| name | draft-ai-phd-reports |
| description | Draft polished AI PhD reports in Markdown from user requirements, outlines, rough notes, existing `.md` text, paper PDFs, or `.bib` files. Use when Codex needs to write or rewrite doctoral proposal reports, daily or weekly or monthly reports, midterm reports, progress summaries, research statements, or other AI PhD-facing academic documents; remove obvious AI-generated tone; preserve a coherent professional narrative; verify factual claims and paper details with live search when needed; and integrate accurate citations or cross-references from user-provided bibliography data or source papers. |
Draft AI PhD Reports
Overview
Write AI PhD reports that read like serious academic prose rather than generic assistant output. Produce a final Markdown document grounded in the user's materials, explicit requirements, verified external facts when needed, and accurate citation handling.
When the user provides an existing Markdown draft, first audit whether its structure, headings, terminology, and Markdown mechanics match the expected report genre and will survive downstream md2all conversion cleanly.
This skill also maintains a reusable session memory whenever the work spans iterative revisions, preference discovery, or repeated user corrections. The memory should capture what the user kept changing, what kinds of outputs the user rejected, and what practices consistently improved the drafts.
Read references/workflow.md first. Load references/report-types.md when choosing a document structure. Load references/style-guide.md before drafting. Load references/citation-and-evidence.md whenever the user provides papers, claims, URLs, or bibliography files. Load references/session-memory-schema.md whenever the task is likely to involve multiple rounds or should remain reusable in later sessions.
Trigger And Scope
Use this skill when the user asks for any of the following in an AI PhD context:
- draft a new report from requirements
- rewrite a rough draft, notes, or fragmented Markdown into polished prose
- produce a doctoral proposal or opening report
- write a daily, weekly, monthly, semester, or annual progress report
- write a PhD midterm report or related assessment document
- integrate references from
.bib, manually supplied citations, or source papers
- take an existing report plus a not-yet-cross-referenced
.bib and weave the bibliography into the body text
- verify research facts, paper claims, dates, benchmarks, or terminology before writing
Do not use this skill for casual blog posts, speculative claims that cannot be sourced, or outputs whose final format is not Markdown unless the user explicitly asks for a later conversion step handled elsewhere.
When the user is revising the same report over multiple turns, or repeatedly clarifying writing preferences, treat session-memory maintenance as part of the core task rather than an optional note-taking step.
Non-Negotiable Rules
- Treat the user's supplied facts, deadlines, and institutional constraints as primary unless they conflict internally or with verified sources.
- Browse when claims, paper metadata, timelines, benchmarks, or field status may be stale or uncertain.
- Do not invent citations, quoted findings, evaluation numbers, project milestones, or advisor expectations.
- Keep the prose professional, specific, and connected. Avoid list-heavy filler unless the document type truly requires concise enumerations.
- When writing in Chinese, follow Chinese academic writing habits rather than English sentence order translated into Chinese; prefer natural transitions, moderate sentence length, and institutionally familiar section framing.
- For Chinese doctoral opening reports and similar formal proposal materials, prefer the institutional self-referential style using 本课题 rather than first-person forms such as 我、我的、本人 unless the user explicitly asks for first-person voice.
- Preserve the user's actual contribution boundaries. Do not exaggerate progress or claim results the user has not achieved.
- When evidence is incomplete, write conservatively and mark the gap instead of hallucinating.
- Output a Markdown file or Markdown-ready body text only.
- For each input Markdown report, infer the expected institutional genre before rewriting. If the text is an opening report, default to a mainland China PhD opening-report frame unless the user specifies another template.
- Audit headings and title hierarchy before drafting. Fix malformed heading depth, vague section names, and title wording that does not fit the report type.
- Audit English terminology inside Chinese text. On first appearance of an important technical term, give the full English form when useful and add the standard abbreviation when needed; later mentions may use the Chinese term, the abbreviation, or the established mixed form consistently.
- Prefer Markdown structures that are deterministic for
md2all and Quarto or Pandoc conversion. Avoid decorative formatting that is likely to map poorly to DOCX styles.
- When the user provides an existing Markdown report, default to a short preflight diagnosis before rewriting so the report type, structure problems, terminology issues, and
md2all mapping are explicit.
- When the user provides a report draft plus an uncited
.bib, treat bibliography integration as a first-class rewriting task rather than a final cosmetic pass.
- In bibliography-integration tasks, try to use the supplied entries as fully as the evidence allows, but do not force a citation into a claim the paper does not support.
- Maintain a session memory when the work includes iterative revision, back-and-forth preference discovery, or repeated critique from the user.
- The memory must distinguish stable preferences from one-off requests. Do not fossilize temporary wording choices into global rules unless the user's behavior across the session supports that inference.
- Record both failure patterns and successful patterns: what the user rejected, why it failed, and what later replacement worked better.
- Never store fabricated facts, speculative advisor preferences, or unverifiable project claims in the memory.
- When a prior memory file exists for the same report family, load it before drafting and use it to avoid repeating previously rejected writing moves.
Workflow
1. Establish The Writing Task
Identify:
- document type
- target audience such as advisor, committee, department, or lab
- language preference
- required sections or formatting constraints
- source materials supplied by the user
- whether factual verification or citation integration is required
- whether a prior session memory exists or a new one should be created for this report line
If the user gives a weak prompt, infer the smallest safe scope and continue.
For an existing Markdown draft, classify the genre before rewriting:
- opening report or doctoral proposal
- midterm report
- periodic progress report
- research statement or summary
Then infer the standard section architecture that a mainland Chinese PhD committee, advisor, or department would usually expect unless the user provides a stricter institutional template.
At the same time, infer the most likely downstream md2all template family:
- opening report or midterm-heavy academic review: prefer
proposal-midterm
- general research progress summary or formal academic writeup: prefer
academic-report
- only fall back to non-academic templates when the genre is clearly administrative rather than research-facing
At intake, also decide the memory mode:
load-and-continue when a prior memory file exists for the same project or report family
start-new when the task is likely to span iterative revisions but no memory exists yet
minimal when the task is clearly one-shot and no reuse value is expected
2. Build The Evidence Base
Gather facts from:
- user notes and draft text
- provided
.md, plain text, PDFs, URLs, or screenshots if available
- provided
.bib files or citation snippets
- live web search for unstable or missing facts
Verify paper titles, authors, years, venues, and claim-level statements before relying on them.
If the user provides a report draft plus a separate .bib that has not yet been cited in the body, explicitly build a citation coverage plan:
- cluster the bibliography by theme
- map each cluster to likely sections or paragraphs in the report
- identify entries that need live abstract or metadata verification before use
- decide whether any section, especially
国内外研究现状, needs substantive expansion so the bibliography can be integrated naturally
At the same time, start a working memory log for the session:
- stable tone and structure preferences
- explicit user dislikes
- corrections repeated across turns
- draft patterns that failed
- draft patterns that worked better after revision
3. Audit The Source Markdown First
Before substantive drafting, inspect the source Markdown for:
- whether the document title names the actual report genre clearly
- whether heading levels are monotonic and logically nested
- whether section names match the report type instead of using generic placeholders
- whether bilingual terminology is introduced consistently
- whether citations, figures, tables, and lists are written in a form likely to render stably through
md2all
- whether the heading and content density match the intended
md2all academic template
If the source already contains institution-specific headings, preserve that frame and improve the wording beneath it. If the source structure is obviously incompatible with the genre, repair the structure first and then rewrite the prose.
For existing Markdown inputs, provide a short preflight block before the rewrite unless the user explicitly asks for draft-only output.
Recommended fields:
- report type judgment
- expected standard structure
- title or heading problems
- terminology or first-mention English issues
- citation or bibliography issues
- recommended
md2all template
- whether YAML front matter needs to be added or cleaned
- whether the draft is suitable for direct polishing or needs structural rewrite first
- whether session memory should be created or resumed
4. Choose The Document Architecture
Select a structure from references/report-types.md. Adapt it to the user's institution and maturity of work rather than copying a rigid template.
5. Draft In Natural Academic Prose
Apply references/style-guide.md. Prefer flowing paragraphs, explicit transitions, and claim-evidence-reasoning links. Use bullet lists sparingly and only when they improve readability or satisfy a required format. If the output is in Chinese, default to formal Chinese academic prose shaped by common mainland university report conventions unless the user specifies another regional style.
Across revision turns, continuously compare new user feedback against the working memory:
- if the user repeats the same criticism, upgrade it into a stable preference or a stable bad case
- if a correction appears once only, keep it local unless later evidence shows it is a broader pattern
- when a new version resolves an earlier complaint, record the successful replacement as a best practice tied to the relevant document type or section
6. Integrate Citations Carefully
Apply references/citation-and-evidence.md. Use only citations that can be mapped to actual sources. If the user gives a .bib, align citation keys with the written text and write the research-status or related-work section around the supplied bibliography rather than around uncited generalities. If the user gives papers without a .bib, cite consistently in Markdown-friendly form and summarize any missing metadata.
For the common case of existing report + uncited .bib, do all of the following:
- rewrite the literature review or research-status section around the supplied entries rather than merely appending citations
- expand or split paragraphs when needed so each citation supports a concrete claim
- try to cover as much of the supplied bibliography as is defensible
- run a final citation-key coverage check so uncited entries are deliberate rather than accidental
7. Perform A Final Audit
Check:
- tone is not robotic or repetitive
- section ordering is coherent
- claims match the evidence
- dates, metrics, and paper details are internally consistent
- Markdown headings and citation markers are usable as-is
- the document title, section names, and heading levels fit the inferred report genre
- key English technical terms are expanded on first appearance when that helps academic clarity
- Markdown is conservative enough for
md2all to render into standard DOCX styles without ad hoc cleanup
- the report would map cleanly to the inferred
md2all template, usually proposal-midterm or academic-report
Output Standard
Default output:
- one complete Markdown document
- with a clear title and section hierarchy
- with integrated citations or reference placeholders that correspond to real sources
- with no meta commentary about being an AI assistant unless the user explicitly asks for process notes
When the task spans meaningful revision rounds, the default deliverable set should also include:
- one session-memory Markdown file using references/session-memory-schema.md
- enough detail to help the next session recover the user's stable writing preferences, repeated objections, and proven fixes
For report draft + uncited .bib inputs, the default deliverable should be:
- one fully revised Markdown document
- with bibliography-aware rewriting rather than citation-only patching
- with Pandoc-style citation keys already inserted in the body
- with Markdown conservative enough for direct
md2all conversion to DOCX
When revising an existing Markdown draft, the default interaction may be:
- a concise preflight diagnosis
- the revised Markdown document
- an updated session-memory file when iterative revision is part of the task
Keep the diagnosis short and operational rather than essay-like.
If the user asks for revision rather than a fresh draft, preserve the original factual content where valid and improve structure, prose quality, and evidentiary support.
References