aminer-grant-application
Skill81 starsChanged 7 days agoMITView on GitHub
[Activation] Use when the user pastes a red-header notice / guideline entry / official format template and asks to extract application requirements; asks "how do I distill the scientific question" or "my research base is X but I want to write Y"; pastes a newly issued policy document to pick a direction; says "check the already-funded projects so I don't collide"; demands "review my proposal as the review expert panel" with both-sides argumentation; asks to "find literature supporting this sentence" or to calibrate text against literature data; says "limit to N words" or "do you still remember". [Scenarios] Full-pipeline executor for grant/funding applications — requirements extraction, direction picking, distilling the scientific question (transfer mode / constraint mode / three-tier), section-by-section drafting with pasted-section iteration, sentence-level evidence pairing and data calibration, overlap check, multi-role simulated review, and peripheral fixed-form deliverables such as budgets / awards / duty reports. [Routing] Pasted-paragraph citation pairing, word-count compression, and data calibration inside the application context stay in this skill — never dispatched out as generic revision; generic iterations of the user's own academic manuscripts outside the application context go to aminer-paper-writing; reviewer-perspective simulation for journal manuscripts goes to aminer-paper-writing; when the direction is not yet set, route upstream to aminer-topic-advisor-selection first.
- 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-paper-deep-readingSkill
- aminer-paper-writingSkill
- aminer-patent-tech-intelligenceSkill
- aminer-pdf-ocrSkill
- aminer-thesis-writingSkill
- 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
5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 84
Every delivery carries its own next steps at the tail: continuable dimensions, to-do list, pre-submission checklist (checking off the …
- Distinctive: line 149
↔ aminer-topic-advisor-selection (topic-setting axis): when the scientific question comes from direction exploration that has not yet …
- Distinctive: line 157
Citation authenticity: write only what retrieval returned.
- Distinctive: line 159
Verify attribution before citing others' results / competitor conclusions.
- Distinctive: line 162
Every delivery carries its own next steps (continuable dimensions / to-dos / pre-submission checklist) — a breakpoint is churn.
≈ “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-grant-application — Grant Proposal Writing (the proposal is a pipeline with a deadline; this skill is its triage executor)
Stance: work is taken on along the chain "extract requirements → pick a direction → distill the scientific question → draft section by section → sentence-level evidence pairing → overlap check → simulated review → peripheral fixed-form deliverables". Payload first: guide the user to paste the notice / guidelines / template / half-finished draft / full text of a funded proposal — look at what was pasted, not at secondhand accounts. The first answer is everything: single-round-then-leave users exist; the first answer delivers the complete artifact per the triage result, never stretching out rounds with counter-questions. A days-before-deadline sprint that produces the entire proposal and a weeks-long marathon polishing the same proposal are both normal cadences; the delivery contract never degrades for speed.
0. Essence of the Scenario
The recurring dominant shape is not policy Q&A but whole-passage pastes of red-header notices, guideline entries, official format templates, half-finished drafts, and other people's funded proposals, demanding constraint extraction, content alignment, sentence-by-sentence evidence pairing, and fault-finding from a reviewer's perspective. The core intellectual work (distilling the scientific question, polishing the entire proposal) concentrates in one ultra-long main session; data verification, format checking, and single papers are loose resupply; the same research topic returns across days and months — the unit of continuity is the research topic, not the session. Failure rate is clearly negatively correlated with instruction specificity: the more explicit the bibliographic details / parameters / scope, the smoother it goes; pasting an entire long document and vague data requests fail most easily. The version chain sits almost entirely in the user's hands: the same section pasted back repeatedly, self-supplied rules re-pasted over and over, conclusions from the previous session pasted back in to resume — this is the number-one pain point this skill must digest.
1. First-Move Triage Table
Seemingly identical pastes are actually multiple task types, each with its own output contract; a wrong shape makes the user resend. Triage before acting:
| Task type | Recognition signals | Output contract |
| Notice / guidelines / template → requirements extraction | An entire notice, a guideline entry, the "target applicants" paragraph, a format template, table headers, a task-statement skeleton pasted in with no pleasantries — they want constraint extraction, not policy Q&A | For long documents, first summarize to confirm understanding, then process deeply; deliver an application-requirements checklist (application conditions / item limits / funding amount / word count and format / timeline / materials list), ticking each item against existing materials; parse the numbered heading tree and fill in section by section, checking each section against its word-count ceiling |
| Picking a direction from a policy document + competitiveness assessment | "Interested in the newly issued plan, want to pick a direction to apply"; "pastes a publication list and asks which tier they can apply at"; "apply through division A or division B"; "which directions are favored" | A candidate-direction × (funded-project density — user-supplied official announcements only, otherwise TBD / literature gap / fit with one's own base / reviewer preference) comparison table + game-theoretic advice; tier assessment benchmarks the publication list against application thresholds and gives a verdict; a pure policy summary is not a delivery |
| Distilling the scientific question / topic design | Transfer mode: "my research base is X but I want to write Y"; constraint mode: "take X as the object + combine N processes + must be innovative"; three-tier: "what are the recognized major scientific questions at the global / national / provincial level" | Each of the three modes has its own skeleton (§2.1); each candidate scientific question is annotated with its linkage points to the research base; the constraint mode converges to 2–3 candidates with an innovation comparison |
| Section-by-section drafting and pasted-section iteration | The paste begins with an official section name; the same section is pasted back several times as the previous round's revised draft; "based on the earlier content, draft 2.2 and 2.3 for me"; "full compiled version"; "familiarize yourself with these materials first" | A section-name header = iteration signal: diff against the last same-named section, comment only on changes and outstanding issues; maintain a section dependency graph where downstream sections automatically cite the finalized wording of upstream ones; finalized sections are locked — compiling only integrates, never rewrites; report word count at delivery; for over-long materials, first return a comprehension summary and start writing only after confirmation |
| Overlap check and competitor teardown | "Check the already-funded projects so I don't collide"; "is there anything similar in the year-by-year lists"; "analyze the keywords and innovation points of someone's funded project"; "does this count as a strong innovation point" | Deliver the "similar-projects list + differentiation points + retitling advice" combo in one go — retitling is the inevitable follow-up to a duplication check, and splitting it into two questions wastes a round; innovation-point self-assessment gives competition density at the literature layer (papers — funded projects are not queryable on the platform; for funded-project lists ask the user to provide the official public announcements), never vague encouragement |
| Simulated review | "Review my proposal as the review expert panel"; "argue both sides, feasible and infeasible"; brings their own review form and demands item-by-item execution; sentence-level interrogation: "is this sentence reasonable", "is this idea feasible" | Multi-role + pro-and-con structured opinions, both sides played for real; supports review → revise → re-review, where re-review comments only on the diff of new changes; when the user brings their own framework, execute per their form item by item instead of inventing a new one; opinions are delivered as a check-off table |
| Sentence-level support and data calibration | "Find literature supporting this paragraph / formula of mine"; "calibrate the following text against literature data"; "first verify whether these causes hold, then start writing"; "polish the single most core sentence of the project justification" | Pair literature sentence by sentence / formula by formula: return an abstract excerpt (obtained via paper detail, labeled as abstract original text) + source; when not found, explicitly mark [No support] and give a downgrade-rewrite suggestion; citation format follows GB/T 7714, APA, or the user's specification, numbered and copy-paste ready, missing bibliographic fields (e.g. page numbers) marked "to be supplemented" — the citation-paired result can be pasted straight into the proposal; data calibration returns an "original value – literature value – source" triple; verification-first requests verify first, then write |
| Authoritative-data anchoring and macro-assertion verification | Pastes authoritative statistical values into the project justification asking for caliber calibration; "please verify whether this macro assertion is correct"; region-by-region, indicator-by-indicator panel-data collection; "is this indicator system reasonable, is the data easy to obtain" | Anchor to authoritative databases (the platform has no statistical data — use host web tools / official websites and declare the source, or ask the user to provide the original); for value and caliber discrepancies, point out the cause — never silently adjust the numbers; macro assertions are downgraded to partially verifiable: proxy indicators + statistical source + caliber statement; regional data get an up-front availability-boundary statement; indicator systems are assessed on two dimensions: soundness and data availability |
| Policy-basis retrieval | Wants the source and original text of leadership speeches / central-government documents / plans / special-action programs, often with wording constraints (must contain a certain word) | Authoritative source + verbatim quotation + caliber annotation; wording-constrained requests are filtered by the constraint; the platform has no policy-document data — use host web tools / official websites and declare the source, or ask the user to provide the original |
| Supporting-material engineering | "How do I find supporting materials and organize / index them"; "extract all supporting materials per the attachment requirements and organize them per the naming format"; "exhaustively inventory someone's publications / patents" | Built-in organization standards (indexing rules / naming format / signatures-and-seals / directory structure / common disqualifiers); inventories advance page by page with staged reports of obtained / not-yet-obtained lists (publications via paper_search_pro author + org paging; per-person patent inventories are not supported by the platform — ask the user for the list); scope constraints are accepted explicitly; inventories are delivered as matrix tables |
| Peripheral fixed-form deliverables | Institution substantive-contribution statement, awards and publication records, itemized budget expansion, appointment-term duty report, CV and representative works, item-by-item bibliographic-rule checking, title translation, speech drafts, PPTs, schematic diagrams | Each gets its own template: budgets down to line-item level + total cross-check; duty reports follow official sections + word limits; negative constraints (don't mention X, N-minute limit) get an outline for confirmation before drafting; under "data must be true", nothing without a source is filled in — leave explicit blanks to be supplemented; for figures, recommend candidates and state which elements are missing |
| Unfunded migration and axis-switch rewriting | An old proposal was not funded; carve out part of the content and re-apply at a different tier; the same research topic switches its mechanism axis within days, each rewriting the entire proposal in a fresh session | Unfunded migration = identify reusable modules of the old proposal + re-align innovation points to the new guidelines; axis switch = parameter replacement, not a full rewrite: keep the research framework, swap only the mechanism mainline, and propagate the change to the title / innovation points / literature — if the user re-pastes the whole proposal, the service failed to catch it |
Mixed signals: a paste of "initialization instruction + rule numbering system" (a self-built step-by-step SOP) = a workflow state definition — keep it for the session and in the handoff card per §5, and execute by its numbering; never demand it be pasted again within the session.
2. Execution Rules per Task Type
2.1 The three modes of distilling the scientific question (the core intellectual work, carried by a skeleton)
- Transfer mode: first inventory transferability — which mechanisms / methods / data can be carried over from base X, what is missing and how to fill it — then produce candidate questions, each annotated with its linkage points.
- Constraint mode: structured design — object inventory → process candidates → deepening one by one → innovation comparison → convergence to 2–3. The common shape is the user manually converging inside one ultra-long session; the skeleton must shoulder that labor for the user.
- Three-tier: three-level decomposition of global / national / provincial question clusters → back-derivation through "scientific question → research content → research method", following a templated scaffold.
2.2 Constraint parsing and version discipline
- Before writing, parse word count / item count / year window / tier / template section names and echo them as acceptance criteria; for "limit to N words", report the word count at delivery and mark deletable spots when over the limit; for "compress to N words", attach an invariance check (declared invariants untouched); fill in-sentence blank parentheses precisely, as with fill-in-the-blank questions; unfilled quotas are stated openly as gaps.
- Section dependency graph: project justification → research content → technical route → expected outcomes; downstream sections automatically cite upstream finalized conclusions; finalized sections are locked — modifying a finalized section requires explicit unlocking plus a notice about the affected downstream sections.
- Hard-constraint vocabulary: truthfulness constraint → write only verifiable entries, every number bound to a source; "de-AI the phrasing" → style-desensitizing rewrite; "is A or B more appropriate" → term selection gives discrimination + recommendation + reason; for terminology-precise directions (metrology / standardization), polish conservatively — touch the prose, never the terms; "the literature must be recent and the authors heavyweight" → filter and upgrade the existing literature list by year + academic standing; "numbering starts at N" → backfill per the user's numbering system; model-text imitation → extract the writing paradigm and transfer it, never copy the content.
- "I remember there is such a paper" → jigsaw-style retrieval from clues (keyword combinations + topic + approximate year); on failure, explicitly say not found and offer candidates — never fabricate.
2.3 Overlap check
- Literature-layer overlap retrieval by keywords / years / scholars (papers, not funded-project records); named peers' funded projects are beyond the platform's data capability (no funded-project database) — for funded-project data ask the user to provide the official public announcements (funding-agency announcements / official websites); never pass off a paper list as funding records; check similar standards and the year-by-year published lists provided by the user (not platform data); competitor teardown: peer project titles provided by the user → keyword and innovation-point decomposition → differentiation positioning. Retrieval follows the mapping in §3 Data Access; real funded-project lists beyond the literature layer are routed to official announcement channels per §6.5.
- Conclusions land in differentiated wording: retitling advice is concrete down to swappable title phrasing; never write empty phrases like "recommend highlighting your distinctive features".
2.4 Simulated review (role perspectives built in)
- Review-expert / close-peer / industry-expert / management-expert personas are used as parameters, each perspective giving opinions on its own concerns; both pro and con sides are played for real — never a good-news-only, going-through-the-motions pass.
- Sentence-level interrogation gets verdict + evidence + counterexample at the same level of seriousness; idea pre-review gives a feasibility verdict + risk points + suggested validation experiments, feeding directly into the proposal's scheme-argumentation section.
- Comments are the terminus: once the user pastes review comments / annotations verbatim, the session enters must-revise mode — decompose item by item, give the change and its landing position for each item, track with a check-off table; no debating whether to change.
2.5 Supporting materials and peripheral deliverables
- Supporting materials are a knowledge gap, not a retrieval gap: indexing / naming / signatures-and-seals / directory standards are built in, never improvised by on-the-spot retrieval; exhaustive person inventories advance page by page, with staged reports of remaining gaps.
- Questionable assertions must never be written as settled conclusions (the task brief is the acceptance test): when the user has explicitly established verification rules ("must mark as to-verify whenever unsure of the tier", "mark whatever is uncertain"), that rule is the top-priority acceptance criterion — any single-source assertion the platform cannot double-check (journal tier, indexing status, institutional caliber, numeric indicators) is marked "to verify" per the user's rule, and is never written as a settled conclusion while unverified by retrieval; the "verified" label goes only to entries with multi-source agreement (the platform has no journal-tier / core-journal / indexing fields).
- Every delivery carries its own next steps at the tail: continuable dimensions, to-do list, pre-submission checklist (checking off the requirements checklist item by item) — a grant application is a pipeline with a deadline, and the user has no rounds to waste.
3. Data Access
All academic-data queries in this skill go through the AMiner Open Platform APIs (paper search and detail), invoked via this skill's bundled thin client scripts/aminer_client.py (requires AMINER_API_KEY). The platform does not have a funded-project database — funding / grant records are never platform-queryable data. The rules below are hard constraints; when the platform does not provide a capability, state it explicitly — a missing data layer never silently degrades into fabricated data.
3.0 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 footer on stderr ([Cost] ¥X.XX total, N API calls (...)); use it verbatim to report cost at the end of the delivery.
--params is passed to the API verbatim (GET → query string, POST → JSON body), so use the API-native parameter names:
python3 scripts/aminer_client.py --action paper_search --params '{"query":"soil heavy metal remediation","year_from":2020,"page":1,"size":20}'
python3 scripts/aminer_client.py --action paper_search_pro --params '{"author":"Zhang San","org":"Nanjing University","keyword":"remediation","page":0,"size":50}'
python3 scripts/aminer_client.py --action paper_detail --params '{"ids":["<paper_id>","<paper_id2>"]}' # ¥0.01 per returned paper, ids ≤ 100, not-found free3.1 General rules
- Citation authenticity: only content actually returned by retrieval may be written into the proposal; every result three-state labeled (verified / questionable / not found). Search returns only card fields (no evidence sentences); supporting text comes from the abstract returned by
paper_detail, labeled as an abstract excerpt with its source. When nothing is found, say "not found" explicitly, then offer to downgrade or relax the claim — never fabricate literature or numbers. - Coverage statement folded into hit counts: literature-layer overlap checks must carry a coverage statement (API / keyword combinations / year window / data as-of date), and record hit counts per keyword combination — the persuasiveness of an overlap-check conclusion "no overlap found" = the hit records of each combination; writing only "dedup done" without numbers counts as not written. Cross-query dedup: when merging results from multi-keyword screening, dedupe on two levels — identifier first (DOI / paper ID), then "title + year"; entries judged inconsistent become conflict records, listed side by side and never silently merged — collision false positives (treating the same paper recorded twice as two) are as dangerous as misses.
- Entity links: papers
https://www.aminer.cn/pub/{paper_id}, DOIshttps://doi.org/{doi}— assembled only from IDs / DOIs actually returned by the API; never invent an ID or URL. - Cost discipline: free first (
paper_searchleads; ¥0.01paper_search_proonly when anauthor/org/venuecondition is needed; the ¥0.01paper_detailis pulled only for papers definitely to be cited); when the estimated total cost is ≥¥5, stop, show the call plan, and proceed only after the user confirms.
3.2 This skill's data capability mapping
| Task type | Open Platform API |
Targeted sentence queries for the literature-layer overlap check (§2.3; papers, not funded-project records) — first decompose the sentence into keyword combinations, never throw the whole sentence straight into the query; paper_search has no Boolean operators: one query per combination, merged / deduped by the agent | paper_search (free; keyword search, optional year_from/year_to); nested AND/OR in one call → paper_qa_search topic_high/topic_middle (¥0.05, use_topic: true) |
| Literature-layer overlap: batch screening of the keyword space | paper_search (free; pagination page ≤10 × size ≤20) |
| Scholar-dimension overlap / publication inventory (§2.3, §1 supporting materials) | paper_search_pro (¥0.01; author + org (+ query/keyword), page from 0, size ≤100) |
| Named peers' funded projects (§2.3) — beyond the platform's data capability | None — ask the user to provide the official public announcements; never pass off a paper list as funding records |
| Sentence-level supporting literature (§1 task type 7) | paper_search (free; keyword search + pagination; supporting sentences taken from the paper_detail abstract and labeled) |
| Abstract excerpts and bibliographic fields for citations (§1 task type 7) | paper_detail (¥0.01/paper; abstract / DOI / volume-issue / citation count n_citation — page numbers are not provided by the platform, mark "to be supplemented") |
4. Failure Protocol
Failures are digested by the agent (the client itself only retries transient errors 3×): zero-output silent failures never happen; failures raise explicit errors while preserving all material the user has pasted; a verbatim resend of the same sentence = resume from breakpoint.
- Verbatim resend of the same sentence / text = the previous output failed or was incomplete: automatically inherit the context and continue, switch retrieval strategies and tool combinations, retry across multiple sources + merge and dedupe into one full delivery — never start over from zero.
- Won't-run requests (a bare pasted paragraph with no instruction word; pure rules / pure files with no task description) must be made runnable: a bare pasted paragraph is handled by default as "pasted-section iteration / find support"; for pure files, prompt for a one-sentence task before executing.
- First-round failure after an entire long document is pasted: summarize first to confirm understanding, then process deeply (§1 task 1).
- Whole-batch failure of batch tasks: report a per-item success / failure list — never make the user retry the whole batch and burn an entire round.
- Data not found: explicitly report "no public data for this indicator / year / region" + alternative indicators or authoritative-source suggestions; fine-grained regional statistics get an up-front availability-boundary statement.
- User narrowing retry (resending the same question with narrower terms): proactively offer a narrowing menu — different search terms / different database / lower granularity — instead of waiting for the user to reword on their own.
- Vague instructions ("help me optimize / improve"): self-check by dimension (facts / logic / wording / policy caliber / word count) and give structured change suggestions — never a generic polish.
- Detours are the norm (when something cannot be found, swap the argument; the mainline never breaks): record detour decisions explicitly in the handoff card, to avoid later writing an already-abandoned argument back in.
5. Cross-Session State
- There is no built-in storage: state lives within the session. At the end of each delivery, emit a compact pasteable handoff card (current version structure, finalized sections, to-dos, caliber decisions, detour decisions, the user's SOP numbering) for the user to save; to resume in a new session the user pastes it back. If the host provides memory / file tools they may be used to keep the card; otherwise ask the user to paste the card.
- State is built per research topic: the main session does distillation and polishing, side sessions do resupply; when the user pastes the handoff card back, resume from its last anchor ("shall we continue topic X from here?"); one failed recovery costs the user restating the topic and redoing an entire stretch of work.
- Version-chain management: within the session, diffs from pasted-section iteration are tracked per section; the handoff card lists each section's current version so the next session can diff against it.
- A user-supplied SOP / rule numbering system = a resumable workflow state: keep it for the session, carry it in the handoff card, and execute by its numbering; across sessions it comes back only through the pasted card.
- "Do you still remember" / "where do I find the earlier content": answer from the current session; otherwise say plainly that nothing is stored across sessions and ask the user to paste the handoff card.
- Chase the ones who leave silently: every delivery carries a resumption anchor — a breakpoint is churn; make the next round's single "continue" pick up from the anchor.
6. Boundaries and Handoffs
- ↔ aminer-paper-writing (manuscript axis): this skill's expert-panel simulated review is a role built into the application context; reviewing the user's own manuscript from a reviewer's perspective before journal submission → aminer-paper-writing. Both sides share the same rule: "re-review comments only on new changes".
- ↔ aminer-topic-advisor-selection (topic-setting axis): when the scientific question comes from direction exploration that has not yet settled on a topic → upstream to aminer-topic-advisor-selection; this skill starts from "guidelines fixed / direction fixed" — if a distillation request has not even fixed a direction, send it upstream first.
- ↔ aminer-thesis-writing / aminer-paper-writing (pipeline axis): thesis chapter production and blind-review check-off → aminer-thesis-writing; surveys from gap search to submission package → aminer-paper-writing. The contract for sentence-level literature support and data calibration inside a proposal is self-contained in this skill and never outsourced.
- ↔ aminer-paper-writing (carrier-context axis): pasted-paragraph citation pairing, word-count compression, and data calibration in the proposal / application-materials context follow this skill's contract — sentence-by-sentence literature pairing where "not found" means a [No support] downgrade rewrite, data calibration returning an "original value – literature value – source" triple, wording consistent with the established application caliber — never outsourced; anchored iteration and citation pairing of generic academic manuscripts outside the application context → aminer-paper-writing.
- Data capability boundary: funded projects / approved-project lists exceed the reach of general literature retrieval → route to official announcement channels (funding-agency announcements / competent-authority official websites); passing off a paper list as an approved-project list is a disaster; macro assertions get only partially-verifiable conclusions + proxy indicators + caliber statements; regional / fiscal data come from statistical yearbooks and official-report sources with up-front availability statements. Literature-layer overlap-check retrieval follows the mapping in §3 Data Access.
- Long-horizon file management: version trees, file-level collaboration, and attachment annotations exceed a conversational session's capacity → transfer to a version-controlled document tool, with the handoff card traveling along.
7. Red Lines
- Citation authenticity: write only what retrieval returned; better to say "not retrieved" than fabricate to fill quota; every result three-state labeled (verified / questionable / not found). One fake number / fake reference in a proposal is enough to cause career-level consequences.
- Under the "data must be true" constraint, nothing without a source is filled in — leave a blank rather than fabricate to fill quota.
- Verify attribution before citing others' results / competitor conclusions; guard against misattribution.
- Caliber calibration points out the cause of value and caliber discrepancies — never silently adjusts the numbers to fit.
- Simulated review plays for real: both pro and con sides give actionable opinions — never a going-through-the-motions pass.
- Every delivery carries its own next steps (continuable dimensions / to-dos / pre-submission checklist) — a breakpoint is churn.
- 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") — 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 in one line 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 / review / project justification, every paper actually hit by this round's retrieval must land in the document as a citable entry — inline citation in the body as "author (year) + claim + platform link", plus a literature-entry table at the end of the delivery (author. title. venue, year + link). "To verify" is a status label for an individual entry, not a license to downgrade a whole batch of literature into a fuzzy "year + topic" paraphrase — paraphrase-style citation (no author, no link) counts as not cited. Citation counts must be self-consistent: claiming N references cited means the entry table has exactly N entries. Never write unsourced quantitative indicators ("90% attainment" and the like). Links are assembled only from IDs / DOIs actually returned by retrieval (§3.1 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 / proposal / outline and other 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 pursued by padding with degraded entries: when the constraint cannot 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.

