ai2-comms-structure-lector

name	ai2-comms-structure-lector
description	Use when reviewing Ai2 drafts for framing, openings, section order, redundancy, narrative flow, reader stakes, or whether the contribution is presented clearly.
version	1.0.0
keywords	["review","structure","framing","narrative","editing","comms"]

Ai2 Comms Structure Lector

Review Goal

Make the draft easier to understand without inflating the claim. Focus on structure, framing, and reader comprehension.

Check

• Does the opening answer what changed, why it matters, who it helps, and what is different?
• Does the opening name the actual artifact, method, result, or release before using abstract mood language?
• Is the main contribution framed correctly: framework, benchmark, model, dataset, result, or adoption pattern?
• Does the opening pose the genuinely open question, not a strawman the audience has already settled? When the field knows the headline (e.g., two architectures score about the same on benchmarks), frame the stakes around the real unknown — how or why they differ — not we don't know whether they differ. A false-uncertainty hook misstates the contribution's stakes; fix it to the question the work actually answers.
• Does the draft avoid a second intro after the first section?
• Are title, subtitle, captions, body, and footnotes repeating the same setup?
• Does each chart, quote, and example earn its place?
• Does each illustrative example use specific, domain-grounded detail (real entity types, realistic query shapes) rather than a generic placeholder? Prefer the concrete domain example; for practitioner audiences, use the domain-precise term with a first-use gloss (e.g. "illegal, unreported, and unregulated (IUU) fishing").
• Is the so what sharper than the background?
• Are polished sentences carrying facts, or only creating a generic sense of importance?
• Did model-assisted brainstorming narrow the piece toward a familiar thesis instead of the source's strongest, most specific contribution?
• Did richer language make a familiar idea look more original than it is?
• Does the draft follow the canonical arc (problem/context → announcement → technical → evaluation → artifacts), or have a justified source-driven reason to deviate?
• Are limitations surfaced inline as research findings, or quarantined into a separate end-of-post limitations section?
• When a headline or update banner states the central fact, does the body restate it? Cut the repeat; let the headline carry the fact.
• When a new update banner supersedes a prior one, does the post replace the prior banner or stack on top of it? Default to replace; stack only when the prior banner is independently true and worth a visible changelog.
• After a plain-language or accessibility pass, did a capability list silently shorten (three authoring paths quietly became two), or did a feature get dropped without a flag? Re-diff against the source. See .agent/skills/ai2-comms-writer/references/revision-passes.md Pass 8.
• In peer or open-tool comparisons, is the peer defined before it is compared? Does any sentence diminish the peer rather than contrast factually? Does the comparison close with an honest [Peer] is the better fit when ... caveat? See .agent/skills/ai2-comms-style-source/references/claims-and-benchmarks.md (Peer And Open-Tool Comparisons).
• For a technical research blog, does difficulty ramp gradually — an accessible intro and a high-level the idea section before the dense formalism, with the dense block led by the most intuitive technical result rather than the most abstract? See .agent/skills/ai2-comms-style-source/references/house-style-patterns.md (Difficulty Ramp For Technical Research Blogs).
• Does the post end with a closing call to action that names the artifact and action — for a paper amplification, a link to the paper? A research blog that never points the reader to the paper is missing its primary CTA; flag the absence even when the prose reads finished. (Claims-lector owns whether the link resolves; structure owns that a CTA is present.) Extension for X and Bluesky threads: the opening post carries the canonical destination link too, so check both ends — a thread about a published blog whose post 1 has no way to reach it is missing its in-timeline CTA. See .agent/skills/ai2-comms-channel-adapter/references/platform-patterns.md (Put the primary link in the opening post).
• Is every load-bearing symbol, acronym, and named quantity defined on first use for the intended audience (n, MSE, bandwidth), rather than introduced cold? See .agent/skills/ai2-comms-style-source/references/voice-and-prose.md (Define Symbols And Acronyms On First Use).
• Setup-payoff: does a framing or setup sentence that promises something deliver it in the body? A line like the probes show where these gaps arise, and why must be followed by the why; a promise of explanation the body never pays off is a dangling setup — either deliver the payoff or drop the promise.
• Redundant preview clause: does a vague setup or gloss clause sit directly before a sentence that already renders the same claim concrete? Cut the gloss rather than trying to make it concrete — the following sentence is the concrete version, so rewriting the gloss only double-states it. Anchor: Asta's literature search, built to work the way an expert does—in steps, with reasoning. Its new default mode ... interprets your query, evaluates whether the results answer it, and keeps searching if they don't → the in steps, with reasoning gloss was cut, not rewritten, because the next sentence (interprets → evaluates → keeps searching) is that concrete version; applied on both the compressed thread and the long post, so it is not channel-specific. Complement of Setup-payoff above — that catches a promise the body never delivers; this catches a preview the very next sentence already delivers.
• Thesis restated too many times: is the piece's central contrast or mechanism re-explained in the intro preview, each results section, and the conclusion? Keep two or three load-bearing statements (the setup, one results instance, the conclusion) and trim the rest to a clause; each restatement should add a new angle, not repeat the same one. This is broader than the second-intro and headline-restated checks above — it catches a thesis spread thin across many sections.
• Conclusion recaps instead of reaching: does the close re-state the body rather than carry the forward takeaway? A conclusion that loops back to the opening tends to re-quote the body's numbers and ledger and re-list the opening's stakes — redundant with what the reader just read. A close should carry the reach, not re-print the spec sheet the body already gave. Test: name what the closing paragraph adds beyond recap. If every sentence points backward to a number, stake, or framing already stated, rewrite it to state the forward takeaway instead. This is narrower than Thesis restated too many times above — that catches one idea spread across many sections; this catches the conclusion specifically doing the body's job over again.
• Heading matches the kind of evidence under it: does a section heading describe the evidence it governs? A controlled or synthetic result (a planted-detail probe) filed under a what real text shows heading blurs the observational-vs-controlled distinction the piece is drawing; give a distinct analysis its own heading.
• Term introduced before use: is a participial descriptor (the reprocessed pool, the deduplicated set, the calibrated score) used as a given before the process it names has been introduced in the passage? Introduce the action — or what was acted on — first, then refer to its output by the participle. See .agent/skills/ai2-comms-style-source/references/voice-and-prose.md (Introduce A Process Before Its Participle).
• Quote density in interview or researcher pieces: is the section narrative carried by a few load-bearing verbatim quotes, or a string of short quotes sandwiched between non-quote fragments? The sandwich reads as a transcript and flows poorly — paraphrase the connective substance and keep only the distinctive quotes. (A Q&A format the user asked to drop in favor of a flowing post is the same flag at document scale.) See .agent/skills/ai2-comms-style-source/references/source-fidelity.md (Paraphrase-Forward Handling Of Interview And Researcher Quotes).
• Named quantity or set referenced before it is introduced: does the copy lean on a specific count or named set (these 16 benchmarks, the three baselines) in a section before that set has been introduced — e.g., the data or setup that names them appears later? Either move the reference after the introduction, or reword to drop the forward dependency (the benchmarks we study). Sibling to Term introduced before use, for named quantities rather than participles.
• Balance survives a positive lean: when negative or failure cases are cut to lean favorable (a make it more positive steer, or dropping the projects that went badly), does at least one genuine limitation survive, and does the synthesis stay two-sided — what it was good for and where it fell short? An all-upside recap reads as a puff piece and undercuts the house transparency value. Anchor: a hackathon recap that cut the weakest tool did poorly projects but kept one honest limitation (a tool's shortfall on simulated data) and a balanced what they agreed on verdict. Pairs with ai2-comms-claims-lector SKILL.md (Favorable-framing re-diff).
• Results-recap drift in an enablement piece: in a testimonial or adoption post, is the piece spending its length on what the study measured — figures, correlation values, model rosters, validation experiments, every case study — when the contribution is what the artifacts made possible? Cut to the artifacts, one or two findings, and the weird specific detail; the verified numbers can stay in the draft's notes rather than the body. Also check the ordering: the artifact map belongs in the first third, not as a closing synthesis, when the piece is a testimonial. See .agent/skills/ai2-comms-style-source/references/release-types.md (Genre Overlays, Testimonial).
• Headings on a short piece: does a sub-1,000-word testimonial or adoption post carry section headings it does not need? At that length the house form is continuous prose; three headings over 600 words chop a piece that reads better as one movement. Flag headings added out of habit, not headings a genuinely long piece needs.
• Significance announced before its substance: does a sentence label a thing's importance (a question no one has answered, its most striking finding, the real surprise) before the reader has been told what the thing is? State the thing, then label its significance. Anchor: It surfaced a question no one had even asked. There are two standard ways … (the label lands before the content, halting) → reordered to setup → finding → Why it flips is an open question. Sibling to Term introduced before use and Named quantity referenced before it is introduced — those concern a name used before its definition; this concerns a significance label placed before its content.

Output

Lead with findings, ordered by importance. Then provide a revised opening or outline only when useful. Do not rewrite the full piece unless asked.

Use references/review-checklist.md for a fuller pass. For the canonical arc, opening forms, and heading conventions referenced above, see .agent/skills/ai2-comms-style-source/references/house-style-patterns.md.

Related Skills

Run alongside ai2-comms-claims-lector for full review coverage: structure handles opening form, arc, redundancy, and framing; claims handles comparison scope, benchmark discipline, approval-sensitive copy, and house-lexicon pairing. The two passes are independent and can run in parallel.

Hand revisions back to ai2-comms-writer. Voice authority is ai2-comms-style-source. If the draft includes benchmark visuals, also route to ai2-comms-chart-designer; if it will cross-post, route to ai2-comms-channel-adapter after both lectors clear it.