structure from one version with wording you write yourself. Copying an entire generated passage is not the goal.

Section-by-section drafting can create repeated definitions, inconsistent terms, or mismatched tone. Maintain a compact continuity sheet:

• canonical terms and definitions;
• claims already established;
• examples already used;
• decisions about point of view and tense;
• unresolved placeholders;
• promised ideas that later sections must fulfill.

Before drafting a new section, provide only the relevant continuity notes. Dumping every previous paragraph into the context can distract the model and expose unnecessary information. After drafting, update the sheet yourself.

Transitions deserve special review. A fluent connector such as “therefore” may imply a logical relationship the evidence does not support. Test whether the prior section actually justifies the next. Replace decorative transitions with accurate ones.

Keep source notes separate from generated text. Record where facts, quotations, and examples came from. A model-generated claim list can help you inspect a section, but it cannot verify itself. If the output introduces a plausible example, delete it or label it until independently sourced.

Follow rules for attribution, confidentiality, academic integrity, and acceptable AI assistance. Some settings require disclosure; some prohibit generated prose; some allow brainstorming but not drafting. The tool’s capability does not determine permission.

Never paste restricted material into an unapproved system. Minimize personal or confidential details and use placeholders where possible. High-stakes claims need qualified review.

A first draft does not need elegant sentences. It is ready for editing when it fulfills the section promise, has a visible line of reasoning, includes the required evidence and qualifications, and contains no hidden unknowns. Keep [SOURCE NEEDED], [DECISION NEEDED], and [EXAMPLE NEEDED] visible. Placeholders are honest work queues.

Finish a coherent rough draft before optimizing every sentence. Premature line editing can waste effort if later evidence changes the argument. The next pass should diagnose clarity and structure separately from voice.