Writing Scannable Prose
Overview
Treat the document as a dependency graph, not a pile of words to shorten: a cut is safe only when nothing depends on what got removed. Watch for orphaning — a claim whose scope, evidence, or antecedent got cut out from under it, leaving a reader confidently wrong, not visibly confused. Verbosity — words carrying no information — is real, but secondary and easy.
A document gets read twice: as a visual surface (headings, bold, layout) and as a linear stream where none of that survives — screen readers, TTS, plain-text extraction. Only markup-level structure survives the second read.
Gate: fires on a document carrying claims, reasoning, or procedure — README, guide, ADR, PR description, changelog, or similar — not on line count. A three-line changelog entry is in scope; a one-line typo-fix PR body is not.
Authoring inputs
Decide two things once, before drafting:
| Decision |
Consumed by |
Controls |
| Reader, and what they decide next |
F1, F2, F4 |
Whether a section survives, what counts as author-facing detail |
| Passage procedural or explanatory |
A4 |
How hard cutting may go within that passage |
Diátaxis mode (tutorial / how-to / reference / explanation) is not an authoring input — A6 alone consumes it, relocating a paragraph misfiled for its section's mode.
Revision protocol
- Derive F1's line. Check sibling artifacts of the same genre, the template, and whether the content already lives canonically; state the result:
Reader: <who> — Decision: <what they do next>. No signal resolves it → say so.
- Run F2 against every section against F1's line: does its absence change the named decision? Mark keep or remove. Then, for each section marked keep, ask what F4 routing would leave it holding and re-test that against F1's line: a section hollowed below materiality is a residue stub reached by hollowing instead of by bad removal, so mark it remove too. This is an assessment, not the edit — F4 routes in step 5, on sections that survive. F1 is unresolved whenever either field is missing — a reader named with no decision, or a decision with no reader — and an unresolved F1 means F2 doesn't run; every section defaults to keep. The same applies to every check that reads F1's line: F4 and A6 both route by what the named reader needs, so an unresolved F1 skips them too and their candidate material stays put. A–E otherwise run normally. Name each skip as a non-blocking escalation — not a second blocking point, since none of it halts the pass.
- For every section marked remove — by either route — run F3: name the section that owns each load-bearing item — and it must be a section surviving this batch, or the item dies with its destination; no owning section → find one, or reclassify it as not load-bearing only when it fails F1's line like any other cut, naming the reclassification to the user beside the removal proposal.
- Emit the removal proposal (shape below) in one batch, then stop for an answer before step 5 — the protocol's one blocking point. Every row needs an explicit disposition; any row the answer does not address is rejected, which is the fail-safe direction. This batch covers both routes to removal, so it is the only proposal the pass emits and the only point it waits. A rejection is final for this pass: the section stays and is not re-proposed, which is what makes the protocol terminate. Nothing marked → straight to step 5.
- If removals were proposed, apply the answer: perform the approved F2 removals and their F3 moves, and retain the sections the user rejected. Then, if F1 resolved, run F4 on every surviving section: route author-facing detail to its destination (table below); a deliberately-not-made decision stays. If F1 did not resolve, F4 does not run and its candidate material stays put. F4 and A6 may not empty a section whose removal was rejected — a section gutted to nothing is an F2 removal the user declined, taken anyway. Step 2 already assessed what routing would leave, so no section is re-tested here and the pass does not wait again.
- Run groups A–E on the surviving content — all of them if F1 resolved; all but A6 if it did not, since A6 routes by the reader's next action. Reuse a passage's procedural/explanatory classification from the authoring phase when one was made there; otherwise — the common case, since authoring rarely runs against an existing draft — classify it in place, immediately before A4 fires, passage grain only, never at document or section grain. A call here that can't be self-verified is non-blocking: apply it provisionally and name it in the report.
- Emit the revision report (shape below): everything steps 1–6 changed and why, including the escalations from steps 2 and 6.
Removal proposal shape (step 4's output):
## Proposed section removals
| Section | Why it fails F2 (against the Reader/Decision line) | Load-bearing items | Destination |
|---|---|---|---|
One row per removed section, empty when nothing was marked — including the F1-undecided case, named in the report. An item F3 reclassifies out of load-bearing status is named beside the proposal too, with the F1 line it failed — approving a removal means approving what goes with it.
Escalation modes
Escalation has two modes, and only one of them blocks:
| Mode |
Applies to |
Behavior |
| Blocking |
F2 section removals only |
Proposed in one batch before A–E begins; the pass waits |
| Non-blocking |
An unresolved F1 line |
Not applied: F2, F4 and A6 are all skipped — no section removed, no F4/A6 candidate material moved. The rest of A–E still runs. Each skip is named in the report |
| Non-blocking |
Any unverifiable A–E judgment call |
Applied provisionally, named in the report; the pass doesn't wait |
The asymmetry is reversibility: an A–E edit is visible in a diff, so a wrong call costs a word. A section removal leaves no trace once gone.
The 28 checks
Six groups, one table each. mech is judgment-free, judg reduces to yes/no, prin makes no enforcement claim; precautionary rests on convergence, not a settled result (evidence-and-limits.md).
Group F runs first: compressing a doomed section risks a residue stub — bulk gone, shell remaining, passing every check while the reader still gets a section they didn't want. A stub has two routes: removing a section badly, which F3 guards, and hollowing one that survived, which F4's re-test guards.
F — Does this section belong?
| ID |
Check |
Tag |
| F1 |
Name the reader and decision in one visible line, from repo signals — everything downstream tests against it |
judg |
| F2 |
Section materiality: absence unchanged decision → the section goes — not compressed, not tightened, not reduced to a summary. Removals proposed, not performed |
judg |
| F3 |
Re-home before removing: every load-bearing item lands in the section that owns it, and the proposal names where. Deleting wholesale takes the survivors with it; keeping a compressed shell fails F2 |
judg |
| F4 |
Detail level: within a surviving section, separate what the named reader needs from what the author wanted on record — each recurring author-facing category has a destination that is not this document. What routing would leave is re-tested against F2 before the removal proposal, so a section hollowed below materiality joins that one batch rather than surviving as a shell |
judg |
F4 routes surviving detail to:
| Category |
Destination |
| Review-round history, per-note feedback tables |
The review threads themselves |
| Approaches tried and rejected |
Commit body, or an ADR if it outlives the change |
| How it was verified — transcripts, run counts, methodology |
Test docstrings, or the test itself |
| Tooling and process narrative |
Nowhere; work-in-progress residue |
| A decision deliberately not made, with its rationale |
Stays — changes what a reader concludes |
These categories make F4 actionable, not exhaustive: content outside the table is still tested against F1's line rather than passing by default. Scope guard, applying to every check in this skill, not just F4. An edit to any file other than the document under revision — F4's ADR or commit body, A6's owning document, E1's CI assertion, E2's runnable quickstart — is proposed, not performed, unless that file is already in the task's scope. Read such files freely; the material stays put until the move is authorized. Non-blocking: it routes to the report like any other non-blocking item, not a second checkpoint.
A — What a cut may touch
| ID |
Check |
Tag |
| A1 |
Name the operation before editing: removing a word, a claim, or a qualifier that scopes a surviving claim |
judg |
| A2 |
Qualifiers (hedges), scope conditions, sample sizes, exceptions, platform caveats stay in the same visual unit as their claim — never demoted to a footnote or a later bullet. Tighten wording, not the fact |
judg |
| A3 |
Orphan check. Every backward pointer ("as noted above", "see §X") and every demonstrative whose nearest antecedent falls inside the deletion range — diff against what was cut, then cut pointer with referent, inline a minimal restatement, or keep both |
mech |
| A4 |
Cut license by content class: procedural cuts hard toward bare imperatives; explanatory cuts filler only — trade-offs, scope conditions, sample sizes, limitations survive. precautionary |
judg |
| A5 |
Dead words: nominalizations back to verbs, expletive constructions, circumlocutions, redundant intensifiers. Default-on, no review |
mech |
| A6 |
Route before deleting: material that is true and relevant but doesn't change the reader's next action moves — footnote, appendix, the document that owns that mode — rather than dying |
judg |
B — Which device carries the logic
| ID |
Check |
Tag |
| B1 |
Reversal test. Reversing two items changes what is true, or what a reader could correctly do → the pair is chained, so prose with a because-class connective; independent items may stay bulleted — burden of proof favors keeping the list. precautionary |
judg |
| B2 |
Label the list relation — order / choose one / all must hold. A bare list implies AND |
judg |
| B3 |
A table earns its place by being countable, not by attribute count: multiply condition cardinalities, check row count, mark gaps impossible per §X or unspecified — never blank, never "N/A". Band continuous conditions first. precautionary |
judg |
| B4 |
A figure needs genuine spatial, sequential, or relational shape; one encoding no structure actively hurts comprehension. Pair a load-bearing figure with a sentence stating what it shows. Diagrams as text source, so they diff |
judg |
| B5 |
Code answers mechanical "how do I call this." Composition across calls, design rationale, and version scope need prose — another snippet will not supply them |
judg |
C — Order and repetition
| ID |
Check |
Tag |
| C1 |
Tail-chain test. Read the last few words of each sentence in a paragraph, alone, top to bottom: advancing facts, or filler and a repeated noun? |
judg |
| C2 |
Subject-swap test. For each passive, does its grammatical subject echo the prior sentence? Yes → leave it, the passive is preserving a topic. No → rewrite active. One lookup, no judgment about agency or vagueness |
mech |
| C3 |
Sections open with the conclusion, so a reader who stops there still has the takeaway |
judg |
| C4 |
Restate only where the thread was plausibly lost — heading intervened, or antecedent distant and ambiguous. One orienting clause or a real link, never a re-explanation. Aggressive on reference/how-to, light on tutorial, never changelogs. precautionary |
judg |
D — Emphasis and the linear channel
| ID |
Check |
Tag |
| D1 |
Bolded spans must be lexically complete statements — strip the bold and nothing is lost |
mech |
| D2 |
Emphasis goes to the claim whose misreading is unrecoverable (data loss, security exposure, binding commitment), not to what feels important |
judg |
| D3 |
The heading outline must convey the document alone: no skipped levels, no "Overview" / "Details" / "More" |
mech |
| D4 |
Nothing load-bearing carried only by bold, color, position, or emoji |
mech |
| D5 |
List length is a function of channel: a list the reader can visually re-consult may run long; an unaided one stays short and gets split with intervening prose |
judg |
| D6 |
Tables need header cells with scope, and a caption |
mech |
E — Staying true
| ID |
Check |
Tag |
| E1 |
Tag facts stable vs. perishable. Perishable → generate from source, assert against source in CI, or point at the constant, not the value |
mech |
| E2 |
Executable examples where the genre allows — the literal quickstart runs in CI |
mech |
| E3 |
Never gate anything on a readability score |
mech |
Do-not-cite blocklist
None of these are sourceable. They stay inline, not in a reference file, to catch the reflex before anyone opens a link:
- A 30%-bold maximum as a ceiling on emphasis density
- "Users read 25% faster with bullets"
- 7±2 as a hard cap on list length
- "Bold enables 30% faster scanning"
- The F-pattern as a design target for layout
Falsification notes
Four precautionary checks rest on convergence, not a settled result. One line each; full notes in evidence-and-limits.md:
- A4 — cut license by content class: a two-content-type RCT, not a spectrum.
- B1 — reversal implies chaining: what causal connectives do cognitively, not a bullets-vs-prose trial.
- B3 — a table earns its place by countability: a completeness argument, not a comprehension study.
- C4 — restate only when the thread is plausibly lost: a screen-deficit study, not restatement itself.
Deference and scope
An existing project template wins on structure — sections, names, ordering; this skill governs the prose inside its slots. quirk:artifacts:adr output keeps its shape.
An explicit user request wins outright: the rules stand down wherever they'd fight it, though they still apply elsewhere — structure, orphan checks, the linear channel.
The skill applies to the whole document, not only text written after it loaded. A README scannable in its back half and dense in its front fails the property it claims.
Skill applies to itself
This file obeys its own structure rules: sections open with the conclusion, no list implies an order where none exists, no bold carries anything load-bearing alone. It deviates where the agent-facing genre differs from prose for a human skimmer — the tables above exist to be grepped, not read as narrative.
Report shape
Step 7's output:
## Revision report
- Removed: <section> — <one-line F2 reason>. Moved: <item> → <destination> (repeat per item)
- Routed (F4): <item> — <category> → <destination>
- Changed (A–E): <check ID> — <what changed> — <why>
- Escalated (non-blocking, F1 undecided): <repo signals checked> — <why no reader/decision line could be derived> — F2, F4 and A6 skipped: no section removed, no candidate material moved; the rest of A–E ran
- Proposed, not performed (out of scope): <item> → <destination file> — material left in place pending authorization
- Escalated (non-blocking, A–E only): <check ID> — <the call> — <how to reverse it>
Not a checklist transcript: each line names a concrete edit and its reason; skip a line type when nothing fired.
Real-user gap
These rules were never validated against real readers, and that can't happen inside a session. Several judg-tagged calls above are best-available procedure, not settled findings — apply them, but hold them as claims, not results.
Links out
Worked before/after examples for all 28 checks: worked-examples.md. Grounded-vs-precautionary status and full falsification notes: evidence-and-limits.md.
1---2name: writing-scannable-prose3description: Use when writing or revising a human-facing technical document — README, guide, ADR, PR description, changelog — or when asked to tighten, shorten, or declutter one ("tighten this", "too long", "hard to scan", "make this scannable"). Separates what the reader's decision depends on from detail they do not want, cutting whole sections that earn no keep; then treats every remaining cut as a dependency problem, so compression removes real bulk without orphaning the claims it leaves behind.4---56# Writing Scannable Prose78## Overview910Treat the document as a dependency graph, not a pile of words to shorten: a cut is safe only when nothing depends on what got removed. Watch for **orphaning** — a claim whose scope, evidence, or antecedent got cut out from under it, leaving a reader confidently wrong, not visibly confused. **Verbosity** — words carrying no information — is real, but secondary and easy.1112A document gets read twice: as a visual surface (headings, bold, layout) and as a linear stream where none of that survives — screen readers, TTS, plain-text extraction. Only markup-level structure survives the second read.1314**Gate:** fires on a document carrying claims, reasoning, or procedure — README, guide, ADR, PR description, changelog, or similar — not on line count. A three-line changelog entry is in scope; a one-line typo-fix PR body is not.1516## Authoring inputs1718Decide two things once, before drafting:1920| Decision | Consumed by | Controls |21|---|---|---|22| Reader, and what they decide next | F1, F2, F4 | Whether a section survives, what counts as author-facing detail |23| Passage procedural or explanatory | A4 | How hard cutting may go within that passage |2425Diátaxis mode (tutorial / how-to / reference / explanation) is **not** an authoring input — A6 alone consumes it, relocating a paragraph misfiled for its section's mode.2627## Revision protocol28291. **Derive F1's line.** Check sibling artifacts **of the same genre**, the template, and whether the content already lives canonically; state the result: `Reader: <who> — Decision: <what they do next>`. No signal resolves it → say so.302. **Run F2 against every section** against F1's line: does its absence change the named decision? Mark keep or remove. Then, for each section marked keep, ask what F4 routing would leave it holding and re-test *that* against F1's line: a section hollowed below materiality is a residue stub reached by hollowing instead of by bad removal, so mark it remove too. This is an assessment, not the edit — F4 routes in step 5, on sections that survive. **F1 is unresolved whenever either field is missing — a reader named with no decision, or a decision with no reader — and an unresolved F1 means F2 doesn't run; every section defaults to keep.** The same applies to every check that reads F1's line: F4 and A6 both route by what the named reader needs, so an unresolved F1 skips them too and their candidate material stays put. A–E otherwise run normally. Name each skip as a non-blocking escalation — not a second blocking point, since none of it halts the pass.313. **For every section marked remove — by either route — run F3**: name the section that owns each load-bearing item — and it must be a section surviving this batch, or the item dies with its destination; no owning section → find one, or reclassify it as not load-bearing only when it fails F1's line like any other cut, naming the reclassification to the user beside the removal proposal.324. **Emit the removal proposal** (shape below) in one batch, then **stop for an answer before step 5** — the protocol's one blocking point. Every row needs an explicit disposition; any row the answer does not address is rejected, which is the fail-safe direction. This batch covers both routes to removal, so it is the only proposal the pass emits and the only point it waits. A rejection is final for this pass: the section stays and is not re-proposed, which is what makes the protocol terminate. Nothing marked → straight to step 5.335. **If removals were proposed, apply the answer**: perform the approved F2 removals and their F3 moves, and retain the sections the user rejected. Then, **if F1 resolved**, **run F4** on every surviving section: route author-facing detail to its destination (table below); a deliberately-not-made decision stays. If F1 did not resolve, F4 does not run and its candidate material stays put. F4 and A6 may not empty a section whose removal was rejected — a section gutted to nothing is an F2 removal the user declined, taken anyway. Step 2 already assessed what routing would leave, so no section is re-tested here and the pass does not wait again.346. **Run groups A–E** on the surviving content — all of them if F1 resolved; all but A6 if it did not, since A6 routes by the reader's next action. Reuse a passage's procedural/explanatory classification from the authoring phase when one was made there; otherwise — the common case, since authoring rarely runs against an existing draft — classify it in place, immediately before A4 fires, passage grain only, never at document or section grain. A call here that can't be self-verified is non-blocking: apply it provisionally and name it in the report.357. **Emit the revision report** (shape below): everything steps 1–6 changed and why, including the escalations from steps 2 and 6.3637**Removal proposal shape** (step 4's output):3839```40## Proposed section removals41| Section | Why it fails F2 (against the Reader/Decision line) | Load-bearing items | Destination |42|---|---|---|---|43```4445One row per removed section, empty when nothing was marked — including the F1-undecided case, named in the report. An item F3 reclassifies out of load-bearing status is named beside the proposal too, with the F1 line it failed — approving a removal means approving what goes with it.4647## Escalation modes4849Escalation has two modes, and only one of them blocks:5051| Mode | Applies to | Behavior |52|---|---|---|53| Blocking | F2 section removals only | Proposed in one batch before A–E begins; the pass waits |54| Non-blocking | An unresolved F1 line | Not applied: F2, F4 and A6 are all skipped — no section removed, no F4/A6 candidate material moved. The rest of A–E still runs. Each skip is named in the report |55| Non-blocking | Any unverifiable A–E judgment call | Applied provisionally, named in the report; the pass doesn't wait |5657The asymmetry is reversibility: an A–E edit is visible in a diff, so a wrong call costs a word. A section removal leaves no trace once gone.5859## The 28 checks6061Six groups, one table each. `mech` is judgment-free, `judg` reduces to yes/no, `prin` makes no enforcement claim; `precautionary` rests on convergence, not a settled result ([evidence-and-limits.md](evidence-and-limits.md)).6263Group F runs first: compressing a doomed section risks a **residue stub** — bulk gone, shell remaining, passing every check while the reader still gets a section they didn't want. A stub has two routes: removing a section badly, which F3 guards, and hollowing one that survived, which F4's re-test guards.6465### F — Does this section belong?6667| ID | Check | Tag |68|---|---|---|69| F1 | Name the reader and decision in one visible line, from repo signals — everything downstream tests against it | `judg` |70| F2 | Section materiality: absence unchanged decision → the section goes — not compressed, not tightened, not reduced to a summary. Removals proposed, not performed | `judg` |71| F3 | Re-home before removing: every load-bearing item lands in the section that owns it, and the proposal names where. Deleting wholesale takes the survivors with it; keeping a compressed shell fails F2 | `judg` |72| F4 | Detail level: within a surviving section, separate what the named reader needs from what the author wanted on record — each recurring author-facing category has a destination that is not this document. What routing would leave is re-tested against F2 before the removal proposal, so a section hollowed below materiality joins that one batch rather than surviving as a shell | `judg` |7374F4 routes surviving detail to:7576| Category | Destination |77|---|---|78| Review-round history, per-note feedback tables | The review threads themselves |79| Approaches tried and rejected | Commit body, or an ADR if it outlives the change |80| How it was verified — transcripts, run counts, methodology | Test docstrings, or the test itself |81| Tooling and process narrative | Nowhere; work-in-progress residue |82| A decision deliberately not made, with its rationale | Stays — changes what a reader concludes |8384These categories make F4 actionable, not exhaustive: content outside the table is still tested against F1's line rather than passing by default. **Scope guard, applying to every check in this skill, not just F4.** An edit to any file other than the document under revision — F4's ADR or commit body, A6's owning document, E1's CI assertion, E2's runnable quickstart — is proposed, not performed, unless that file is already in the task's scope. Read such files freely; the material stays put until the move is authorized. Non-blocking: it routes to the report like any other non-blocking item, not a second checkpoint.8586### A — What a cut may touch8788| ID | Check | Tag |89|---|---|---|90| A1 | Name the operation **before editing**: removing a word, a claim, or a qualifier that scopes a surviving claim | `judg` |91| A2 | Qualifiers (hedges), scope conditions, sample sizes, exceptions, platform caveats stay in the same visual unit as their claim — never demoted to a footnote or a later bullet. Tighten wording, not the fact | `judg` |92| A3 | **Orphan check.** Every backward pointer ("as noted above", "see §X") and every demonstrative whose nearest antecedent falls inside the deletion range — diff against what was cut, then cut pointer with referent, inline a minimal restatement, or keep both | `mech` |93| A4 | Cut license by content class: procedural cuts hard toward bare imperatives; explanatory cuts filler only — trade-offs, scope conditions, sample sizes, limitations survive. `precautionary` | `judg` |94| A5 | Dead words: nominalizations back to verbs, expletive constructions, circumlocutions, redundant intensifiers. Default-on, no review | `mech` |95| A6 | Route before deleting: material that is true and relevant but doesn't change the reader's next action moves — footnote, appendix, the document that owns that mode — rather than dying | `judg` |9697### B — Which device carries the logic9899| ID | Check | Tag |100|---|---|---|101| B1 | **Reversal test.** Reversing two items changes what is true, or what a reader could correctly do → the pair is chained, so prose with a *because*-class connective; independent items may stay bulleted — burden of proof favors keeping the list. `precautionary` | `judg` |102| B2 | Label the list relation — order / choose one / all must hold. A bare list implies AND | `judg` |103| B3 | A table earns its place by being countable, not by attribute count: multiply condition cardinalities, check row count, mark gaps `impossible per §X` or `unspecified` — never blank, never "N/A". Band continuous conditions first. `precautionary` | `judg` |104| B4 | A figure needs genuine spatial, sequential, or relational shape; one encoding no structure actively hurts comprehension. Pair a load-bearing figure with a sentence stating what it shows. Diagrams as text source, so they diff | `judg` |105| B5 | Code answers mechanical "how do I call this." Composition across calls, design rationale, and version scope need prose — another snippet will not supply them | `judg` |106107### C — Order and repetition108109| ID | Check | Tag |110|---|---|---|111| C1 | **Tail-chain test.** Read the last few words of each sentence in a paragraph, alone, top to bottom: advancing facts, or filler and a repeated noun? | `judg` |112| C2 | **Subject-swap test.** For each passive, does its grammatical subject echo the prior sentence? Yes → leave it, the passive is preserving a topic. No → rewrite active. One lookup, no judgment about agency or vagueness | `mech` |113| C3 | Sections open with the conclusion, so a reader who stops there still has the takeaway | `judg` |114| C4 | Restate only where the thread was plausibly lost — heading intervened, or antecedent distant and ambiguous. One orienting clause or a real link, never a re-explanation. Aggressive on reference/how-to, light on tutorial, never changelogs. `precautionary` | `judg` |115116### D — Emphasis and the linear channel117118| ID | Check | Tag |119|---|---|---|120| D1 | Bolded spans must be lexically complete statements — strip the bold and nothing is lost | `mech` |121| D2 | Emphasis goes to the claim whose misreading is unrecoverable (data loss, security exposure, binding commitment), not to what feels important | `judg` |122| D3 | The heading outline must convey the document alone: no skipped levels, no "Overview" / "Details" / "More" | `mech` |123| D4 | Nothing load-bearing carried only by bold, color, position, or emoji | `mech` |124| D5 | List length is a function of channel: a list the reader can visually re-consult may run long; an unaided one stays short and gets split with intervening prose | `judg` |125| D6 | Tables need header cells with `scope`, and a caption | `mech` |126127### E — Staying true128129| ID | Check | Tag |130|---|---|---|131| E1 | Tag facts stable vs. perishable. Perishable → generate from source, assert against source in CI, or point at the constant, not the value | `mech` |132| E2 | Executable examples where the genre allows — the literal quickstart runs in CI | `mech` |133| E3 | Never gate anything on a readability score | `mech` |134135## Do-not-cite blocklist136137None of these are sourceable. They stay inline, not in a reference file, to catch the reflex before anyone opens a link:138139- A 30%-bold maximum as a ceiling on emphasis density140- "Users read 25% faster with bullets"141- 7±2 as a hard cap on list length142- "Bold enables 30% faster scanning"143- The F-pattern as a design target for layout144145## Falsification notes146147Four `precautionary` checks rest on convergence, not a settled result. One line each; full notes in [evidence-and-limits.md](evidence-and-limits.md):148149- **A4** — cut license by content class: a two-content-type RCT, not a spectrum.150- **B1** — reversal implies chaining: what causal connectives do cognitively, not a bullets-vs-prose trial.151- **B3** — a table earns its place by countability: a completeness argument, not a comprehension study.152- **C4** — restate only when the thread is plausibly lost: a screen-deficit study, not restatement itself.153154## Deference and scope155156An existing project template wins on structure — sections, names, ordering; this skill governs the prose inside its slots. `quirk:artifacts:adr` output keeps its shape.157158An explicit user request wins outright: the rules stand down wherever they'd fight it, though they still apply elsewhere — structure, orphan checks, the linear channel.159160The skill applies to the whole document, not only text written after it loaded. A README scannable in its back half and dense in its front fails the property it claims.161162## Skill applies to itself163164This file obeys its own structure rules: sections open with the conclusion, no list implies an order where none exists, no bold carries anything load-bearing alone. It deviates where the agent-facing genre differs from prose for a human skimmer — the tables above exist to be grepped, not read as narrative.165166## Report shape167168Step 7's output:169170```171## Revision report172- Removed: <section> — <one-line F2 reason>. Moved: <item> → <destination> (repeat per item)173- Routed (F4): <item> — <category> → <destination>174- Changed (A–E): <check ID> — <what changed> — <why>175- Escalated (non-blocking, F1 undecided): <repo signals checked> — <why no reader/decision line could be derived> — F2, F4 and A6 skipped: no section removed, no candidate material moved; the rest of A–E ran176- Proposed, not performed (out of scope): <item> → <destination file> — material left in place pending authorization177- Escalated (non-blocking, A–E only): <check ID> — <the call> — <how to reverse it>178```179180Not a checklist transcript: each line names a concrete edit and its reason; skip a line type when nothing fired.181182## Real-user gap183184These rules were never validated against real readers, and that can't happen inside a session. Several `judg`-tagged calls above are best-available procedure, not settled findings — apply them, but hold them as claims, not results.185186## Links out187188Worked before/after examples for all 28 checks: [worked-examples.md](worked-examples.md). Grounded-vs-precautionary status and full falsification notes: [evidence-and-limits.md](evidence-and-limits.md).
zpyoung