report-with-html — скилл для Claude Code: создание HTML-отчётов
Скилл report-with-html для Claude Code создаёт автономные визуальные HTML-отчёты, интерактивные дашборды и верстаки для ревью. Запускается по прямому запросу на веб-отчёт или визуализацию данных.
MIT 143 установок ★ 1 540 обновлён 2026-11-22
npx skills add daymade/claude-code-skills --skill report-with-html -a claude-codeЧто делает report-with-html
Инструмент генерирует автономные HTML-страницы с отчётами, архитектурными схемами или сравнением версий. Архитектура документа строится вокруг вопросов читателя, а каждое число требует единицы измерения, временного окна и источника данных. Оценочные суждения и недоказанные утверждения отсекаются по правилам валидации.
Для отрисовки графиков скилл требует предварительного подключения data-visualization-discipline, который определяет типы диаграмм и палитру. В состав входят готовый шаблон assets/report-template.html, правила плотной компоновки и скрипт scripts/preview_report.py для локального HTTP-превью.
Готовый результат проверяется непосредственно в браузере. Инструмент не используется для интерфейсов продуктов, ответов в чате или Markdown-заметок, если пользователь прямо не запросил оформление в виде веб-отчёта.
Какие задачи решает
- создание автономных HTML-отчётов
- собрать интерактивный дашборд для читателя
- оформить сравнение версий с доказательной базой
- запустить локальный HTTP-просмотр верстки
Примеры запросов агенту
- Собери автономный HTML-отчёт по результатам бенчмарка с таблицами и источниками
- Подготовь визуальный дашборд в HTML для сравнения архитектуры двух сервисов
- Запусти локальное превью готового отчёта через скрипт preview_report.py
Как установить report-with-html
1Через Skills CLI (Vercel)
npx skills add daymade/claude-code-skills --skill report-with-html -a claude-code2Вручную: положить файлы
~/.claude/skills/report-with-html/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/report-with-html/
1Через Skills CLI (Vercel)
npx skills add daymade/claude-code-skills --skill report-with-html -a codex2Вручную: положить файлы
.agents/skills/report-with-html/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add daymade/claude-code-skills --skill report-with-html -a cursor2Вручную: положить файлы
.cursor/skills/report-with-html/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install daymade/claude-code-skills report-with-html2Вручную: положить файлы
.github/skills/report-with-html/SKILL.md1Через Skills CLI (Vercel)
npx skills add daymade/claude-code-skills --skill report-with-html -a gemini-cli2Командой платформы
gemini skills install https://github.com/daymade/claude-code-skills --path report-with-htmlВручную: положить файлы
.devin/skills/report-with-html/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Через Skills CLI (Vercel)
npx skills add daymade/claude-code-skills --skill report-with-html -a opencodeClaude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о report-with-html
Как установить report-with-html в Claude Code?
Выполните команду npx skills add daymade/claude-code-skills --skill report-with-html -a claude-code в терминале или сохраните файл в ~/.claude/skills/report-with-html/SKILL.md.
Что делает report-with-html?
Формирует автономные HTML-страницы с доказательными отчётами, схемами механизмов и таблицами, проверяя вёрстку в браузере.
Что нужно для работы report-with-html?
Для проектирования графиков требуется подключить скилл data-visualization-discipline, а для локального просмотра нужен Python для запуска scripts/preview_report.py.
Исходный текст
SKILL.md · MIT
---
name: report-with-html
description: >-
Produces standalone reader-facing HTML reports/dashboards and review workbenches — evidence-backed,
browser-verified. Use for a visual HTML report or requests for “可视化报告”“图表报告” or a
standalone HTML dashboard; load data-visualization-discipline before charts. Not for product UI,
chat or Markdown unless asked to become a report, or PowerPoint.
---
# report-with-html
Create a reader-facing HTML artifact whose information architecture follows the
reader's question—not the author's implementation checklist.
## Trigger boundary
Use this skill when the requested artifact is itself a visual HTML report or
interactive information tool. Typical outputs include:
- a system, architecture, journey, causal, decision, comparison, or gap view;
- a client-facing data report or one-pager;
- a data/archive browser or review workbench;
- an explicitly requested HTML/web presentation.
**If the user asks for a “可视化报告”“图表报告” or a standalone HTML dashboard,
load this skill even when the request also sounds like frontend design. Load
`data-visualization-discipline` before designing any chart. Product UI and React/
Vue dashboards stay with their product design-system owner; this skill owns only
the standalone HTML report artifact.**
Do not trigger it for a normal status reply, generic review, code diff, issue list,
or Markdown document. Those tasks may be reported in chat unless the user explicitly
asks to turn the result into a visual HTML artifact.
For restoring or hosting an existing HTML page without changing its content,
take [Existing HTML restoration](references/local-http-preview.md#existing-html-restoration)
and stop after its delivery checks. This transport-only route does not enter the
new-report design, masked-render or reader-review workflow below. If the request
also changes the report's conclusions, layout or interactions, use the applicable
authoring/maintenance workflow for those changes.
For a local report that needs HTTP preview, use
`references/local-http-preview.md` and its bundled `scripts/preview_report.py`.
Serve only the reviewed explicit file
snapshot; verify its HTTP readiness and rendered content/action in the permitted
browser.
For a course/archive browser, use the [hierarchical reader](references/archive-reader.md)
and its reusable builder/PDF adapter. Preserve current source navigation, historical
nodes and stable material IDs; verify lesson resources, actual PDF pages and media
play/seek through the existing source-owned asset server.
For bookmark, annotation or passage readers, follow
[reading-surface disclosure](references/reading-surface-disclosure.md).
The default delivery is the verified HTML opened in a browser. A PNG is a QA artifact
and is delivered only when the user asks for an image. If the user needs editable
PowerPoint, finish and verify the HTML first, then use `deck-creator` (Route C).
## SSOT map
**Visualization judgment lives in another skill — load it before designing any chart or
geometry.** `data-visualization-discipline:data-visualization-discipline` owns which form
fits the content, which chart type and summary statistic to use, whether segments may be
aggregated, how many categorical colors are allowed, cross-view color identity, table
discipline, the antipattern list, and the pre-delivery self-check gate. This skill does not
restate those rules; a second copy would drift. This skill owns the report page itself:
register calibration, the component shelf, the approved corpus, page order, the decision-card
contract, interaction contracts, and browser verification.
- Reader-facing principles, content boundary, decision-card contract, accessibility:
`references/design-principles.md`
- **Layout density (铺满画布 / 消灭空洞), page order as narrative rhythm, the HTML table
micro-spec, and form → component implementation:** `references/visualization-patterns.md`.
**Not conditional on having already chosen a form** — most of it applies to every page
regardless of what you draw.
- Approved corpus and register calibration:
`references/approved-examples.md`
- Reusable interactions and their keyboard contracts:
`references/interaction-components.md`
- Parameter exploration, selected-record follow-up and recoverable workbench state:
`references/interactive-analysis.md`. Load before authoring those interactions;
keep static reports on the ordinary path.
- Numbered hover/focus source previews and original-content adapters:
`references/citation-source-preview.md` (candidate/pilot; load when citations use previews).
- Photos, screenshots and explanatory diagrams: `references/visual-evidence.md`.
Load before authoring when objects, physical choices, visual defects or relationships
need to be seen; do not wait until images have already been selected.
- Long-lived/generated-page maintenance:
`references/long-lived-report-maintenance.md`
- Version-comparison reading order, paired evidence and full-text inspection:
`references/version-comparison-reading.md`. Load when the page explains what
changed between editions; keep source interpretation separate from presentation.
- Fresh-context review payload:
`references/independent-review-prompt.md`
- Warm-paper starter implementation:
`assets/report-template.html`
**Read these before authoring** — the principles, the patterns file, the closest
approved example, and `data-visualization-discipline`. The rest are activated by what the
artifact actually contains.
Do not let "read only what's activated" decide the required sources above. That filter operates over needs
you already know you have, so it is structurally unable to surface *a step you don't know
exists* — and it hands the skip a legitimate-sounding rationale, which is worse than forgetting.
These are the sources that tell you which steps exist.
## Product contract
### Reader and evidence first
Write down the audience, the decision or understanding the page must enable, and the
authoritative sources before choosing components. Retrieve current facts during this
session. Every number needs a unit, date/window, and traceable source; unknown values
stay unknown.
Screen copy may contain a factual takeaway when the same page visibly supports it
with data, mechanism, or primary evidence. Remove unsupported value judgments,
slogans, maxims, superlatives, and unproven causal claims. A grep pattern such as
`不是.*是` only surfaces *candidates*; its presence alone is never a failure. Judge each
candidate with the test in `design-principles.md` §3 — delete the sentence, and ask
whether the reader loses a fact they could have checked on this page, or only an opinion.
The first stays, the second goes.
Use complete, audience-readable language. Internal codes, private shorthand, naked
algorithm scores, and unexplained abbreviations do not belong on the page.
### Let the question choose the skeleton
- “How does it work?” → journey, architecture, or mechanism.
- “What should we decide?” → options, trade-offs, decisive unknowns, recommendation.
- “Why did it happen?” → timestamped causal chain plus falsifying evidence.
- “Which is better?” → entities × criteria on a common baseline.
- “Let me inspect this collection.” → list/detail, search/filter, state, provenance,
deep links, and review-output return path.
Implementation state is a badge on that skeleton, not the skeleton itself. A missing
dependency, unverified result, or unknown must remain visible.
For technical selection or purchase comparisons, use the concrete-plan reading order
in `references/visualization-patterns.md` before arranging sections. Distinguish choosing
what to investigate next from authorizing a purchase or deployment.
### Geometry and real evidence
Follow `data-visualization-discipline` for quantitative encoding, chart form,
color, statistics, and table discipline; do not restate those rules here. For UI
or visual-product review, real screenshots are the primary evidence: crop the
exact defect with enough surrounding context and keep before/after crops
identical.
Use an installed mature charting capability or the current project's declared
chart stack for statistical charts. If neither exists, research and select a
maintained library before implementing; do not hand-build a chart engine.
### Interaction and accessibility
For calculators or workbenches that continue from the reader's parameters or
selection, load `references/interactive-analysis.md` before implementing controls,
state recovery or the return to the conversation. Keep context preparation,
state saving and an explicit message send distinct; use a manual handoff in a
standalone browser.
**Evidence comes to the reader; the reader does not travel to the evidence.** When the
page cites a clause, a screenshot, or a source passage, embed it at the point of citation
— an in-place overlay or drawer that returns the reader to where they were. An anchor
that throws the reader elsewhere on the page and expects them to navigate back is
prohibited, not merely discouraged; `design-principles.md` §13 holds the rule and the
user's own words for it. Hash anchors remain correct for nav and for external deep links:
what is banned is being forced to leave mid-read.
Copy approved interactions from `assets/components/` instead of rebuilding them.
Behavior is part of the component contract; register-specific styling may change.
Copy each component through its complete contract:
1. Read the component file's own header contract — it states prerequisites the registry
table does not repeat (required CSS tokens, required DOM structure, which element
types it scans). `citation-drawer`, for instance, only linkifies inside `<section>`
elements and needs a generator-produced `#s9`.
2. Make those prerequisites true in your page.
3. **Click the component's primary interaction in a real browser before delivery.** A
component can be present, id-unique, tag-balanced and completely inert — every static
check passes while the feature does nothing. Only the click distinguishes those two
worlds.
Every interaction must work without a pointer:
- use native buttons, links, and dialogs when available;
- provide a visible close control and Escape handling for modal surfaces;
- keep focus inside an open modal and restore it to the invoker on close;
- expose names, states, and relationships with semantic HTML and ARIA;
- preserve visible focus, logical tab order, and usable narrow-screen geometry.
## Workflow
Choose verification from the actual changed axes and still-valid evidence, not from the
number of edits. For the first delivery of a complex report, retain both independent
reader comprehension (step 9) and load-bearing source checks (step 10). They need not
always be two agents: one independent agent may first return and freeze a genuinely
page-only blind read, using the canonical prompt verbatim, and only then receive the
sources and acceptance packet. Never expose author findings, conclusions, or proposed
fixes to that agent before its blind read is frozen.
For a later delta to an accepted report, invalidate only evidence covering changed
claims, visual encoding, layout, interactions, or decision structure. When input,
configuration, criterion, and relevant state are unchanged and no new failure signal
exists, keep unrelated evidence valid. Re-render and recheck affected axes after chart,
layout, or interaction changes. Repository-required checks remain required. Stop when
the named result is decided; do not add a reviewer merely because a small fix landed.
1. **Define success.** Record audience, question, source set, required evidence,
delivery form, and what would make the artifact fail. The gate template (step 5) also
always carries a `读者已知清单` slot — fill it when the declared audience is narrower than
"any reader" (an internal, already-informed reader rather than a stranger): the specific
terms, scripts, or systems that audience is assumed to already know, in checkable
backtick-quoted form. When the audience genuinely is a stranger, write
`不适用: <why>` — the slot still has to be filled, just with that answer. Write this before
step 9's first review, not after one returns findings — see `references/audience-triage.md`
for why the ordering is load-bearing and how step 9 uses this list.
Resolve the actual reader and intended use from the original request, later corrections,
and the current project contract. Distinguish a publication-ready draft from an actual
external publication, and distinguish a **one-off static report** from a **recurring live
decision cockpit**: aesthetic polish loops must never downgrade a recurring live tool into a
hardcoded dead snapshot, and every interactive control must map to a concrete reader decision
rather than demo chrome (such as replay players or ungrounded sliders) added to chase a
visual score. Set evidence precision for that use: for private investigation,
retain traceable original identifiers, paths, and logs rather than silently substituting
a public, redacted edition. Follow the authorized project's existing delivery directory
or document library; a scratch directory is a working location. If no destination is
established, choose a reversible local default and state it. Ask only for missing inputs
that materially affect the result and cannot reasonably be defaulted. For an investigation,
record complete readability and proof of the core causal claim as separate acceptance
results. A readable report does not close the investigation while necessary, authorized
actions can still advance that claim; unavailable historical evidence remains unknown.
2. **Calibrate.** Read the mandatory sources — `design-principles.md`,
`references/visualization-patterns.md`, the closest entry in `approved-examples.md`, and
`data-visualization-discipline` — then any component reference the artifact activates.
Use the nearest approved register; do not copy a skin blindly.
3. **Plan visual evidence and 造量纲 — before a line of HTML.** First follow
`references/visual-evidence.md` for sections about concrete objects, physical choices,
visual defects or relationships. Put representative images at the relevant options
and explain connections with diagrams; identify unavailable evidence explicitly.
Record the selected visuals in the existing page intent, not a second inventory.
Section by section, name what in the content is
**ordered / has magnitude / has direction / has a counterpart**, and write that quantity
down. Each geometry then carries it in the markup:
```html
<div data-geometry="退款家数刻度轴"
data-derived-from="退款家数(0–30 家线性,共同基线;位置=家数÷30)"> … </div>
```
**No supported quantity or relationship means no invented geometry.** Keep that
content as prose; a verified photo or screenshot can still show the object or evidence.
Wrapping a sentence in a coloured box does not make it a diagram.
**A number and the coordinate that encodes it come out of the same script.** As soon as
any geometry is positioned by a computed value — a bar length, a tick at `left:30.80%`,
an axis label, a line's endpoints — write a small no-dependency script that prints *both*
the finding and the coordinates, paste from its output, and ship the script next to the
page. Computing the value in prose and then hand-writing the SVG attribute is a two-copy
fact with no `grep` that can catch the drift: it is how the same quantity ends up rounded
two different ways in two sections, and how an axis label lands at the position of a
smaller number — a lie factor authored by the very page arguing against lie factors.
Shipping the script also makes the drawing auditable: a reader re-runs it and diffs.
Step 5's gate asks for that script's path and stats the file, so this rule has a machine
consumer rather than being one more thing only you referee.
4. **Author.** If a calling Skill supplies a user-approved local report form that still
fits this report's output contract, use that form as the starting point. Otherwise
start from `assets/report-template.html` or an approved example. The calling Skill
owns template approval and storage; this Skill owns report generation and visual QA.
Keep customer facts and conclusions tied to current evidence, and do not copy the
caller's template into this Skill's update-owned `assets/`. Before updating an accepted
report's facts or content, follow the [incremental maintenance recipe](references/long-lived-report-maintenance.md#63-已验收报告的有界增量更新).
Use actual shelf components for new or changed interactions; retain the approved
embedded implementation during a data-only update. If the page embeds mutable source documents, follow the long-lived-report
reference and keep generated and authored regions separate.
5. **Run the delivery gate — it produces the artifacts the judgment steps otherwise skip.**
```bash
uv run <skill-dir>/scripts/delivery_gate.py init page.html [--height 9600]
# → renders page.png AND page--masked.png (all text transparent — the squint test,
# headless; see the script header for the technique's provenance)
# → scans for ≥120 CSS px whitespace voids (visualization-patterns.md 版面填充纪律)
# → writes page.html.gate.md: the delivery gates + the stage-1 questions
# (`STAGE1` in delivery_gate.py owns the exact list), every slot empty
# Read the masked PNG segment by segment, fill every slot, then:
uv run <skill-dir>/scripts/delivery_gate.py check page.html # exit 1 until it passes
```
`check` fails on: a missing or stale masked render (so 遮字自测 cannot be claimed without
the evidence), a percentage-driven visual with no `data-geometry`/`data-derived-from` (so
造量纲 cannot be skipped — pure-layout boxes opt out **visibly** with `data-layout`), any
unfilled slot, and a gate file older than the page (so "filled the checklist, then kept
editing" cannot pass). It asserts these steps **happened**; it says nothing about whether
the conclusions are right — that is still steps 9 and 10.
6. **Render a fresh artifact.** Step 5 already produced `page.png` at the height you passed
it; re-render here only when you changed the page since, and **pass the same height** —
`render_report.sh` overwrites in place, so a shorter H silently replaces a good render with
a truncated one.
```bash
<skill-dir>/scripts/render_report.sh page.html page.png 1300 2400 # H is an example, not a default
```
`<skill-dir>` is this skill's own directory — its absolute path is given to you when
the skill loads. Bare relative paths like `report-with-html/scripts/…` only resolve
from one particular parent directory and fail with exit 127 anywhere else.
The renderer fails without a new valid PNG and never treats an existing output as
success. Set `CHROME_BIN` to an explicit managed Chrome/Chromium executable in CI
or a shared-browser environment; an invalid override is a hard failure.
Then split it and read every segment — a tall PNG exceeds what one Read can resolve:
```bash
uv run <skill-dir>/scripts/crop_segments.py page.png --segments 5
```
**That step is also where truncation is caught.** Chrome screenshots exactly the
`--window-size` you passed and never grows to fit, while the renderer's PNG check only
proves the bytes are a valid image — so a page taller than your H loses its tail
silently, and §6 of `design-principles.md` puts the strongest conclusion last. The
splitter warns when content runs to the final canvas row, which means the canvas cut
through the page rather than the page ending inside it. It is checked here rather than
in the renderer because the only non-mutating way to measure page height is a second
`--dump-dom` call, and that hangs on some pages — a gate that can never measure is a
gate that always skips.
**That warning has one false positive, and its obvious remedy does not terminate under
it — so read the warning's own two branches before you re-render.** If the page sizes
anything in `vh` (a scrolling panel at `max-height:78vh` is the common case), headless
`1vh` is `H/100`, so raising H grows that element too and content stays pinned to the
canvas bottom at *every* H. Discriminate once by rendering at a clearly different H and
comparing the reported content bottom: unchanged means it was real truncation and is now
fixed; still ≈ the new canvas height means it is the `vh` loop. Do not keep doubling H —
an operator who sees the same warning at 3000 and at 40000 learns to ignore it, and a
gate trained into noise no longer catches the real truncation it exists for.
Also test the narrow-screen layout with real device emulation and measure overflow.
Device emulation is not optional phrasing: headless Chrome clamps `--window-size` to a
~500 CSS px minimum width (measured 2026-09-17: asking 390 yields `innerWidth=500`;
`--force-device-scale-factor` does not shrink the CSS viewport either). For 390-class
phone widths use Playwright/CDP `Emulation.setDeviceMetricsOverride` (or Ego) — never
report a `--window-size=390` run as a phone-width test.
Measuring painted bar spans for the proportion gate: anchor on the *container* (find the
track box first, then measure dark pixels inside it), never cluster content pixels
globally — edge antialiasing creates phantom clusters and a long bar hides its own track
row (two real failures from 2026-09-17).
7. **Exercise the journey.** In a real browser, follow the reader's route—not merely
a click list. Test understanding order, the obvious next action, keyboard access,
modal focus, deep links, filters, and error/empty states. Use
`frontend-visual-qa` for micro, macro, and intent review. Every shelf component you
embedded gets its primary interaction clicked here, and you write down what you
observed — "the drawer opened and highlighted clause 7" is evidence; "the component
is present" is not. Check that actual images load, important details are readable,
captions remain next to their visuals, and the delivered local file has no missing assets.
8. **Reconcile and verify.**
For a merge or a current-state update, apply the semantic acceptance contract in
`references/long-lived-report-maintenance.md` §6.1. Keep the information-point
mapping with the gate evidence; a source-link index is not the requested synthesis.
- For a rewrite or merge, run:
```bash
uv run --python 3.12 python <skill-dir>/scripts/reconcile_content_diff.py old.html new.html
```
It opens both pages in Chrome and compares their initial, computed reader-visible
text. This does not traverse tabs/accordions or other interaction-created states;
cover those states in the Journey/browser harness.
- For generated long-lived pages, run the page generator and then its `--check`
mode. Missing/duplicate anchors or stale output are failures.
- Search for stale facts, shorthand, unsupported claims, broken file paths, and
duplicated version/status values.
9. **Independent reader review — and leave a file saying it happened.** The reviewer
must start with no inherited conversation or author reasoning: in Codex use
`fork_turns: "none"`; in Claude Code start a separate non-fork `general-purpose`
agent and put all required reader context in its prompt. Give that reviewer the HTML plus its
rendered screenshots — and nothing else (the table below says "the page and nothing else";
screenshots are the page, a text-only reviewer answers the visual questions by reading
source, which is indistinguishable in the write-up from having looked), using
`references/independent-review-prompt.md` verbatim. Reader comprehension failures
are defects — with one exception, `读者已知清单` below, that a page not declaring a
narrow audience never gets to claim. Treat proposed implementation changes as
hypotheses and reproduce them before changing the artifact.
Confirm the reviewer can actually open a browser before you send it the page — several
of the review questions are about focus, Escape and keyboard order, and a text-only
reviewer will answer them by reading the source. You cannot introspect a subagent's
capabilities before spawning it, so make it the reviewer's own first action: tell it to
attempt the browser step and **report "browser unavailable — items 7 and 8 answered from source"
in its output** if it cannot. An unmarked source-read answer is the failure; a marked one
is a known gap you can cover another way. That answer is indistinguishable in
the write-up from one produced by clicking, and it lands in exactly the blind spot this
skill warns about: every static check passes while the feature does nothing.
Write the outcome to a file next to the page (or in your review archive): **how the
reviewer was spawned (agent type, and that it was not a fork)**, the prompt you gave,
each finding with its disposition and reason, and what could not be checked. Record the
spawn details because a fork's write-up reads identically to an independent one — the
only thing that distinguishes them is a fact about how it was launched, and if that is
never written down, the independence requirement above is unverifiable by anyone
including you. The file itself is the point: a review that leaves nothing behind is
indistinguishable from a review you meant to run and didn't, and that is the failure
mode this step actually has. It is not that authors disagree with the step; it is
that under delivery pressure the step is the cheapest thing to silently drop.
**Triage every "I don't understand this" finding against `读者已知清单` before it counts —
provisionally; step 10 independently re-derives it, this isn't a self-check.** The
reviewer's zero-context persona is deliberate and does not change — but a term that is on
the list committed in step 1 is a **phantom problem** (Sauro, MeasuringU): the reviewer
correctly has no way to know it, and the declared audience isn't actually confused by it.
Quote the reviewer's finding **verbatim** next to the list entry it's disposed against —
"reduces to a known term" is an interpretive claim, not a file-exists check, and the quote
is what lets someone other than you judge whether the reduction was honest. Log the
disposition as `audience-declared-known` with that quote and reason; it does not count
toward the same-axis-recurrence test below. A flagged term that is *not* on the list either
stands as a real finding, unchanged from today, or gets added now — in which case say so
directly in the gate file next to the entry (`补记于 cycle N 审阅之后`) rather than blending
it into the original list silently. Do not backdate it: `init --force`'s preserved
"上一轮的答案" appendix already carries the list as it stood before this cycle, so a
silently-blended addition is trivially caught by diffing the current list against that
appendix — say it plainly instead of relying on someone doing that diff. Full mechanism and
the grounding for why this pass is provisional rather than final are in
`references/audience-triage.md`.
**Deciding whether to run it again is an observation about invalidated evidence, not an
edit count.** Compare the delta with the claims and reader decisions the recorded review
covered. A number, term, or label change triggers a fresh page-only review only when it
invalidates the previous reader's interpretation of an affected claim, visual encoding,
layout, interaction, or decision structure; not every number or term change does. Reuse
unrelated review evidence when its input, configuration, criterion, and relevant state
are unchanged and no new failure signal exists. A local NAS-name correction, a
normalized-copy hash label, or an actionable gap-card correction requires readback of
every affected projection and its meaning, not another unchanged full QA pass. Typo and
punctuation fixes need only the affected readback. Chart, layout, or interaction changes
require a fresh render and rechecks of the affected axes. Stop once the named finding is
resolved; do not dispatch a new reviewer just because its small fix landed.
**A required re-review has no ceiling of its own — pair it with a budget, or "any
invalidating edit needs review" mints an unbounded stream of fresh rounds.** A local fix
whose affected readback resolves the named finding without invalidating the blind read
does not start another cycle. When re-review is required, a converging review and a
runaway one are indistinguishable from inside a single round — the failure has been
observed at both small scale (three full rounds to close one low-stakes correction) and
large (21 rounds on a real decision document, stopped only because a human asked why).
This is the Loop Contract from
`claude-code-hooks`'s SKILL.md rule 7 — "If the hook demands remediation, prove the
loop terminates"
(`daymade-claude-code/claude-code-hooks/SKILL.md` (daymade/claude-code-skills),
read there for the general framework and its worked failure cases) — instantiated for
this step instead of a hook:
```text
LOOP KEY: this delivery's lineage (the HTML frozen at round 1 plus every descendant
edit) + one failure axis: can a fresh, uninformed reader use this page?
FIRE T: after audience triage (above), a review still answers review-prompt item 9
"no" — or surfaces any finding that amounts to the same thing (the reader
could not actually make the decision this page exists to produce) —...
© daymade · MIT · daymade/claude-code-skills