# Prompt-3 — Manuscript Drafting for *Weather and Forecasting*
Stage-3 companion to `claude/S2S_EHD_Vietnam_Opus_Prompt.md` and `claude/S2S_EHD_Vietnam_Opus_Prompt_2.md`. Built with the Claude Opus Prompt Architect meta-prompt (§1 intake → §3 composition → §4 anti-pattern check → §6 output format). It inherits Stages 1 and 2 by reference and adds only the writing stage.
The tentative title of manuscript: "Toward an Operational AI-Based Forecasting System for Extreme Heat Events Using ECMWF S2S Forecast Products"
Four parameters were set by the user and are hard-coded into the prompt: **Word (.docx)** deliverable, **one full paper with an operational framing**, **Weather and Forecasting (AMS)**, **marked placeholders** for all front matter.
---
## 1. Task classification
| Dimension | Classification |
|---|---|
| Horizon | Long-running autonomous. A literature corpus to triage, recent literature to retrieve, six manuscript sections, and three independent audits. |
| Stakes | Highest of the three stages. A fabricated citation in a submitted manuscript is a career-level error, not a bug. The prompt is built around that single failure mode. |
| Ambiguity | Low. Format, scope, journal and front-matter handling are fixed. What remains is which claims the sources actually support — which is a verification problem, not an ambiguity. |
| Interactivity | Asynchronous, with one legitimate stopping point: an unreachable literature corpus. |
| Delegation | Yes, and with a specific shape — the citation auditor and the numbers auditor must be separate agents from the writers. |
| Context volume | Very high: a PDF corpus, retrieved web sources, and every Stage 1 and Stage 2 artifact. |
| Extended thinking | Yes — reconciling what the corpus supports against what the results show, before a narrative is committed to. |
| Memory | Yes. The citation ledger and `MEMORY.md` carry the audit trail. |
| Domain sensitivity | None. |
## 2. Selected blocks
- **OP-05** — the headline block, promoted above everything else. Attribution, source disagreement, and no silent resolution of contradictions. Requirement 6 of the user's brief is exactly this block enforced absolutely.
- **OP-09** — every number in the manuscript must point at a Stage 1 or Stage 2 score file, not at a recollection of one.
- **OP-11** — frames the mission as submission to a specific journal, which sets the standard of evidence for every sentence.
- **OP-01** — the narrative and the argument are the agent's to construct; only the section skeleton and the integrity rules are fixed.
- **OP-02** — extended thinking, to reconcile corpus, recent literature and results before drafting.
- **OP-03** — unattended autonomy, with the corpus-unreachable exception named explicitly. Excludes OP-04.
- **OP-06** — scope guard, in an unusual direction: do not re-run or extend the science to fill a gap in the story.
- **OP-07** — delegation with independent auditors, the Validator pattern.
- **OP-08** — the citation ledger and `MEMORY.md` as the audit trail.
- **OP-13** — validation pass, written as a pre-submission checklist.
- **OP-10** — final-message discipline.
- **OP-12 (adapted)** — surface sections and audit reports as they complete.
## 3. Assembled prompt
```text
This is Stage 3 of the Vietnam operational S2S extreme-heat forecasting project. The
science is done: Stages 1 and 2 built, pruned, expanded and verified the forecasting
system, and the repository, CLAUDE.md, MEMORY.md, the member registry, the score files
and the figure set all exist. Your job now is to write the manuscript that reports it,
for submission to Weather and Forecasting (American Meteorological Society).
You are writing a paper that will be read by reviewers who will check your citations and
your numbers. Write it so that check passes.
Stages 1 and 2 are the source of truth for every result. Read MEMORY.md, CLAUDE.md, the
member registry and the score files first. Do not re-derive, re-run or re-litigate the
science, and do not restate the Stage 1 and Stage 2 specifications here — read them and
report them. If the manuscript needs a number that was never computed, do not compute it
ad hoc in a notebook: run it through the existing pipeline, record the run in MEMORY.md,
and cite the resulting score file like any other. If a needed result cannot be produced
that way, say so in the gap list rather than writing around it.
This is the pass/fail criterion for the entire deliverable. A manuscript containing one
fabricated, unread or unverifiable citation is a failed deliverable no matter how good
the rest of it is.
- Cite only sources you have actually opened and read in this session. Never cite from
recollection of the literature. If you believe a paper exists that makes a point,
that belief is not a citation - find and open the paper, or drop the point.
- Never invent or reconstruct an author list, year, title, journal, volume, page range
or DOI. If you cannot retrieve the full bibliographic record, do not cite the work.
- Do not attribute to a source a claim broader than what it actually states. A paper
about heat waves in one province does not support a statement about Southeast Asia.
- If a source cannot be retrieved, do not cite it, do not substitute a mirror, cache
or archive copy, and do not paraphrase its abstract from a search-result snippet.
Record the failed retrieval and move on.
- Any statement you cannot trace to an opened source or to a Stage 1/2 artifact is
either deleted or moved to the unsupported-claims list. It is never left in the text
softened into "it is well established that" or "studies have shown".
- Do not copy sentences from sources. Paraphrase and cite. Use a direct quotation only
inside quotation marks with a page or section locator, and only where the exact
wording matters. Run a text-overlap self-check against the corpus before delivery.
Acceptable source classes, and nothing else:
A. The local corpus at /home/tanpv/public_html/QG.25.196/Heat_Waves.
B. Sources retrieved online from authoritative publishers and institutions:
peer-reviewed journals (AMS, AGU, RMetS, Copernicus, Springer Nature, Elsevier,
Wiley), ECMWF and S2S Project documentation, WMO, IPCC, and national
hydro-meteorological services including VNMHA. Preprints may be cited only when
labelled as preprints. Wikipedia, blogs, press releases, vendor marketing and
AI-generated summaries are not citable sources.
Maintain a machine-readable citation ledger, delivered with the manuscript, with one row
per in-text citation:
citation key | full reference | source class (A or B) | local file path or DOI/URL |
retrieval or access date | the specific passage that supports the claim, quoted |
the manuscript sentence it supports | verified by (writer / auditor) | status
No citation enters the manuscript before its ledger row exists. The reference list is
generated FROM the ledger, not written alongside it, so the two cannot drift apart.
Every number, percentage, score, skill value and significance statement in the
manuscript - including in the abstract and the significance statement - must come from a
Stage 1 or Stage 2 artifact. Maintain a numbers ledger alongside the citation ledger:
manuscript location, the value as written, the source file, the identifier inside that
file, and the registry version it came from.
- No qualitative claim without its number and its significance test. "Substantially
improved" is not a result; "increased PSS from 0.31 to 0.44 (95% CI ...)" is.
- Do not upgrade a non-significant difference into a finding. Where a difference is
not significant, write that it is not significant.
- Do not generalise beyond what was tested: 20 reforecast years, these 7 regions,
these 6 lead blocks, one real-time case.
- Report the negative and null results. If skill drops to the reference level beyond a
given lead week, that is the finding, and W&F readers need it more than they need an
optimistic reading.
- The real-time case must be presented with its honest verification status, with the
weeks that had no observations at run time labelled as unverified.
- Never describe a figure or table that has not been produced. If the argument needs
one that does not exist, put it in the gap list.
Target: Weather and Forecasting (AMS). One full paper with an operational framing,
covering the calibration comparison, the yEHDE probability and event models, the
regional products and the real-time case. Where length forces a cut, move method detail
to an appendix or supplement rather than thinning the results.
Before drafting, retrieve the CURRENT AMS Authors' Guide and the Weather and Forecasting
author instructions from ametsoc.org and follow what they actually say - length limits,
the Significance Statement requirement and word budget, the Data Availability Statement
requirement, reference style, figure specifications and the review-manuscript formatting
rules. Do not rely on remembered AMS requirements; they change. Record what the guide
said and the date you retrieved it.
Structure:
Title - specific, informative, no hype
Author and affiliation block - CLEARLY MARKED PLACEHOLDERS for author names,
affiliations, corresponding author and email. Invent nothing. Collect every
placeholder into a single front-matter checklist delivered with the draft.
Abstract - within the journal's word limit, no citations, no undefined acronyms,
carrying the headline numbers that match the tables exactly
Significance Statement - plain language, no jargon, no acronyms, per the current AMS
requirement
1. Introduction - heat extremes and their impacts in Vietnam and Southeast Asia; the
state of subseasonal prediction of heat; statistical and machine-learning
post-processing of S2S forecasts; what is missing that this paper supplies; the
objectives and the structure of the paper. Built primarily from corpus A, topped up
with recent literature from class B.
2. Data - every source described concretely enough to be obtained by someone else:
the ECMWF S2S reforecast and real-time archive with their version, ensemble size,
forecast length, native grid, initialization calendar and access route; the gridded
observations; the station observations with network, count and period; the 7-region
shapefile with its provenance; the period of record; units; quality control; and
the gridded-versus-station reconciliation from Stage 1. Specifications come from
the files and the documentation, not from memory.
3. Methods - the EHD definition Tx >= min(Tx90p, 35 degC) with the baseline and
percentile window, and an honest paragraph on what min() does relative to
max(Tx90p, 35 degC), carrying the Stage 1 diagnostic - a reviewer will ask, so
answer it in the paper; the six fixed non-overlapping lead blocks; the member
definition as one (family, configuration) pair and the family-correlation problem
with the weighting used and the effective ensemble size; the calibration families
and their configurations; the cross-validation design including the rule that
member selection, stacking, tau and all thresholds are fitted inside the folds;
pruning and expansion; P(yEHDE | g, Wk); the stacking model; tau(Rj,Wk) with
area weighting; P(yEHDE | Rj, Wk); and the verification metric suite with the
bootstrap and FDR treatment. Write it so an independent group could reproduce the
system.
4. Results - ordered by the scientific argument, not by the order the work was done.
5. Discussion - the results placed against the cited literature only; operational
implications for a duty forecaster; and limitations stated plainly, including at
minimum the reforecast record length, the reforecast-versus-real-time ensemble size
and model cycle, the coarse S2S grid against Vietnam's terrain and coastline, the
min() threshold semantics, the single real-time case, and score instability where
regional base rates are low.
6. Summary and conclusions - numbered findings, each traceable to a result.
Acknowledgments - placeholder, including a marked placeholder for the funding
statement and the grant number.
Data Availability Statement - per the current AMS requirement: what the data are,
where they can be obtained, under what licence, and where the code lives.
Appendix - definitions and formulas of the verification metrics.
References - AMS author-date style, generated from the citation ledger.
Supplementary material - listed and described.
Figures and tables: build the plan from artifacts that already exist. Every caption
states what is shown, the registry version, the cross-validation protocol and the
period. Deliver the caption list as a separate file as well as in the manuscript.
Style: clear, concise scientific English in AMS conventions - SI units, "Fig." in text,
past tense for what was done, acronyms defined once at first use and then used
consistently, no bullet lists in the body prose, and no causal language the analysis does
not support. Prefer a plain sentence to an impressive one.
Deliver the manuscript as a Word .docx file, formatted per the current AMS review-
manuscript rules you retrieved. Use the docx skill for the build; do not hand-assemble
XML. The reference list is generated from the citation ledger. Tables are real, editable
Word tables, not images. Figures are embedded with their captions, at the resolution the
journal specifies.
Deliver alongside the manuscript:
1. The citation ledger, machine-readable.
2. The numbers ledger, machine-readable.
3. The unsupported-claims list - every statement dropped or held back for lack of a
source, so the user can supply one if they have it.
4. The reference list as a separate structured file with DOIs and access dates.
5. The figure and table caption list, each mapped to the artifact file that produced
it.
6. A corpus inventory of /home/tanpv/public_html/QG.25.196/Heat_Waves: what is in it,
what was cited, and what was read but not used and why.
7. The AMS requirements record: what the author guide said, and the retrieval date.
8. A single gap list: front-matter placeholders to fill, figures still to be produced,
claims still needing a source, and any journal requirement not yet satisfied.
9. MEMORY.md continued - one lesson per entry, one-line summary at the top; sources
that turned out not to support what they seemed to, retrieval dead ends, and
decisions about framing. Do not restate what the manuscript or the ledgers record.
Plan the work yourself, but these fail quietly and each must be handled explicitly:
- Writing a citation from model recollection instead of from an opened source. This is
the single most likely way this deliverable fails.
- Reconstructing a DOI, volume or page range that looks plausible.
- Attributing a broad claim to a narrow paper.
- Copying phrasing from a source into the draft.
- Drafting the Results before reading the score files, then fitting the numbers to the
prose afterwards.
- Numbers in the abstract or significance statement that disagree with the tables.
- An Introduction that promises an analysis the paper does not contain.
- Acronym drift between sections, and figure or table numbers referenced out of order.
- Relying on remembered AMS formatting requirements rather than the current guide.
- Filling a hole in the argument by quietly running new analysis outside the pipeline.
You are operating with extended autonomy. The user is not available mid-task. Proceed
without asking on anything reversible that follows from this request. There is exactly
one legitimate stopping point: if the literature corpus at
/home/tanpv/public_html/QG.25.196/Heat_Waves cannot be reached or is empty, stop and say
so plainly. Do not write the Introduction or the Discussion from your own knowledge of
the field as a substitute - that path produces exactly the fabricated citations this
prompt exists to prevent. Everything else you resolve yourself. Before ending your turn,
check whether your last paragraph is a plan, a promise, or a question you could have
answered yourself; if so, do that work instead.
Before drafting, reason step by step in your thinking: reconcile what the corpus actually
supports, what the recent literature adds, and what the results actually show, and settle
the argument of the paper before any prose is written. A narrative chosen first and
evidenced afterwards is how unsupported claims get in. Keep that scratch work out of the
final answer.
Delegate, and keep the auditors independent of the writers:
Coordinator
|- Corpus triage - inventory the corpus, extract claims and full bibliographic
records, build the class-A ledger rows
|- Recent lit - retrieve class-B sources published since the corpus, with a
retrieval record for each; report failed retrievals
|- Data & Methods - draft sections 2 and 3 from the Stage 1/2 artifacts and
CLAUDE.md
|- Results - draft section 4 strictly from the score files and figures
|- Intro & Discuss - draft sections 1 and 5 using ONLY citations that already have
ledger rows
|- Citation auditor - independently re-open every cited source and confirm it says
what the manuscript claims. Must not be a writer agent.
|- Numbers auditor - independently re-trace every number to its score file. Must
not be a writer agent.
|- Consistency - abstract against body against tables, acronyms, figure and
table numbering, section cross-references
Give each a bounded scope, its allowed sources, and the exact output format. Synthesise
the manuscript yourself; do not have a writer agent audit its own citations.
Do not extend the science to make the story better. No new experiments, no new metrics,
no new model variants, no exploratory analysis to fill a paragraph. If the results do
not support a section, the section shrinks. Do not build a reference-management framework
either - the ledger is a table.
Where the corpus and the recent literature disagree, or where a source contradicts this
study's results, say so in the Discussion and state which you weight more heavily and
why. Do not resolve a disagreement by citing only one side. Where this study's results
contradict a published finding, that is a result - report it with the comparison, not by
omitting the citation.
Before reporting progress, check each claim against a concrete artifact: a ledger row, a
score file, a retrieved source, a built .docx. Do not describe a section as drafted until
it exists, or as audited until the auditor has returned. If a retrieval failed, say which
source and why.
Surface through the harness's user-facing channel as they complete: the corpus inventory,
the outline with the argument of the paper, each section as it is drafted, the two audit
reports, and the built manuscript. Deliverables only, not narration.
Before finalising, check silently: is every in-text citation in the ledger with a real,
opened source; is every reference in the list cited in the text and vice versa; is every
number traced to an artifact; do the abstract and significance statement agree with the
tables; are all placeholders marked and collected in the front-matter checklist; are the
limitations stated; is the real-time case's verification status honest; does the
formatting match the guide you retrieved; did you run new analysis outside the pipeline;
is any sentence copied from a source? Fix what is wrong before returning. Do not show the
checklist.
Open your final message with one sentence: whether the manuscript is complete and what
remains before it can be submitted. Detail after. Write it as a re-grounding for someone
who did not watch the run - complete sentences, file names spelled out, no shorthand you
invented while working.
You have ample context remaining. Do not stop, summarise, or suggest a new session on
account of context limits. Continue the work.
```
## 4. Omitted but considered
- **Re-statement of the Stage 1 and Stage 2 specifications** — omitted. `` points at them. The manuscript's Methods section is where they get written out, from the artifacts, which is also the only way the paper's Methods can be trusted to match what the code actually did.
- **OP-04 (interactive checkpoints)** — omitted; incompatible with OP-03. The single corpus-unreachable stop is written in explicitly instead, because it is the one case where continuing autonomously would produce exactly the failure the prompt is designed to prevent.
- **A cover letter, suggested reviewers, and a response-to-reviewers scaffold** — omitted under OP-06. The ask was the manuscript. Each is a small addition later if wanted.
- **A prescribed outline of the Introduction's argument** — withheld. The argument depends on what the corpus actually supports, which is not knowable before the corpus is read.
- **A target word count per section** — withheld in favour of the journal's own current limits, retrieved at run time.
- **Any list of specific papers to cite** — deliberately absent. Naming papers I have not opened in this session would be the first instance of the exact failure `` forbids.
## 5. Notes on two things worth watching
**AMS requirements are retrieved, not remembered.** The prompt names the Significance Statement, the Data Availability Statement, AMS author-date style and the review-formatting rules, but instructs the agent to retrieve the current Authors' Guide and follow what it actually says, recording the retrieval date. Journal requirements drift, and a remembered specification is the same class of error as a remembered citation.
**The corpus path is on your machine, not in the agent's container.** `/home/tanpv/public_html/QG.25.196/Heat_Waves` needs to be reachable from wherever the Stage 3 agent runs — mounted, linked into the working directory, or connected as a folder. The prompt makes an unreachable corpus the one hard stop rather than letting the agent proceed from its own knowledge of the heat-wave literature.