Outlining : IMRAD structure, claims, and the interview

This is the spec phase. Its output is the input to drafting. Nothing is written as manuscript prose here. Three deliverables:

1. .writing/outline.md, IMRAD structure with 3-7 key claims per section.
2. .writing/claims/section<NN><slug>.md, one file per section, a YAML list of claim stubs prose will later bind to.
3. .writing/metadata.yaml, authors, availability, writing_profile, and reporting fields filled.

The outline is a claim spine, not a table of contents. Every bullet is an assertion the paper will make ("We show X"), not a topic ("Overview of methods"). The claim-first discipline only works if outlining produces real claims.

Two load-bearing contracts

• Filename stems. Sections whose stem ends in abstract, references, or _acknowledgments are exempt from paragraph claim-tags. Every other section's .tex file must carry % claim: id-tagged prose. If you add a new exempt section, treat its slug like abstract/references. The contract covers .tex files, not .md.
• Abstract is citation-free. Any abstract stem emits no \cite/\citep/\citet/\nocite/any \cite command and no % claim: tag, it summarizes the paper's own claims in prose. It still carries BPMRC structural tags. Do not create claims/section00_abstract.md.

Optional term-definition-before-use: copy templates/glossary.md to .writing/glossary.md to activate % define:/% use: ordering checks; add a glossary entry when you add the term's claim stub.

Process

Step 1 : Capture the contribution

Lock down three things and write them to .writing/findings.md:

• Core contribution in one sentence. Everything downstream anchors to this.
• Target venue or venue class, which sets the reporting guideline and structural norms.
• Unit of evidence (dataset, trace, benchmark, proof), what counts as EVIDENCE later.

The contribution sentence usually names the system or method. Those names are domain terms: record them in the repo-root CONTEXT.md as they are decided (superpower-writing:domain-glossary).

If the contribution is not yet decided, do not force a half-formed idea through IMRAD, that produces a fake outline masking missing thinking. Route to superpower-writing:idea first (it generates candidate directions, checks novelty against prior work, and evaluates the survivor against a top-venue bar, all advisory). The surviving direction becomes the contribution sentence. When the contribution is already decided (the common case), skip ahead.

Step 2 : Ground in the literature

Before proposing claims, synthesize the landscape with superpower-writing:literature: what is known, contested, missing. Record each finding in .writing/findings.md under Research Findings (citation + one-sentence takeaway + which IMRAD section it informs). Two or three passes is normal; a single sweep usually misses a baseline or a competing result.

Step 3 : Draft a rough IMRAD skeleton

Pick the paper type first: technique (new method, the default) or benchmark/dataset (the artifact is data + an evaluation protocol). They use different skeletons; do not mix them. Ask via AskUserQuestion when unclear.

Write the skeleton into .writing/outline.md with 3-7 claim bullets per section, using the section frameworks below as readable guidance, they encode real rhetorical structure, not a checklist to satisfy mechanically. For the full template and examples of any framework, read the matching file under ../drafting/references/section-standards/ (resolve by slug: <NN><slug>.md, else the single *<slug>.md).

Adapt to venue variations: combined Results and Discussion (interleave [R]/[I]), conclusion folded into Discussion, ML-conference format (Related Work in appendix). Medical/biology venues use IMRAD-strict (skip Background/Motivation/Related Work; follow CONSORT/STROBE/PRISMA). For a benchmark paper, §Methods and §Results become evaluation-framework / construction / findings sections, read ../drafting/references/section-standards/benchmark_README.md first.

Step 4 : Interview to sharpen (interleaved loop)

The skeleton is cheap and reviewable; sharpen it before it fans out into dozens of claim files. Interview the user with AskUserQuestion, one or two questions per round, refine the skeleton in place after each answer, and loop until it converges. Probe these dimensions, self-resolve any the draft, findings, or literature already answer; spend questions only on what genuinely needs the user:

• Contribution boundary. Is every bullet in service of the one-sentence contribution? What is deliberately out of scope?
• Claim → evidence. For each key claim, what evidence backs it, a dataset, a run you have, a proof, a citation? Which claims have no data yet?
• Venue framing. What does this venue expect structurally, and how should the contribution be framed for its reviewers?
• Novelty. What is new relative to the closest prior work, stated concretely (not "we are the first")?
• Reviewer attacks. What are the three most likely reviewer objections, and does the outline pre-empt each?
• Narrative spine. Does every Introduction gap have a Discussion bullet answering it, and vice versa?

Record non-obvious answers (rejected framings, known weaknesses, missing data) to .writing/findings.md. Only once the skeleton holds up does it earn claim stubs.

Step 5 : Materialize claim stubs

For every bullet, add a YAML entry to the matching .writing/claims/section<NN><slug>.md (stem matches the manuscript file that will pair with it):

- id: meth-c1
  CLAIM: 1,247 jobs from the Google cluster trace 2019
  EVIDENCE:
    - type: dataset
      ref: google-trace-2019
    - type: citation
      doi: 10.xxxx/...
  STATUS: stub

• id: short, section-prefixed, unique (meth-c1, res-c2). Prose binds via % claim: meth-c1.
• CLAIM: one sentence restating the bullet, no hedging, no citations in this line.
• EVIDENCE: what must be true to survive verification. type ∈ citation (needs doi/arxiv/ref), dataset (ref), figure/table (id), analysis (description; on systems papers use the structured source fields from systems-evidence-contract.md), artifact (first-party ref + role).
• STATUS: always stub here. Drafting advances it to evidence_ready; claim-verification to verified. Leave placeholders like doi: 10.xxxx/... when the source is unknown, but never invent DOIs.

Stubs are intentional: the claim-first discipline forbids prose against stubs, so claim files must mature before prose references them.

Step 6 : Fill metadata.yaml

Replace every TODO/empty key with a real value: authors (name/affil/orcid/coi), preregistration, data/code availability, and reporting_guideline (CS/ML/systems papers normally none). Verification checks presence, not truthfulness, the user owns that; use null for fields that truly do not apply. Ask via AskUserQuestion for anything you cannot infer; never fabricate names, ORCIDs, grants, or URLs.

writingprofile gate. Set writingprofile: systems or default. Fill it from the user's explicit statement or a clear systems venue; when unsure, ask once, do not infer from reportingguideline: none (ML and theory papers use none too) and do not scan the body to guess. For an older project missing the field, pause the first relevant flow, ask once, write it, then continue. Structured analysis evidence is required on all profiles; writingprofile only toggles the systems wording and figure references.

Optional Zotero seeding: when metadata.yaml has zotero.enabled: true and a collectionkey, seed citation EVIDENCE from that collection via zoterogetcollectionitems (record source: zotero + zoteroitemkey). Skip silently when disabled. Never push new items during outlining, that is a drafting concern.

Step 7 : Self-review (readable checklist, not a linter)

Before handing off, read the outline once and confirm:

• Placeholders. No stray TODO/xxxx beyond intentional evidence placeholders; unresolved items marked [NEEDS-EVIDENCE] with a one-line rationale in findings.md.
• Scope. Every claim serves the one-sentence contribution; tangential bullets move to future-work notes.
• I↔D narrative. Every Introduction gap has a Discussion answer, and every Discussion interpretation points back to a Results bullet.
• Framework fit. Each section roughly matches its framework's shape and order (e.g. Introduction runs Territory → Niche → Occupy; Results has one [R] per [RQ]). This is a judgment read, not a grep, fix obvious breaks, do not police label counts.
• Challenge → module → contribution spine. Each prior-work limitation maps to a challenge, each challenge to one Methods core module, each module to a contribution, each contribution to a section. A spine that does not close end to end reads as incoherent regardless of how strong any single piece is, reconcile before drafting.

Step 8 : Handoff

Update .writing/progress.md (mark outlining complete with section and claim counts). Then hand off to superpower-writing:drafting, which resolves each claim's evidence and writes the LaTeX prose section by section.

Key principles

Claims, not topics. A bullet must have a truth value and bind to evidence.

The interview is the point. A well-structured outline of the wrong paper is the expensive mistake. The interleaved interview exists to catch a wrong contribution, a missing baseline, or an unanswerable reviewer attack while the outline is still cheap to change.

Metadata gate now, not later. Unresolved authorship, availability, and writing_profile compound; resolving them while the paper's shape is malleable is cheap, at the submission deadline it is not.

Don't fabricate. No invented DOIs, ORCIDs, or grant numbers. When information is missing, ask or mark [NEEDS-EVIDENCE], fabrication in outlining contaminates every downstream stage.