Journalism Writing

journalism-writing

Writing craft drawn from news/explanatory journalism and argumentation traditions — narrative lead (Lede→Nut graph→Kicker), dialectical structure (thesis·antithesis·synthesis), argument quality (Toulmin), and fairness (BBC due impartiality). criteria.json is the SoT for each criterion's definition, comparator, and source. The shared parsing and wiki-global state that the deterministic checks (judge=A) rely on are injected by the orchestrator (the skill is content-type-agnostic). Examples are illustrative of the target English prose.

Dialectic structure (jrn.thesis-antithesis · jrn.c-section-size · jrn.c-stance-naming · jrn.monitoring-balance)

Develop an issue as Hegelian thesis → antithesis → synthesis. State thesis and antithesis with explicit Position A / Position B bold labels (ko rendering: A 입장 / B 입장); add a C — Mediation label (ko: C 중재) only when a genuine convergence exists. The C paragraph must not run longer than the longer of A and B, so the convergence is not mistaken for the main clash. If C is not a synthesis but a meta-critique (weakening both sides at once, flagging interest bias), move it out of the dialectic frame — a meta-critique in the C slot breaks the three-part structure.

Synthesis does not pick a winner. In Hegel's terms it sublates — cancels and preserves — identifying what each side correctly grasps. Concretely, place each side's monitoring point (what one would observe if that side were right) symmetrically (jrn.monitoring-balance); a monitor skewed to one side hides an editorial verdict under hedged wording (combine with BBC due impartiality). e.g. ✅ "If tighter regulation is right, we would observe reduced consumer harm; if looser regulation is right, increased new entry" (winning conditions symmetric on both sides) / ❌ "Regulation blocks innovation, so abolishing it is right" (one-sided verdict).

Argument quality (jrn.toulmin-claim · jrn.rebuttal · jrn.qualifier)

Check each side's support structure with the Toulmin model (claim · grounds/data · warrant · qualifier · rebuttal · backing).

• Claim-Warrant (jrn.toulmin-claim) — every side pairs its claim with the grounds (data) that support it; no side asserts a claim with no grounds. e.g. ✅ "Regulation slows innovation — the grounds: new licensing waits average 18 months" (claim + grounds) / ❌ "Regulation slows innovation" (ungrounded assertion)
• Rebuttal acknowledgment (jrn.rebuttal) — concede a real weakness for each side, grounded in one of: (i) a limit the side itself admits, (ii) an internal contradiction in its logic, (iii) a design limit of its evidence (sample/timing/method). Re-citing the opposing side's evidence is NOT a rebuttal — it merely repeats the clash and loses the Toulmin value of a flaw seen from within the side. No side may be left perfectly defended. e.g. ✅ "However, this measurement is a first-generation adoption sample, so whether the same result holds at maturity is untested" (a design limit of one's own side) / ❌ "The opposing side also has many failure cases" (re-citing the opponent's evidence — not a rebuttal)
• Qualifier (jrn.qualifier) — every claim holds only conditionally; include at least one scope qualifier ("in the short term"·"on this metric"·"within 5 years"·"under this study design") so the claim is not over-generalized.

Fairness (jrn.due-impartiality)

When aggregating many topics/sides into one piece, keep length and references from skewing to one side — BBC due impartiality is proportionate to weight, not a mechanical 50:50, and privileges no side. The deterministic check signals via a max/min reference-ratio ceiling (default 3.0, see criteria.json), but a hub that is intrinsically more referenced can be normal, so human review accompanies it.