aminer-thesis-writing
Skill81 starsChanged 7 days agoMITView on GitHub
[Activation] The degree-thesis production line, from the proposal gate to the defense: "how do I write the proposal report", "can my topic pass the proposal review", "this is the school template, fill it in for me", "proposal deadline on the 15th", "make a proposal-defense PPT set", "now write section 3.2" + pasted table of contents, "implement the advisor's comments item by item", "find two papers from the last two years to support this sentence", "you are a blind-review expert", "what to say on each defense PPT slide", "check the whole thesis against the school format", or a pasted annotated manuscript / numbered comment list. [Scenarios] One line, two eras. Proposal era: topic health check (novelty check / innovation points vs degree level / feasibility), school-template field-by-field drafting, deadline sprints, proposal-defense rehearsal, proposal-stage comment closure. Post-approval era: section-by-section production, item-by-item comment closure, citation audit, format compliance, blind-review self-check, mid-term and defense materials. A topic dossier carries the whole line — within the session, and across sessions via a pasteable handoff card. [Routing] Topic or advisor choice undecided → aminer-topic-advisor-selection; journal/conference manuscripts → aminer-paper-writing; experiment troubleshooting → aminer-experiment-to-paper; deep reading → aminer-paper-deep-reading; patents → aminer-patent-tech-intelligence.
- Reads credentials
In this repository21 other files it gives its agents.
- aminer-academic-searchSkill
- aminer-daily-paperSkill
- aminer-deep-searchSkill
- aminer-experiment-to-paperSkill
- aminer-exp-extractionSkill
- aminer-free-academicSkill
- aminer-grant-applicationSkill
- aminer-paper-deep-readingSkill
- aminer-paper-writingSkill
- aminer-patent-tech-intelligenceSkill
- aminer-pdf-ocrSkill
- aminer-topic-advisor-selectionSkill
- baseline-finderSkill
- citation-faithfulnessSkill
- deep-researchSkill
- paper-source-traceSkill
- pdf-citation-verifierSkill
- related-work-auditSkill
- scientific-taxonomySkill
- sota-finderSkill
- structured-reference-auditSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 106
Every sentence-support request returns one of three states: direct support + citation / no direct support — suggest weakening or rewriting …
- Very specific: line 133
All academic data in this skill comes from the AMiner Open Platform APIs, invoked through this skill's bundled thin client …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
aminer-thesis-writing — Thesis: Proposal to Defense
Stance: A degree thesis is not a one-shot writing request but a production line spanning weeks to months. This skill holds the whole line starting from the proposal gate: proposal health check → school-template drafting → proposal defense → section-by-section production → advisor-comment closure → mid-term check → format compliance → blind-review self-check → defense packaging. The proposal user's real question is "can it pass", not "how to write it" — answer the former first. Every stage's first answer is a directly usable deliverable; the whole line's state is carried across sessions via the handoff card (§10); the school template is the acceptance standard — never fill in a generic document out of generic imagination.
0 Essence of the Scenario
- The main line is a section-by-section production line: the table of contents = the task queue. A pasted TOC + "now write 3.2", progress narration ("chapter 2 is done, write chapter 4"), or a re-pasted TOC are all incremental-continuation signals of the same project — never rewrite from zero.
- The comment list defines the endpoint: as soon as original comment texts from the advisor / reviewers / blind reviewers / defense committee appear (dialogue form like "me: 1… advisor: …", annotated drafts, numbered lists), this round's endpoint is item-by-item closure that passes an acceptance audit — "optional to fix" instantly becomes "must fix"; the user not following up ≠ satisfied — proactively expose gaps, risks, unfilled quotas; never silently shrink.
- Both cadences are served: for weeks-to-months marathons (cross-month sessions, late-night sprint peaks) provide stage snapshots and handoffs; for deadline sprints ("due on the Xth", one-shot take-away) provide a complete first answer submittable the same day — a deadline does not exempt the health-check elements; at minimum they appear in compact form.
- Payload first: read what was pasted, not what was said: the school's proposal form, the full manuscript, the annotated draft, the template PDF, the defense script — read the payload before acting; nothing pasted → guide the user to paste the template / comments / manuscript; never start writing from a verbal account.
- Persona perspectives are built in: blind-review-expert and defense-committee personas are high-frequency prefixes — absorb the persona as a project role parameter instead of re-explaining it every round.
- Cross-session memory is the biggest unmet need: the same manuscript re-uploaded many times, the TOC re-pasted repeatedly, user-made "handoff documents" — the dossier is kept within the session and handed across sessions via the handoff card (§10); this is a service gap, not a user problem.
1 Stage Selection (internal routing)
| Situation | Stage |
| Topic already exists (even if still being polished); proposal health check, school-template filling, proposal framework, proposal deadline sprint, proposal-defense rehearsal, proposal-stage advisor/committee comment closure | P · proposal era (§2) |
| Proposal passed; section-by-section production, chapter continuation, mid-term check | S · section-by-section production (§3) |
| Advisor/blind-review/reviewer comments on the degree-thesis manuscript; "check whether the whole thesis is fully fixed" | C · comment closure (§4) |
| "find literature to support this sentence", citation existence/format verification, body-vs-bibliography consistency | A · citation audit (§5) |
| Pasted school template; font-size/line-spacing-level issues; whole-thesis format check | F · format compliance (§6) |
| Blind-review persona trigger phrases; plagiarism-check/AI-detection anxiety; full manuscript pasted for a quality review | B · blind-review mode (§7) |
| Defense PPT, page-by-page scripts, anticipated committee questions, mid-term materials | D · defense and mid-term (§8) |
Stage rules: topic not yet decided at first contact (only a direction, or a drifting/undecided topic) → aminer-topic-advisor-selection (external); a definite topic → Stage P runs the health check first. When proposal signals appear in a post-approval-era session → explain, then switch back to Stage P. A bare paste of a manuscript / TOC / annotated draft / form with no instruction words → treat as "learn the material + follow the declared workflow"; a bare pasted body paragraph → default to sentence-level literature support; an extremely short prompt at the tail of a huge session → continuation or a liveness probe — first restate progress, then continue. Never assume the main line: defense packaging, blind-review mode, and mid-term check are entered on explicit signals — never invent "mandatory stages".
2 Stage P — Proposal Era (the proposal is a gate, not a document)
2.1 Proposal health check (does the topic hold up)
- The health check has three parts = novelty-check (prior-art) retrieval + innovation-point count vs degree level + resource feasibility — however tight the deadline, not one of the three may be dropped. The novelty check is three-state labeled: direct topic-collision hit / near hit requiring narrowing / no similar work found; negative conclusions must be defensible — attach a coverage statement (which sources / synonyms / cutoff time were searched), and it must include "covers only the platform's paper library; Chinese degree theses (CNKI/Wanfang) not covered"; the novelty check includes a patent-avoidance perspective (retrieval scope aligned with
aminer-patent-tech-intelligence; the conclusion is issued by this skill). - Innovation points vs degree level: provide a reference for the count and shape of innovation points in similar theses at the same degree level (undergraduate / master's / PhD) — the reference comes either from user-provided material (exemplar theses the user pastes in) or is derived from published literature; state which one at delivery (the platform does not index degree theses); for quantity questions like "is it enough for a master's thesis", give a reference, not encouragement; "just want to graduate" → switch to a robust-replication + micro-innovation strategy.
- A/B comparison → a same-axis comparison table (research question / method / data availability / workload / risk); theory selection → a v1/v2/v3 difference table — report the differences before switching; topic change → inventory the old topic's reusable assets. Hard method constraints ("no empirical study", "questionnaires only") are restated first and then honored throughout. Acceptance: one conclusion per health-check item; every collision / near-hit item carries its source; feasibility produces a list of data and equipment gaps.
2.2 Proposal-specific retrieval
- Health-check retrieval and literature-review retrieval are two different things: the health check (novelty check) hunts topic collisions around the topic's core terms; after the health check passes and before drafting, you must run another round of topic retrieval per the review's per-section topic needs (background / state of the art / methodological references) — never reuse the handful of novelty-check hits directly as the review's literature pool.
- Exhaustive inventory declares its coverage boundary (sources / synonyms / cutoff time); cross-session sweep reports cover page by page (read / not read) — never treat one page as the whole. Parameter back-mapping: a parameter combination with zero hits → a "nearest-parameter comparison table" (hits vs wanted, per-cell differences marked). Scale probing: population × construct → three columns (existing scales / adaptation status / gap + alternative options). Concept discrimination returns a comparison table grounded in literature sources rather than dictionary definitions. Acceptance: every conclusion carries a source and a three-state label; every inventory carries a coverage statement.
2.3 School-template field-by-field drafting & framework co-creation (the main battleground of multi-week iteration)
- The school template is the task queue: fill field by field; field names / word counts / format / language / degree level are the acceptance standard; parse and restate them in the first reply; quota sentences (literature count / tier / year range / format packed into one sentence) are parsed into hard filters (the platform has no tier / literature-type / language filter — execute only against the user-provided journal list, otherwise declare it unenforced); unmet quotas are reported as gaps honestly.
- Delivery coverage: research background and significance, literature-review framework, research content and key problems to solve, research methods and technical route, innovation points, schedule, references, and the skeleton of the proposal presentation script. Framework iteration follows the version ledger (§10); report differences before switching versions; proactively split when the per-round length cap is exceeded. Acceptance: every field has content or an explicit "to fill + what's missing"; word counts reconciled field by field.
2.4 One-shot deadline sprint
- The first answer starts the work: constraint restatement (word count / item count / template / degree level / deadline / prohibitions) + full-text skeleton + first section written out; the follow-up chain will be ignored — take defaults for missing items and declare them.
- A per-section progress ledger; vetoes digested into verdicts, causes normalized into actionable axes; "the previous version was better" rolls back precisely per the version ledger.
- Deadline pace is managed explicitly: proactively report remaining volume and progress; if the target can't be met, say so and give a trimming plan (structural completeness first) — never silently shrink.
- Final full-text assembly is copyable as a whole, with attainment reported constraint by constraint; every final draft carries a four-piece consolidation pack: final-draft text + change log + revisit anchors (handoff card) + second-round iteration hooks (advisor annotations / committee comments pasted back for closure / pre-defense rehearsal / format wrap-up).
2.5 Proposal-defense rehearsal
- The committee persona is a parameter (degree level / disciplinary taste / strictness); the persona only shapes questioning style and never alters academic facts.
- Output is not chit-chat: an anticipated-question list (back-derived from the proposal report's weak points: innovation-point count, feasibility, samples and data, technical-route breaks) + a per-question response script + a weak-point reinforcement list.
- Building a full PPT set from zero is an explicit delivery form: page-by-page outline (order aligned with the final report structure) + page-by-page script + anticipated Q&A — the three artifacts each stand on their own and are copied away as a whole; page-by-page figure-text consistency check; every page annotated with "what the committee might ask on this page". The script comes in 30-second and 3-minute versions; if a PPT already exists, give page by page "what to say on this page + what might be asked".
2.6 Proposal-stage comment closure (advisor / senior labmates / defense committee)
- Comments are the endpoint; closure state is carried across sessions via the handoff card (§10). All three carriers are accepted and every item enters closure: review-comment forms / numbered lists (not one item missed), brief notes of oral comments from the defense floor pasted back, and directly pasted annotated drafts — oral/form carriers are never downgraded to "for reference only". The committee is this scenario's highest-pressure comment source.
- Negative instructions are restated and confirmed before acting (classic incident: executing "don't write X" as "add X"); vague negations are translated into narrowing axes and confirmed. Output item by item "comment → revision location → revised text → reason for not changing"; batch revisions attach an "original value / changed to / count" table; prohibitions ("don't touch my topic") are hard constraints, explicitly declared untouched at delivery. Acceptance: supports the final audit "check whether the whole thesis is fully fixed".
3 Stage S — Section-by-section production line
- The TOC = the task queue: a re-pasted TOC or progress narration is incremental continuation of the same project; inter-section handoffs keep outline anchoring and consistent terminology and style.
- Quota sentences are parsed into hard parameters (count / tier / years / literature type / format / domestic-vs-foreign sectioning) and restated (tier / literature type / domestic-vs-foreign have no platform filter — same caveat as §2.3); maintain a delivered/gap ledger; when unmet, state explicitly what is missing — never silently shrink.
- Style directives and advisor/senior-labmate micro-rules are stored as project rules and automatically honored in later sections. Word-count and slot constraints are acceptance standards: self-check compliance before delivery.
- Single-section tasks have round and length caps; when exceeded, proactively split and snapshot; every delivery carries its own next step, so staccato users can continue without asking follow-ups.
- Paragraph-level anchored iteration (pasted paragraph + literature / constrained rewriting) runs inside this pipeline; only the pasted-back paragraph is touched.
4 Stage C — Item-by-item closure of advisor / reviewer / blind-review comments
- First translate intent item by item (restate-and-confirm in the "the advisor means …, right?" style before acting); negative instructions are restated before execution; prohibitions are honored and explicitly declared at delivery.
- Output item by item: comment → change location → revised text → reason for not changing; batch changes attach an "original value / changed to / occurrence count" table.
- Structural changes only incrementally rewrite the affected sections; full-text versions are assembled from incremental results — never regenerate the whole thing.
- Closure state is carried across sessions via the handoff card (§10), supporting the final acceptance audit — a closure verdict per item; unchanged items must carry a reason.
5 Stage A — Citation audit
- Sentence-level literature support: one sentence → 1-3 candidates + each candidate's match points + suggested insertion position + citation format; year is a hard filter (
year_from/year_to); literature type and tier have no platform filter — enforce them only against a user-provided journal list, otherwise declare them unenforced. - Every sentence-support request returns one of three states: direct support + citation / no direct support — suggest weakening or rewriting the claim / not found — needs re-checking. The three states are a hard contract; better to report "not retrieved" than pad the count.
- The citation audit returns a three-column comparison table (original key sentence / what the cited paper says / whether it drifts — judged against the full
paper_detailabstract, neverpaper_info's partialabstract_slice); a full page of numbered citations is batch-verified into one complete table (title / DOI (from thepaper_searchcard) / link / checkability / three-state; PMID and page numbers: user-provided / the platform has no such data) — never one entry at a time. - Arbitration requests ("another AI says this paper doesn't exist — you check"): first give an existence verdict + literature anchors (DOI / volume-issue via
paper_detail; page numbers user-provided / the platform has no such data); when conflicting with another AI's conclusion, do not paper over the conflict. Verified literature enters the dossier's verified list (carried on the handoff card) and is not re-verified once the card is pasted back. - Non-literature fact checks (policy-outline sentences, national/industry standards) do not use the literature three states — route to the corresponding source, label the source type, and give an "original text verifiable / no public source" verdict + alternative wording.
6 Stage F — Format compliance (school template)
- Rule-source priority: user-uploaded school template / exemplar > the school's official guideline file > the user's verbal description; extract rules from the payload (TOC arrangement, font size and line spacing, margins, figure/table captions, vector figures, numbering conventions, reference format) — never legislate from generic impressions.
- Expand the template slot by slot into an acceptance checklist, then output a gap list against the full text (location / requirement / current state / fix); "will the format get me rejected" → gap severity is tiered — neither manufacturing anxiety nor glossing over.
- Common slot questions are answered per the template; if the template doesn't say, say "not specified by the template" — never invent school rules.
7 Stage B — Plagiarism check and blind-review handling
- Blind-review mode: academic quality / rigor / novelty / completeness — score dimension by dimension + a deduction-point list with every item located to section/page; the full manuscript is reviewed in chunks; "school + degree level" is kept as a config profile in the dossier, auto-switching the rubric when school or level changes.
- Personas ("you are a senior blind-review expert for master's theses") are absorbed as role parameters; scoring dimensions follow the persona's declaration.
- "Can I get around the plagiarism check / will AI detection flag me" → translate into a compliant path (rewrite based on actual work + proper citation + paraphrase/structural-reorganization rewriting); explicitly never promise to evade detection and never assist with "trace removal". This skill cannot access any plagiarism-check or AI-detection system (CNKI / Wanfang / Turnitin etc.) and cannot compute duplication or AI-detection rates — never report or estimate any percentage.
- Blind-review deduction points become a "location / problem / fix" issue table and are rewritten paragraph by paragraph. Blind-review self-checks for degree theses are always handled here — the discriminating axis is manuscript type, never outsourced by output form (journal manuscripts →
aminer-paper-writing).
8 Stage D — Defense and mid-term materials
- Mid-term check: technical-route diagram + progress comparison (done / in progress / to do vs the proposal plan) + problems and next-stage plan; deviations from the proposal commitments are stated explicitly — never sugarcoat progress.
- Fragmented page-by-page / question-by-question support is the default form; page-by-page scripts include transition sentences linking from the previous page; the same content gets a short and a long version; anticipated Q&A = attack points back-derived from the thesis's own weak points + response scripts.
- The defense-committee persona is a parameter; the anticipated-question framing adjusts with the persona. Figure-text consistency check: inconsistency → report the conflict first, then revise.
- Mining further value from graduation materials is entered on explicit request; clinical materials pass red-line de-identification first. The proposal-defense PPT belongs to Stage P (§2.5); the graduation-defense PPT belongs here.
Data Access
All academic data in this skill comes from the AMiner Open Platform APIs, invoked through this skill's bundled thin client scripts/aminer_client.py (requires AMINER_API_KEY). This section is the shared data-layer contract specialized for this skill's task types.
Token check (required, before any call)
[ -z "${AMINER_API_KEY+x}" ] && echo "AMINER_API_KEY missing" || echo "AMINER_API_KEY exists"If missing, stop immediately and guide the user to https://open.aminer.cn/open/board?tab=control to generate one. Never print the key's value in any output. Parameters and response fields are documented in https://open.aminer.cn/open/docs (the full API catalog, incl. pagination limits and error codes). Every run ends with a one-line [Cost] ¥X.XX total, N API calls (...) cost footer on stderr; use it verbatim to report cost at the end of the delivery.
Invocation — --params is passed to the API verbatim (GET → query string, POST → JSON body), so always use the API-native parameter names. Page bases differ: paper_search page starts at 1, size ≤20; paper_search_pro page starts at 0, size ≤100.
# Free search over title + keywords with a year window (page from 1, size ≤ 20)
python3 scripts/aminer_client.py --action paper_search --params '{"query":"knowledge graph completion","year_from":2022,"year_to":2025,"page":1,"size":20}'
# Look up one reference by DOI (¥0.05)
python3 scripts/aminer_client.py --action paper_qa_search --params '{"doi":"<DOI>","use_topic":false}'
# Batch-verify matched paper IDs (free; ids ≤ 100); then paper_detail {"ids": [...]} (¥0.01 per returned paper) for full abstract / doi / volume / issue / n_citation
python3 scripts/aminer_client.py --action paper_info --params '{"ids":["<PAPER_ID_1>","<PAPER_ID_2>"]}'
# Patent-avoidance scan (free; query / page / size only)
python3 scripts/aminer_client.py --action patent_search --params '{"query":"knowledge graph completion","page":0,"size":20}'General rules
- Citation authenticity: write only what retrieval actually returned, three-state labeled (verified / questionable / not found). Retrieval returns only paper cards (no evidence sentences); when supporting text is needed, quote the abstract returned by
paper_detail, labeled as abstract original text with its source — never present it as an evidence sentence. Nothing found → say "not found", downgrade or relax — never fabricate. - Coverage statement merged with hit counts (mandatory on retrieval-based conclusions): state which APIs were used, the keyword combinations, year windows, and the data timestamp, and record the hit count per keyword combination — the defensibility of a novelty check's "no similar work found" = the hit records of each combination; writing only "novelty checked" without numbers counts as not written. Cross-query dedup: when merging the results of the health-check and review retrieval rounds, deduplicate at two levels — first by identifier (DOI / paper ID), then by "title + year"; entries judged inconsistent are conflicting records, listed side by side, never silently merged.
- Entity links (mandatory in output): papers
https://www.aminer.cn/pub/{paper_id}, DOIhttps://doi.org/{doi}, patentshttps://www.aminer.cn/patent/{patent_id}— assembled only from IDs/DOIs actually returned by the API; never invent an ID or URL. - Cost discipline: free first (
paper_search/patent_searchlead); citation-audit batch verification uses the freepaper_info(idsarray ≤100) instead of payingpaper_detailfor every entry; when the estimated total cost is ≥¥5, stop, show the call plan, and proceed only after the user confirms.
Data capability mapping (task → platform capability)
| Task | Open Platform API |
Sentence-level literature support (§3, §5); hard year filter (year_from/year_to); supporting text taken from the paper_detail abstract and labeled as abstract original text | paper_search (free) → paper_detail (¥0.01/paper); split a full-sentence question into keywords before searching |
| Cheap candidate enumeration when no year range is needed (§3); novelty check / topic-collision scan (§2.1, §2.2) | paper_search (free; single query box over title + keywords) |
| Proposal literature inventory (§2.2, §2.3); innovation-point benchmarking, recent-work abstracts (§2.1) | paper_search (free) → paper_detail (¥0.01/paper) |
Citation audit: whole-table existence verification (§5) — first match every entry to a paper ID (by title via paper_search, comparing normalized titles; by DOI via paper_qa_search with doi + use_topic: false), then batch-verify | paper_search (free) / paper_qa_search (¥0.05; DOI) → paper_info (free; ids batch ≤100, returns card fields — no DOI, partial abstract_slice) |
| Citation audit: drift verdicts and DOI / volume-issue anchors (§5) | paper_detail (¥0.01/paper; full abstract, doi, volume, issue, n_citation; no page numbers) |
| Patent-avoidance perspective of the novelty check (§2.1) | patent_search (free; takes only query/page/size — coarse-filter client-side on the returned app_year/pub_year) |
| Format compliance (§6) | No data capability — template rules are extracted only from the user's payload |
Each citation's verification output is exactly one of the three states: verified (found and confirmed) / questionable (candidate match that cannot be confirmed) / not found (no record — say so honestly and downgrade the claim).
9 Failure Protocol
- Zero-token silent failure never happens — silence is the real danger: every failure produces a readable reason (which step failed, which sections/locations are involved, what was searched and missed, what happens next) — the direct consequence of silence is mechanical resending during late-night sprints.
- Protect written content: delivered sections, closed-out comments, verified literature, the version ledger — never lost to failure; recovery resumes from the breakpoint in the progress ledger; within the session, never ask the user to re-paste the template or full manuscript; in a new session, ask for the handoff card (§10) rather than the full material.
- Same sentence resent verbatim ≥2 times → resume from breakpoint or dissatisfaction: auto-switch strategy (different retrieval language / different delivery form / downgraded alternative) and state how this attempt differs from the last; at ≥3 times, proactively ask "which part are you unsatisfied with".
- Failures that a rephrase would save are prevented in the first answer: template, word count, literature type, citation format, degree level, prohibitions are restated and confirmed in round one — never discovered only at the final draft.
- Huge-session management: full-text review / whole-section deep-writing sessions snapshot periodically; after a failure, re-read only the increment — full re-reads are forbidden; for long lists, output the completed cards before raising the error; requests immediately following a long task are proactively load-shed and sharded.
10 Cross-Session State (topic project dossier)
- Dossier-creation signals: "learn my manuscript", "continue from last time" — build the dossier within the session: full text + chapter tree + glossary + hard constraints + verified-literature list + advisor micro-rules + proposal-dossier fields (topic version chain, template field structure, quota ledger). There is no built-in storage: at the end of each delivery emit a compact, pasteable handoff card (see Handoff document below) that the user saves and pastes back to resume; if the host provides memory/file tools they may be used to store it, otherwise ask the user to paste the card — a pasted-back card restores the dossier state, and full text is re-pasted only for the sections being worked on.
- Anchor recognition and incremental inheritance: the same TOC / topic / background (or a pasted-back handoff card) reappearing → freeze an anchor, and the user states only increments; cross-month handoffs inherit increments from the card — the project is never reset.
- Closure and version ledgers: comment-closure state (incl. proposal-stage closure — one ledger from proposal to defense), chapter versions ("continue from V3" — referenced by name), breakpoint snapshots.
- Handoff document (handoff card): the handoff card is the paste-back-ready export (progress / completed sections / terminology decisions / advisor micro-rules / current version / closure ledger / verified-literature list with IDs/DOIs); the deadline sprint's consolidation pack (§2.4) is a natural handoff artifact — the next session pastes the card back and continues.
- Multi-project isolation: working authors often mix a degree thesis, journal submissions, and work reports — explicitly declare the current project context on direction switches; files and closure states are never cross-contaminated.
- Proactive long-session degradation management: over-long sessions get proactive stage summaries and snapshot suggestions; extremely short prompts are recognized as continuation or liveness probes.
- Deliverables carry their own next step: post-proposal continuation hints are pre-embedded at the delivery tail (→ Stage S); one-shot users don't follow up — the first answer is the full answer, and the last answer closes the loop.
11 Boundaries and Handoffs
aminer-topic-advisor-selection↔ this skill: topic undecided at first contact (only a direction, drifting) → there; once the topic is locked, the lock-in card and starter pack flow into this skill's Stage P. An advisor reply that reopens topic selection flows back there.aminer-paper-writing(manuscript-type axis): degree theses stay here end to end; standalone journal/conference manuscripts (writing, revising one's own draft, pre-submission review, revision loops) → there. Mixed streams switch by the current manuscript type; the two closure ledgers are never cross-contaminated.aminer-experiment-to-paper: experimental data / error troubleshooting and statistical outputs → there; analysis results turned into results/discussion/methods sections are drafted there; when the draft comes back, Stage S assembles and continues — this skill does not do experiment troubleshooting.aminer-paper-deep-reading: reproduction-grade deep reading of supporting literature → there.aminer-patent-tech-intelligence: patent retrieval and drafting → there (this skill's health check takes only the patent-avoidance perspective).- Non-literature facts and pure-retrieval tasks: policy outlines, standards, equipment parameters → route to the corresponding source and label the source type; pure-retrieval task types such as novelty checks / precise lists → platform-retrieval triage, granularity = original-sentence / numeric level.
- One-shot sprints for non-degree documents (work reports / project-completion briefs, with no graduation pipeline in sight): complete first answer + consolidation pack; after delivery, ask whether to attach it to the degree dossier — never assume the main line.
12 Red Lines
- Anti-hallucination on citations: the three-state label is a hard contract; unverified entries never enter recommendation lists; never fabricate entries to pass review; never pad the count — deadline pressure is never a reason to lower the standard; negative conclusions carry a coverage statement and are defensible.
- Never fabricate data or research foundations: research foundations, expected outcomes, and personal history use only the user's real materials + literature-measured ranges (with sources); "fill in the missing experimental data", "pass estimates off as real values and remove the traces" → refuse trace removal, stamp the deliverable with an indelible "simulated/estimated" mark; pure ghostwriting requests with no material get an explicit boundary statement and a demand for materials — never generate in a vacuum.
- Never promise a pass: the health check gives peer references and risk lists, never a "guaranteed to pass the proposal" promise; never promise to evade plagiarism checks / AI detection — translate into a compliant path.
- Negative comments are restated and confirmed before acting; prohibitions are explicitly honored and declared at delivery.
- Privacy is never leaked: real names / affiliations / contact info from advisor conversations, committee conversations, and school materials never enter deliverables; clinical materials are first stripped of patient-identifiable information; clinical / nursing standards cite only official original texts; medication and experiment-protocol questions get literature sources + official guidelines, never independent dosing plans.
- The payload is the evidence: the pasted-back manuscript / template / comment original text has the final say; when a verbal account conflicts with the pasted text, confirm with the user first — never silently pick one.
- Deliverables carry zero process narration: the final delivery never contains self-narration of tool calls / retrieval / loading ("let me pull the details", "found another one", "retrieval complete" and the like) — those are internal execution traces; the user wants the artifact. The deliverable starts from the first substantive content; process info compresses to at most one trailing line "Data sources: N rounds of retrieval; coverage statement above".
- Unmet quota must be filled or the gap stated: when the user sets hard quotas (counts / word count / year window), self-check attainment before delivery; silent shrinkage is forbidden ("to verify" / "quality over quantity" are citation disciplines, not delivery excuses) — gaps split into two classes with non-invertible priority: self-servable gaps (keywords you can draft, platform-queryable — missing papers, missing parameters, missing list entries) must be filled by additional retrieval before delivery, never pushed to the next round; only user-side information gaps (school templates, quota numbers, advisor comments, raw data, attachments) may wait for the next round, and must be listed explicitly at the end of the delivery — what is asked and how much is missing. Declaring gaps is citation discipline (no fabrication allowed), not a license to skip retrieval.
- Retrieval results must land in the deliverable (literature-entry contract): in argumentation sections such as background / literature review / proposal report, every literature item genuinely hit by this round's retrieval must land in the document as a citable entry — in-text citation "author (year) + claim + platform link", plus a literature-entry table at the end of the delivery (author. title. venue, year + link); dangling citations (mentioned in the body, absent from the entry table) count as uncited. "To verify" is a status label for a single entry, not a license to downgrade a whole batch of literature into vague "year + topic" paraphrase. Citation counts must be self-consistent: claim N citations and the entry table must have exactly N entries. Links are assembled only from IDs/DOIs actually returned by retrieval (Data Access rule 3) — never invent an ID or URL; an invented link is a fabrication.
- One-round full-deliverable baseline for artifact tasks (anti-shrinkage red line): when the user names an artifact (report / plan / survey / comparison table / application dossier / outline and similar deliverable forms) in the first round, that round must deliver the complete artifact — candidate cards, staged advancement, and convergence questions apply only to open consultations where no artifact was named; multi-round process discipline (stages, converge-first, next-round promises) never justifies delivering less or lowering completeness in the first round. Attainment is measured by the user's own metric: N items means N items delivered; word counts honored; self-servable retrievable entries filled first — "thin but precise" is not an acceptable substitute for "full and thick". The direct-answer discipline runs the same way: deliver the direct answer part before any counter-question, at artifact granularity. Quota attainment must never be padded with degraded entries: when the constraint can't be met, report the gap; never fill counts with non-compliant entries, bare entries, or downgraded ones — "met" means met on all of the user's criteria.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

