---
name: research-report-writing
description: >
  The prose craft of a research report: paragraphs that carry an argument,
  finding sentences that are precise about base and population without becoming
  unreadable, numbers and quotes integrated so they evidence rather than
  decorate, interpretation marked consistently, and the tells of AI-generated
  research writing detected and removed. Use for "write up the findings",
  "draft the chapter", "make this section read properly", "rewrite this so it
  is not so dense", "the writing is flat", "this reads like AI wrote it".
category: 11 Reporting and Storytelling
ref: 11.03
tier: 1
inherits: [K2, K3, K4, K5]
---

# Research Report Writing

## 1. One-line description
Turns a structured, evidenced argument into research prose that is precise about what was measured, readable by someone who was not on the project, and honest about the point at which reporting becomes arguing.

## 2. What this skill is used for

**The research problem it solves.** Research writing has a specific and permanent problem: **precision and readability pull against each other, and neither can be sacrificed.** Every qualification that makes a sentence defensible (the base, the population, the period, the question wording, the uncertainty) makes it harder to read. Every simplification that makes a sentence land removes something that made it true. Writers resolve the tension by picking a side, and both sides produce bad reports.

Picking precision produces the report nobody finishes: every claim carrying four parentheses, every verb hedged, and a reader who cannot tell after three pages whether anything was found. Picking readability produces the report that says more than the study did: the qualifications migrate to the appendix, the associations become drivers, and the prose gets more confident exactly where the evidence gets thinner. The professional resolution is neither. It is a set of specific techniques for placing precision where it does not block the sentence, and this skill is those techniques.

There is now a second problem on top of the first. AI-generated research prose has a recognisable signature: paragraphs of uniform length, three-part lists everywhere, transitional sentences that connect nothing, restatement presented as synthesis, and a confident summary attached to a weak finding. It reads fluently and says less than it appears to. Readers are increasingly good at spotting it, and a client who spots it stops trusting the analysis underneath, which is a reputational cost out of all proportion to the writing itself.

**Where it sits.** Reporting. Downstream of narrative, architecture and compilation, all of which decide what is said and in what order. This skill decides how each sentence is built. Upstream of design and QA.

**Typical use cases.**
- Drafting the findings chapters of a report from an evidenced structure.
- Rewriting a section that is accurate and unreadable, or readable and overclaimed.
- Bringing several authors' chapters into one register without changing their claims.
- Writing about a finding that is genuinely uncertain, without the prose collapsing into hedge.
- Removing the signature of AI-generated prose from a draft, whether or not AI wrote it.
- Writing a chapter where the evidence is qualitative and the temptation is to let the quotes do the arguing.

**Who uses it.** Researchers at every level who write, which is all of them; research directors editing other people's chapters; consultants who are judged on the document as much as the analysis; anyone whose draft has been returned with the comment that it is dense, or thin, or does not sound like them.

## 3. When to use it

- A structured, evidenced plan exists and the chapters have to be written.
- A draft is accurate but unreadable, and the fix must not remove its accuracy.
- A draft is readable but has quietly acquired claims the evidence does not carry.
- Several people wrote sections and the report reads as several people.
- A chapter rests on qualitative material and the quotes are currently doing the work the prose should do.
- A finding is uncertain and the writing has to convey that without becoming unusable.
- A draft has the signature of machine-generated prose and needs it removed.
- A section opens and closes flatly, so the reader cannot tell what it established.

## 4. When NOT to use it

- **The structure has not been decided.** Prose written before the argument exists produces sections that are individually well-made and collectively pointless, and rewriting them later costs more than planning would have. The order and the chapter logic are **12.01 Research Report Architecture**; the argument itself is **11.01 Research Narrative Development**. Come here with both.
- **The evidence has not been assembled and mapped.** Writing is where unsupported claims enter a report most easily, because a sentence that needs a connective supplies one. Where inputs are heterogeneous and no evidence map exists, that assembly job is **12.03 Research Report Compilation**, which owns inventory, source codes, the conventions register and the evidence map. **The boundary, stated in both directions: 12.03 assembles evidence into a structure and owns the conventions that hold across sections; 11.03 is the prose written inside that structure.** A single-source report from one clean analysis by one analyst can come straight here, because the compilation overheads earn nothing.
- **The analysis is incomplete.** Prose fills gaps more smoothly than any other research activity, because fluency and connection are the same operation. Do not write toward a cut that has not been run.
- **The deliverable is a summary.** The executive summary is a distillation with its own selection rules, cap on conclusions and confidence requirements, and it is **11.02 Executive Summary Development**. This skill governs the construction of any given sentence, including the ones in a summary, but not what a summary contains.
- **The deliverable is slides.** Slide headlines, one idea per slide and the sequencing of a live argument are **11.05 Research Presentation Development**. Report prose written onto slides produces the document nobody can present, and slide fragments assembled into a report produce prose with no connective tissue.
- **The problem is visual.** Where the section is unreadable because a table is doing a chart's job, the fix is **11.04 Data Visualisation and Chart Selection**, not more sentences. Prose that explains a chart the reader cannot parse is compensating for the wrong thing.
- **The request is to make a weak finding sound stronger.** This is the commonest writing request and it must be declined as written, per K4 §9. Offer instead: state the finding at its confidence, say what it would take to establish it, and let its decision relevance carry it. Writing is where confidence inflation actually happens, sentence by sentence, and the request to "make it punchier" is usually a request to inflate.
- **The report is finished and needs checking rather than writing.** Do not rewrite as a way of reviewing: it loses the original's decisions and puts a second version into circulation. That is **12.06 Research Report QA**.

## 5. Required inputs

**Required.** Without these the skill cannot run. If absent, ask. If no answer is available and work must proceed, state the assumption at the point where it bites, per K5 §5.

- **The section plan**: what each section establishes, in what order, and what it hands to the next. Prose cannot supply a structure it was not given without inventing one.
- **The findings to be written**, each as a claim with its source reference, base description, base size and confidence level. A finding without a base cannot be written into a defensible sentence.
- **The conventions in force**: terminology, base definitions, rounding, comparison rules, confidence vocabulary, quote editing convention. Where a conventions register exists from **12.03**, use it. Where none exists, write one before drafting, because inconsistency is created in the first chapter and discovered in the fourth.
- **The verified quote set**, with participant identifiers and segment attributions checked, per **07.04**. Quotes are not selected during writing; that is how attribution errors and convenient selections happen.
- **The audience and their vocabulary.** A report for a clinical audience, a board, an operational team and a public consultation are four different registers, and register is a decision, not a default.

**Optional, and what each one adds.**

- **The narrative statement** from **11.01**: supplies the section openings and closings, because each section's job in the argument is what its first and last paragraphs have to say.
- **A previous report for the same client**: fixes register and terminology without a discussion, and reveals what the audience is already used to reading.
- **The client's own documents** (strategy papers, operational reports): supply the vocabulary the reader uses, which is usually not the vocabulary of the questionnaire.
- **The full transcripts** behind the selected quotes: allow a quote's context to be checked when the surrounding prose is written, which is where quotes most often acquire a meaning they did not have.
- **A named editor**: makes the single-voice pass possible, and one editor with a register beats three writers with a style guide.

## 6. Questions to ask before starting

1. **Who reads this, and what do they already know?** Determines register, how much explanation each concept needs, and which terms can be used unglossed. *Default if unanswered:* write for an intelligent reader in the client organisation who was not involved in the project.
2. **What is this section's job in the argument?** Determines its opening claim and its closing hand-off. *Default:* take it from the section plan; if the plan does not say, the plan is incomplete and prose will not fix it.
3. **What conventions are in force for terminology, bases, rounding and confidence words?** Determines consistency, which is the mechanical part of good writing. *Default:* write the register yourself before drafting and apply it without exception.
4. **How will this be read: on paper, on screen, in a meeting, or extracted in pieces?** Determines paragraph length, how self-contained each sentence must be, and whether a claim can rely on the paragraph above it. *Default:* assume paragraphs will be quoted out of context.
5. **Is there a house style or a client style guide, and does it conflict with anything required here?** Determines what can be negotiated. *Default:* follow house style on everything cosmetic and never on anything that would remove a base, a caveat or a confidence marker.
6. **Which findings in this section are weak, and how is that going to read?** Determines where the writing work actually is, because weak findings written well are the hardest thing in the report. *Default:* identify them before drafting so their sentences are built deliberately rather than smoothed.

## 7. Step-by-step methodology

**The governing tension.** Precision and readability pull against each other in every sentence of a research report. The resolutions below are all versions of one principle: **put the claim in the sentence and the machinery next to it.** A reader should be able to read the argument at speed and drop into the evidence when they want to check.

**Step 1. Set the register before drafting.** Register is a decision with two failure modes at either end. **Academic hedging**: "it may tentatively be suggested that there is some indication that respondents could potentially be experiencing a degree of difficulty". Four hedges on one claim, which communicates nothing about how strong the claim is because everything gets the same treatment. **Consultancy inflation**: "customers are crying out for a fundamentally reimagined experience". Emphatic, unfalsifiable, and evidentially empty. The professional register sits between: **declarative sentences, hedged only where the evidence is genuinely uncertain, and hedged differentially so the hedge carries information**, per K3 §1. Fix three things now: the person and tense (a consistent choice, applied throughout), the term for each concept, and the confidence vocabulary from K3 §4. *Correct result:* a half-page register note, written before the first chapter, that a second writer could apply.

**Step 2. Build every paragraph as a unit of argument.** The default research paragraph has four moves: **claim, evidence, qualification, consequence.** Not every paragraph needs all four, but a paragraph with none of them is description, and a run of them is what makes a section feel like it is going somewhere.

> Fee levels are not what stops smallholders enrolling. Cost was cited by 14% as the main barrier, behind travel time and paperwork (Q11, all non-enrolled, n=604). The measure is stated reason rather than observed behaviour, so it captures what people are willing to say. The consequence for programme design is that a fee waiver addresses the fourth-largest barrier.

Claim first, so the reader knows what the numbers are for. Evidence with its reference, so it can be checked. Qualification adjacent to the claim rather than in a footnote, per K4 §4.3. Consequence, so the paragraph has done something. The commonest defect is the paragraph that opens with the evidence and ends with the claim, which forces the reader to hold numbers with no idea what they are for. *Correct result:* a paragraph whose first sentence, read alone, tells the reader what the paragraph establishes.

**Step 3. Write the finding sentence.** A finding sentence must be precise about **base, population and period** without becoming a legal document. Three techniques do most of the work.

**Put the specificity in the reference, not in the clause.** Compare "among those respondents who had used the service in the previous twelve months and who completed the full battery, 38% reported difficulty" with "38% of recent users reported difficulty (Q14, used in past 12 months, n=411)". The second is more precise, because the reference is unambiguous, and it is readable, because the clause is short.

**Name the population in the noun, not in a subordinate clause.** "Recent users", "non-enrolled smallholders", "callers who reached an adviser" are populations carried by two or three words. This is where terminology consistency pays: a lexicon built once means every population has a short name.

**Say what was measured when the measure is not what the reader will assume.** Stated intention, recalled behaviour, attributed reason and satisfaction with a specific event are all routinely read as something stronger. One clause fixes it: "asked to choose their main barrier, 14% chose cost", which is honest about the fact that a forced choice among listed options is not a spontaneous account. Do this where the distance matters, per K3 §3.4, and not everywhere, or you have chosen hedging.

*Correct result:* a sentence a critic could check and a stakeholder could read aloud.

**Step 4. Integrate numbers so they are legible.** Numbers stop being read at roughly three per sentence, and a paragraph of percentages becomes a table the reader has to build themselves. Rules that hold: **one number per sentence carries the claim, and the rest go in a reference or a chart.** Round to the precision the method supports, per K3 §4.4, so whole percentages unless there is a reason. Give a comparison its comparator in the same sentence ("38% of recent users, against 22% of lapsed users") rather than in the next one. Use a verbal expression only with its exact figure on the same page, per K2 §7, so "nearly two thirds" is permitted when 62% appears in the reference and prohibited when it does not. Never let a number arrive without its base: if the sentence cannot carry it, the reference must. And where a difference is untested, say so in three words ("untested") rather than choosing a comparative that implies a test, per K4 §3.1. *Correct result:* a paragraph a reader can follow at speed, where every number is checkable and none is doing rhetorical work.

**Step 5. Integrate quotes so they evidence rather than decorate.** A quote earns its place by showing something the prose cannot: a mechanism, a way of framing, a word the participant chose that the researcher would not have. A quote that restates the sentence above it is decoration, and decoration is expensive, because it takes the space a second piece of evidence would have taken.

The working pattern is **claim, quote, what the quote shows.** The prose makes the claim; the quote provides the instance; a short sentence says what to notice, because a quote left to speak for itself is interpreted by the reader in whatever way they were already inclined to. Attribution carries a participant ID and the segment, per K2 §4.2, and never a fabricated descriptor. Prevalence goes with the quote where a theme is claimed: "14 of 22 participants described the same sequence", so the reader can tell an illustration from a pattern. Editing follows K4 §2.3 and the convention is stated once in the report. Two prohibitions specific to writing: **never edit a quote while editing the prose around it**, which is where quotes acquire the report's voice, and never introduce a quote with a sentence that says more than the quote does. *Correct result:* every quote in the section could be removed and the claim would still stand, but the section would be poorer.

**Step 6. Mark interpretation, consistently.** K2 §3.2 gives the signal words for each level, and the requirement of this skill is that they are used **consistently and exclusively**: "suggests" always means moderate confidence and never appears attached to a strong finding; "shows" is reserved for what the data shows; "what appears to be happening" marks an insight and not a finding. Once fixed, the reader learns the code within a chapter and can navigate the report by it. The rule that catches most errors: **a finding and an interpretation never share a sentence without a signal word between them.** "Satisfaction fell because the fee changed" hides the join. "Satisfaction fell 11 points over the period. The timing points to the fee change as the most likely explanation, though this design cannot separate it from the seasonal pattern" shows it, per K2 §3.2. The first sentence in which judgement enters must be identifiable, per K2 §3.3, and in a chapter that is usually the third or fourth sentence, not the last. *Correct result:* a reader can mark every sentence in the section as reporting or arguing, without ambiguity.

**Step 7. Write about uncertainty without becoming unreadable.** Four techniques, in order of preference. **State the finding, then the limit, in two sentences** rather than one qualified sentence: the claim lands, the limit lands, neither is buried. **Quantify the uncertainty where it is quantifiable** ("between roughly a quarter and a third, on this base") rather than hedging the verb, because a range is information and "may be around" is not. **Name the alternative explanation** rather than gesturing at caution: "the alternative, that these customers were already dissatisfied, is not ruled out by this design" tells the reader something they can weigh. **Say what would settle it**, which converts a limitation into a next step. And apply the discipline that makes all four work: **hedge differentially.** If every claim in the chapter is hedged, the reader cannot find the strong ones and discounts everything equally, per K3 §1, which is the more common AI failure and the less often recognised one. *Correct result:* a reader can rank the section's claims by firmness from the language alone.

**Step 8. Write section openings and closings that carry the argument.** An opening paragraph states what the section establishes, in a sentence that could be disagreed with. Not "this section examines enrolment barriers" (a label, and the heading already said it) but "enrolment is not blocked by cost, and the two barriers that do block it are both administrative". A closing paragraph states what has been established and hands to the next section, ideally by raising the question the next section answers. Openings and closings are where the argument is visible, and they are the parts a skimming reader actually reads. A section whose opening is a label and whose closing is a summary of itself has left the argument to the reader. *Correct result:* the openings and closings, read alone in sequence, reconstruct the report's argument.

**Step 9. Run the de-AI pass.** These tells appear in machine-generated research prose and in tired human prose, and readers now recognise them. Detect and remove each one.

- **Uniform paragraph length.** Every paragraph three or four sentences, every section the same length. Real argument is uneven: some points take two sentences and some take twelve. Vary deliberately, and let the shortest paragraph in a section be the one carrying its most important claim.
- **Tricolon everywhere.** Three-part lists in sentence after sentence ("faster, simpler and more transparent"). One is rhetoric, five in a chapter is a tic, and the third item is usually invented to complete the pattern. Cut to the items that are actually evidenced, whether that is one, two or four.
- **Hollow transitional sentences.** "This finding has important implications for how the organisation approaches the challenge." It connects nothing, asserts nothing, and can be deleted with no loss. Test every transition: does it state a relationship? If not, delete it; if it does, it needs evidence, per K4 §5.1.
- **Restatement disguised as synthesis.** A paragraph opening "taken together, these findings suggest" and then restating the findings with no addition. Genuine synthesis says something none of the findings says alone. If you cannot name what it adds, delete the paragraph.
- **The confident summary of a weak finding.** A small-base or single-item result, written in the same declarative register as the strong ones, usually at the end of a section where summary voice takes over. Check every summary sentence against the confidence of the finding beneath it, per K3 §7.
- **Two further tells worth removing.** The negative-parallel construction ("this is not about X, it is about Y"), which manufactures profundity by inventing a position nobody held; and the long dash used as an all-purpose connector in place of deciding whether the relationship is causal, exemplary or contrastive.

*Correct result:* prose whose shape follows the argument rather than a template, in which every transition, list and summary sentence earns its place.

## 8. Analytical framework

The paragraph, which is the unit of research writing:

    Claim → Evidence (with reference) → Qualification → Consequence

And the section, which is the same structure one level up:

    Opening claim → Paragraphs that establish it → Closing that hands to the next question

**Applying it.** Write the claim first, always, at both levels. The commonest structural defect in research prose is the inductive paragraph, which walks the reader through the evidence and reveals the point at the end. It reads as suspense to the writer and as fog to the reader, who has been holding numbers without knowing what they are for. Answer-first at the paragraph level is the same discipline as answer-first at the report level, applied where most of the words are.

**The precision ladder.** For any claim, precision can live in four places, and the craft is choosing the lowest one that is sufficient:

1. In the sentence, where it blocks reading.
2. In a short reference after the sentence, which is where most of it belongs.
3. In the chart's base note, for anything the sentence does not depend on.
4. In the appendix, for anything a reader would only want when checking method.

Anything that changes how the claim should be read stays at levels 1 or 2, per K4 §4.3. A base of 42 goes next to the number, not in the appendix.

**The two-pass method.** Draft for argument, edit for precision. Attempting both at once produces sentences that are hedged into incoherence during composition. In the first pass, write the claims and mark evidence slots. In the second, fill the references, check every number against source, and verify every quote. **Never combine the second pass with a style edit**, because that is precisely where quotes get smoothed and bases get dropped.

## 9. Output format

Research prose has no single template, but a section has a recognisable shape.

**Section opening.** One paragraph. The claim the section establishes, and why it matters to the argument. No preamble about what the section will cover.

**Body paragraphs.** Each a unit of argument: claim, evidence with reference, qualification, consequence. Length varies with the weight of the point. Every proportion carries its base; every comparison carries its comparator; every untested difference says so.

**Evidence integration.** Numbers at no more than one load-bearing figure per sentence, with the rest in references or a chart. Quotes in the claim, quote, significance pattern, with participant ID, segment and prevalence where a theme is claimed.

**Interpretation.** Marked with the K2 §3.2 signal words, used consistently, and never in the same sentence as the finding it rests on without a signal between them.

**Uncertainty.** Stated in the K3 §4 register for the confidence level, with the alternative explanation or the specific limit in the same passage for anything moderate, and the validation named for anything low.

**Section closing.** One paragraph. What has been established, and the question the next section answers.

**When the evidence is thin.** The prose gets shorter, not vaguer, and this is the single most useful discipline in the section. A finding on a base of 40 gets one sentence with its base and a directional reading, not three paragraphs of interpretation. A section with two findings is two paragraphs and a closing, not a chapter padded with context. Where a section exists in the plan and the evidence for it does not, say what was found and what was not, per K3 §5.2, and do not write connective prose across the gap: connection is the easiest thing to write and the hardest thing to detect afterwards, per K4 §1. The tell that this discipline has failed is a section whose word count is stable while its evidence base varies.

## 10. Quality checks

Run on every section before it is presented. Sits on top of K4 §8.

1. Does every paragraph's first sentence state what the paragraph establishes?
2. Does every proportion carry a base description and base size, in the sentence or its reference?
3. Is every population named precisely, and does the name match the questionnaire routing?
4. Does any sentence claim causation, behaviour from stated preference, change over time, or reach beyond the sample, per K4 §3?
5. Is every untested difference described as untested rather than with a comparative that implies a test?
6. Are the K2 §3.2 signal words used consistently, with one confidence register per word?
7. Can every sentence be marked as reporting or arguing, and does no sentence contain both without a signal between them?
8. Does every quote carry a participant ID and segment, match the source word for word, and show something the prose does not?
9. Is prevalence stated wherever a quote is used to evidence a theme?
10. Is one term used for each concept throughout, matching the conventions register?
11. Is hedging differential, so a reader can tell the strong claims from the weak ones by the language?
12. Does every section open with a claim rather than a label, and close by handing to the next question?
13. Does any transitional sentence assert a relationship that no evidence supports?
14. Does any paragraph beginning "taken together" add something none of its findings says alone?
15. Are paragraph lengths uneven in a way that follows the argument, rather than uniform?
16. Was any quote, base or number altered during a style edit?

## 11. Common failure modes

| Failure | How to recognise it | How to prevent it |
|---|---|---|
| **The inductive paragraph** | Evidence first, point at the end; the reader holds numbers without knowing why | Claim first, always. Step 2 |
| **Hedge saturation** | Every verb qualified; the reader cannot rank any claim | Hedge differentially per K3 §1. Strong findings are stated declaratively |
| **Consultancy inflation** | Emphatic, unfalsifiable language: "crying out for", "fundamentally reimagined" | Every sentence must be checkable against a finding. If it cannot be false, delete it |
| **Number fog** | Four or five percentages in a paragraph, none obviously load-bearing | One load-bearing number per sentence; the rest to references or a chart |
| **Decorative quoting** | The quote restates the sentence above it | Claim, quote, significance. A quote must show what the prose cannot |
| **Quote drift on rewrite** | Quotes read fluently and in the report's voice | Never edit a quote while editing the prose around it. Verify after any style pass |
| **The vanishing base** | A percentage in the prose with its base only on the chart two pages later | Base travels with the number, per K3 §7 |
| **Signal word inconsistency** | "Suggests" attached to a strong finding in one chapter and a hypothesis in another | Fix the K2 §3.2 register once and apply it without exception |
| **Uniform paragraph length** (AI tell) | Every paragraph three or four sentences regardless of what it carries | Let length follow the weight of the point. Step 9 |
| **Tricolon tic** (AI tell) | Three-part lists in sentence after sentence; the third item unevidenced | Cut to the items the evidence supports, whatever the number |
| **Hollow transition** (AI tell) | "This has important implications for how the organisation approaches..." | Every transition states a relationship or is deleted, per K4 §5.1 |
| **Restatement as synthesis** (AI tell) | "Taken together, these findings suggest" followed by the findings again | Name what the synthesis adds. If nothing, delete the paragraph |
| **Confident summary of a weak finding** (AI tell) | A small-base result written in the same register as the strong ones, usually at a section's end | Check every summary sentence against the confidence of the finding beneath it |
| **The label opening** | Sections beginning "this section examines..." | Open with the claim the section establishes |
| **Padding a thin section** | Word count stable while evidence base varies | Short sections are correct when the evidence is short, per K4 §1 |

## 12. AI guardrails

Skill-specific only. Universal prohibitions are inherited from K4. Quote editing follows K4 §2.3; signal words follow K2 §3.2; confidence registers follow K3 §4.

1. **Never write a transitional or connective sentence that asserts a relationship no evidence establishes.** Fluency is the mechanism by which unsupported claims enter reports, because the sentence that joins two paragraphs is the one nobody checks.
2. **Never adjust a quote, a base description or a number during a style edit.** Voice consistency applies to the analyst's words and never to the evidence.
3. **Never hedge a high-confidence finding, and never state a low-confidence finding declaratively.** Both are calibration failures, and the first is as damaging as the second because it destroys the reader's ability to discriminate.
4. **Never write a sentence containing both a finding and its interpretation with no signal word between them**, per K2 §3.2.
5. **Never use a comparative that implies a statistical test where none was run**, per K4 §3.1. "Higher" with both bases shown is permitted; "significantly higher", "notably higher" and "clearly ahead" are not.
6. **Never let a rounded verbal expression appear without its exact figure and base on the same page**, per K2 §7.
7. **Never introduce a quote with a claim larger than the quote supports**, and never present a single quote as evidence of a theme without its prevalence.
8. **Never produce a paragraph whose function is to fill a section for which the evidence is thin.** Write the shorter section, per K4 §1.
9. **Never write a synthesis paragraph that restates its inputs.** If you cannot name what it adds beyond the findings it draws on, it adds nothing.
10. **Never smooth an inconvenient finding into a subordinate clause.** A finding that cuts against the section's claim gets its own sentence, at minimum, per K4 §4.1.

## 13. Best-practice principles

1. **Precision and readability are both non-negotiable, so the craft is placement.** Claim in the sentence, machinery next to it. Almost every technique in this skill is a version of that.
2. **Write the claim first, at every level: sentence, paragraph, section, report.** Suspense is for fiction. A reader holding evidence without a claim is doing your work.
3. **The hedge must carry information.** Hedging everything is the same as hedging nothing, and it is the more common failure in machine-assisted writing.
4. **A quote is evidence, not seasoning.** If the paragraph would say the same thing without it, it is decoration and it is occupying space that a second piece of evidence needed.
5. **Consistency is invisible and its absence is not.** One term per concept, one signal word per confidence level, one base convention. A reader who notices your terminology has stopped reading your argument.
6. **The base travels with the number, into every sentence it appears in.** A percentage that has lost its base has become an opinion.
7. **Short sections are a sign of discipline, not thinness.** The report should be as long as the evidence, and a section that ran short because the finding was small is correct.
8. **Openings and closings are the report a skimming reader actually reads.** Write them last, when you know what the section established, and make them carry the argument.
9. **Vary paragraph length on purpose.** Uniformity signals that the shape came from a template rather than the argument, and readers now read it that way.
10. **Read it aloud.** Sentences that cannot be spoken are usually sentences carrying three qualifications that should have been two sentences.
11. **The sentence you are most pleased with is the one to check.** Elegance in research prose is frequently the sound of a qualification having been dropped.
12. **Fluency is not neutrality.** Good prose makes any claim more believable, including the unsupported ones, which is why the evidence check runs separately from the style pass and never inside it.

## 14. Worked example

*Fictional scenario, used for illustration only. The organisation, findings, figures and quotes below are invented for the purpose of demonstrating method.*

**INPUT.** An international NGO has completed a study into low enrolment in a subsidised agricultural insurance scheme for smallholder farmers. The chapter to be written covers barriers to enrolment. Inputs: a section plan (the chapter must establish that cost is not the primary barrier and that the two barriers that dominate are administrative), four findings with references, a verified quote set from 22 depth interviews, and a conventions register fixing "smallholder", "enrolled", "non-enrolled" and whole-number rounding. A first draft exists, written by a colleague, and it is accurate and unreadable.

**PROCESS.**

*Step 1.* The draft's register is diagnosed as academic hedging: 11 of its 14 claims carry a modal verb, including the two strongest. Because everything is hedged, a reader cannot tell that the cost finding rests on a base of 604 and the trust finding on 22 interviews. The register note fixes declarative statement for the two strong findings and reserves "appears to" for the interview-based reading.

*Step 2, and the rewrite.* The draft's opening paragraph reads: "This section examines the range of factors that may influence enrolment decisions among smallholder farmers." It is a label. It becomes: "Enrolment is not blocked by cost. The two barriers that dominate are both administrative, and both are within the programme's control."

*Steps 3 and 4, and the judgement call.* The draft sentence reads: "Among those respondents who reported that they had not enrolled in the scheme and who were asked about barriers, a total of 14.3% indicated that cost was the main factor preventing enrolment." Three problems: the specificity is in the clause rather than the reference, the decimal implies a precision the method does not support per K3 §4.4, and "indicated that cost was the main factor" is a forced choice among listed options presented as a spontaneous account. It becomes: "Asked to choose their main barrier from a list, 14% of non-enrolled smallholders chose cost, behind travel time and paperwork (Q11, non-enrolled, n=604)." Shorter, more precise, and honest about the measure.

*Step 5, and a second judgement call.* The draft carries a quote after the cost finding: "It is expensive for us." It restates the sentence above it and evidences nothing the number did not. It is cut. In its place goes a quote that shows a mechanism the survey could not: "They asked for the land papers. My father has them and he is in the city." Followed by the significance sentence: "Documentation requirements assume a household structure that many enrolments do not have (14 of 22 participants described a document held by someone else)." Prevalence is stated so a reader can tell a pattern from an instance.

*Step 7.* The trust finding rests on 22 interviews and is an interpretation. It is written at moderate confidence with its alternative in the same passage, and with what would settle it: "The most likely reading is that the paperwork is read as a test of eligibility rather than a formality. The alternative, that it is simply burdensome, is not ruled out, and the two imply different fixes. A short enrolment experiment varying only the document requirement would separate them."

*Step 9, the de-AI pass.* Three removals. A transitional sentence, "these findings have significant implications for how the programme approaches enrolment", asserts nothing and goes. A tricolon, "faster, simpler and more transparent", loses its third item, which nothing in the evidence supports. A closing paragraph beginning "taken together, these findings suggest" restates the three findings and adds nothing; it is replaced by a closing that names what has been established and raises the question the next section answers: whether the two administrative barriers can be removed without weakening the eligibility check.

**OUTPUT.** A chapter of nine paragraphs of visibly uneven length, opening on a claim and closing on a question, in which every proportion carries its base, two quotes carry participant references and prevalence, the interpretation is marked and stated at moderate confidence with its alternative and its test, and no transitional sentence asserts a relationship the evidence does not support. Word count is 40% below the draft it replaced.

## 15. Advanced usage

**Editing another writer's chapter into the register.** Separate the two operations completely. First, a claims pass: read for what each paragraph asserts, and mark anything unsupported, over-hedged, or carrying an interpretation without a signal. Second, a language pass on the marked sentences only. Rewriting whole paragraphs for style is how a report acquires claims nobody made and loses references nobody noticed. Where a rewrite touches a sentence containing a quote or a number, re-verify against source afterwards, without exception.

**Writing for a hostile or technical reader.** Where the report will be read by someone looking for a flaw (a regulator, an opposing expert, a peer reviewer), raise the precision ladder by one level throughout: what would normally sit in a reference goes in the sentence, and what would sit in the appendix goes in a reference. Accept the loss of fluency, and compensate by shortening. A dense, short document reads better to that reader than a fluent, long one.

**Multi-language and translated reporting.** Prose written for translation should avoid constructions that carry meaning through rhythm, and idioms that will be rendered literally. More importantly, quotes translated from another language carry an additional evidential step: mark them as translated, retain the original where the wording is load-bearing, and route cultural interpretation to a human reader with the relevant context, per K5 §2.2.

**Writing where the finding is uncomfortable.** Write it in the plainest available sentence, with its full evidence and its confidence. Softened prose around a difficult finding reads as uncertainty, and the finding then gets attacked on its evidence rather than debated on its implications. Where the decision to include it is a materiality judgement rather than an evidential one, mark it per K5 §2.1 rather than resolving it in the wording.

**When the standard approach does not fit.** Some audiences genuinely require a narrative or first-person register: an ethnographic account, a participatory research report written back to a community, a document whose credibility rests on the researcher's presence being visible. The evidential rules do not relax, but their placement changes: references move to endnotes, prevalence moves into the prose ("most of the people I spoke to"), and the report states its convention once at the front. What never changes is that the reader can tell observation from argument, per K2 §3.3.

## 16. Skill chain

**Recommended previous skills:**
- **11.01 Research Narrative Development.** Hands over the spine and the section-level claims, which become the openings and closings that carry the argument.
- **12.01 Research Report Architecture.** Hands over the chapter plan, page budgets and the job of each section.
- **12.03 Research Report Compilation.** Where inputs are heterogeneous, hands over the evidence map, source codes and conventions register that this skill writes inside. Compilation owns assembly and conventions; this skill owns the prose.
- **07.04 Quote and Evidence Extraction.** Hands over verified quotes with participant IDs, segments and prevalence, so quotes are not selected during drafting.

**Recommended next skills:**
- **12.02 Research Report Design.** Takes the written chapters and applies layout, typography and the visual system without altering claims.
- **11.02 Executive Summary Development.** Takes the checked chapters and distils them, using this skill's sentence craft under its own selection rules.
- **12.06 Research Report QA.** Takes the draft and checks it independently, which the writer's own passes do not replace.
- **13.03 AI Output Verification.** Run against any draft where AI assisted the writing, specifically for unsupported connective claims and confidence inflation.

**Runs well alongside:**
- **11.04 Data Visualisation and Chart Selection**, because a section that will not read is sometimes a section whose chart is doing the wrong job.
- **13.04 Bias Detection**, run against a draft for framing effects introduced in the writing rather than the analysis.
- **K2 §3**, which defines the signal words this skill applies, and **K3 §4**, which defines the confidence registers, and **K4 §2.3**, which defines the permitted quote edits.

---
A Yazi Supplied Skill and resource.