Immutable. This exact content is served forever at /api/v1/blob/b6e9caf9572e6716.
---
name: awesome-abstract
description: >
Reconstruct a paper's argument, audit the evidence chain, then write or
revise a top-tier CS/AI/ML/NLP/CV abstract as a closed-loop scientific
argument rather than a compressed introduction. Use when the user asks
to write, rewrite, diagnose, or audit an abstract; mentions 顶会摘要 /
Gap–Method–Evidence–Resolution / 证据链 / closed-loop abstract; or wants
a conference or journal abstract that closes one specific limitation.
Not for sentence-only polish (use nature-polishing) or full-manuscript
drafting of intro/method/experiments (use nature-writing).
version: 1.0.0
---
# Awesome Abstract
A top-tier abstract is not a compressed introduction. It is a closed-loop
argument: one limitation of prior work is named, the method addresses that
limitation directly, and the evidence shows the limitation is resolved.
```text
Find ONE precise Gap
→ design the Method around it
→ collect Evidence that tests it
→ write so the reader can see the Gap close
```
Do not jump to prose. Reconstruct the argument, audit it, then write.
## When to use
- Write, rewrite, or critique a CS/AI/ML/NLP/CV abstract
- The current draft reads like a shortened intro or an experiment inventory
- The user wants the Gap–Method–Evidence–Resolution loop made explicit
## When not to use
| Request | Use instead |
|---|---|
| Sentence-level English polish only | `nature-polishing` |
| Full intro / method / experiments draft | `nature-writing` |
| Novelty / prior-art overlap | `scoop-check` |
| Literature search | `paper-search` |
## Hard rules
1. Do not invent results, metrics, datasets, models, comparisons, or limitations.
2. Do not draft the abstract until the Argument Card and evidence map exist.
3. One core Gap. Secondary results may appear only as evidence that this Gap
closed, not as a second unsolved problem.
4. Every major concept introduced early is an **anchor** and must be answered
later. Orphans are deleted, weakened, or given evidence.
5. Claim strength may not exceed evidence scope.
6. The last sentence must resolve the original Gap, not merely report strong
performance.
7. If the loop cannot close, deliver the diagnosis and a scaffold with
`[EVIDENCE NEEDED: …]`. Do not hide the hole in fluent prose.
## Mode
| User signal | Mode |
|---|---|
| No usable abstract | `write` |
| Pastes or names a draft | `revise` |
| 检查 / 诊断 / audit / 证据链 | `audit` |
| 先拆逻辑 / extract only | `extract` |
Default: `revise` if a draft is present, else `write`.
Honor a stated venue or word limit. Otherwise write **150–220 words**, one
paragraph. Use two short paragraphs only for journal abstracts when the venue
needs a slightly wider close.
## Workflow
Execute in order. Mark each phase done before the next. Open reference files
only when the phase says to.
### Phase 1 — Extract
Read the source (paper, notes, tables, or draft). Prefer intro + method +
experiments + conclusion. Treat an existing abstract as a claim to verify, not
as ground truth.
Fill the **Argument Card**. Tag each field `stated` or `inferred`.
```text
SETTING field + specific technical setting (not "AI has progressed")
PRIOR what existing work does, or what proxy it uses
GAP Existing approach A fails to capture B because C
NEED capability required to close the Gap
METHOD name + what it does in relation to the Gap
CLAIMS[] each claim, including strength words (better / robust / …)
EVIDENCE[] each result + scope (datasets / tasks / models)
RESOLUTION one sentence: how the Gap is actually closed
UNCLOSED claims without evidence, or evidence that does not test the Gap
```
If there are no results, stop after this phase. Ask only for the missing
evidence, or return a scaffold. Do not write a results sentence.
### Phase 2 — Select ONE Gap
Ask: **If this paper could solve only one problem, what is it?**
Compress the winner to:
> Existing approach A fails to capture B because C.
A usable Gap is small, specific, important, testable, and solvable by *this*
method. Reject field-wide complaints, laundry lists, and gaps the method
does not touch.
If several candidates survive, rank by (1) alignment with the method,
(2) testability, (3) specificity, (4) importance. Keep one. Demote the rest
to evidence or drop them.
### Phase 3 — Build the maps
**Anchor map.** For every major term in SETTING / GAP / NEED, write the later
payoff. Typical shape:
```text
early concept later answer
───────────────── ──────────────────────────
technical setting → efficiency / measurement / …
failed prior proxy → inconsistency, then replacement
target quantity → the proposed signal or method
named failure mode → the result that reduces it
```
An early term with no later answer is an orphan.
**Evidence map.** For every major claim:
```text
Gap → Claim → Experiment → Result → Interpretation → Resolution
```
Open [references/evidence-audit.md](references/evidence-audit.md) and fill the
claim–evidence matrix, including the generalization ladder.
### Phase 4 — Loop gate
Walk this chain out loud before writing:
```text
BACKGROUND → GAP → MOTIVATION → METHOD → EVIDENCE → RESOLUTION
│
└── must restate the Gap
```
Fail the gate if any of these hold:
- Method does not follow from Gap (missing Motivation)
- Evidence does not test the Gap (inventory instead of a chain)
- Resolution is "SOTA / strong results" while the Gap was a different problem
- A strong claim has no row in the matrix, or its scope exceeds the data
Repair the card, or label the break, before Phase 5.
### Phase 5 — Draft
Open [references/six-anchors.md](references/six-anchors.md). Write in argument
order, not paper-section order.
Audience: a competent CS researcher **outside** this exact sub-area. Explain
the object before naming it. Keep one stable term per concept.
In `revise` mode, keep only sentences that still earn a seat on the loop.
Do not preserve a polished sentence that breaks the argument.
If a sentence-level example would help, open
[references/worked-example.md](references/worked-example.md).
### Phase 6 — Audit, then deliver
Open [references/anti-patterns.md](references/anti-patterns.md) and run the
checklist below. Fix or weaken before showing the abstract.
```text
□ Background is the technical setting, not a field pep talk
□ Exactly one core Gap, written as A fails to capture B because C
□ Motivation makes the method the obvious next move
□ Method is defined by how it attacks the Gap
□ Every major claim has a matrix row
□ Every early anchor is recalled
□ Numbers support a claim (generality, effect, comparison) rather than decorate
□ No claim outruns its evidence
□ Final sentence resolves the original Gap
□ A non-specialist can state the problem after one read
```
In `audit` or `extract` mode, do not rewrite unless asked. Still return the
card, maps, and verdict. For each failed box, give a one-line repair.
## Output
Return these blocks in order. Omit `Abstract` only in `audit` / `extract`.
1. **Abstract** — plain prose, not a code block.
2. **Argument card** — one line each for SETTING, GAP, NEED, METHOD, RESOLUTION.
3. **Claim–evidence matrix** — `Claim | Evidence | Scope | Status`
(`supported` / `overclaim` / `missing`).
4. **Audit** — `pass` or `fail`, plus the loop breaks or weakenings applied.
5. **Cuts** — dropped claims, orphan anchors, and why.
If the source is Chinese or mixed, write the English abstract first, then add
short Chinese notes on structural choices only.
Do not invent a prettier loop than the paper can defend.