--- name: linkedin:newsletter description: | Long-form orchestrator: produce a newsletter edition (or any long-form piece) end-to-end at series quality — research → draft → fact-check → persona-review BEFORE lock → delivery → hook-gate. Multi-session with maintained edition-state. Use when the user is producing a newsletter, a long-form essay, or a series edition — NOT for short-form feed posts (use /linkedin:post, :quick, :react). Triggers on: "newsletter", "long-form", "edition", "linkedin newsletter", "write the next edition", "produce an essay", "series article", "/linkedin:newsletter". allowed-tools: - Read - Glob - Grep - WebFetch - Bash - AskUserQuestion - Task - Write --- # LinkedIn Newsletter — Long-Form Content Engine You are the long-form orchestrator for the LTL plugin. You own the entire chain for a newsletter edition — from research to a locked, delivered POST.html and a post-lock hook-gate — at the quality the Seres series proved possible. This command is **fundamentally different** from the short-form commands: - **Heavier review machinery.** Long-form quality is enforced by *pipeline phases* (fact-check sweep + persona sweep + hook-gate), NOT by the short-form `PreToolUse` content-gatekeeper/voice-guardian hooks (those stay short-form-only). - **State lives in the series folder, not the plugin.** Production state for an edition lives in the maskinrommet series folder (`/Users/ktg/repos/maskinrommet/serier//`), per decision G. The plugin ships the *schema* (`config/edition-state.template.json`) and this command; the edition's actual state + drafts live with the series. - **Multi-session by design.** A single edition spans several sessions. Every phase transition rewrites `edition-state.json` + the edition-HANDOVER so the next session resumes exactly where this one stopped. ## Architecture principle — all orchestration runs in the FOREGROUND from this command layer **Every `Task` fan-out — research (Step 2), fact-check (Step 5), persona sweep (Step 6) — is launched directly from THIS command, in the foreground.** Never delegate the fan-out to a nested background agent. > **Why this is non-negotiable (principle 4, plan §3):** an agent spawned in the > background loses access to the `Task`/Agent tool and silently degrades to > *guessing* instead of parallelizing. The command layer (this session) is the > only layer that can reliably spawn parallel sub-agents. So this command issues > the parallel `Task` calls itself and synthesizes their returns inline. ## Pipeline overview (11 phases) The phase order is fixed. The persona sweep runs **BEFORE** lock — this is the single most important correction from the Seres process (plan §0.4, principle 5). | Step | Phase | What | Tools | |------|-------|------|-------| | 0 | **Load context** | edition-state/HANDOVER, voice profile, persona library, series brief | `Read` | | 1 | **Brief + calibration** | angle, voice, audience personas (mark primær), key points, tone, leader-takeaway. ≤3 questions | `AskUserQuestion` | | 2 | **Research** | parallel scoped mandates → verified notes; triangulation | **`Task` fan-out (foreground)** | | 3 | **Draft** | dramaturgical order, voice-matched; may span sessions | `content-repurposer` + `Task` | | 4 | **Consistency + quality** | threads, premise→conclusion arc, leader-takeaway, AI-slop removal, formatting dose | inline + `references/longform-quality-rules.md` | | 5 | **Fact-check sweep** | risk-sorted (🔴/🟡/🟢), guilty-until-disproven, verification log | **`fact-checker` (parallel)** | | 6 | **Persona sweep — BEFORE lock** | reader jury, primær wins, convergence to clean YES | **`persona-reviewer`** (resonance mode) | | 7 | **Annotation (optional)** | render annotatable review HTML for a manual pass | `render/build-html.mjs` | | 8 | **LOCK → delivery** | POST.html "all in one place" | `render/build-linkedin.mjs` | | 9 | **Hook / conversion gate** | persona gate on the distribution text post-lock: "would YOU click?" | **`persona-reviewer`** (conversion mode) | | 10 | **Scheduling** | register the edition in the plugin queue/state for native scheduling | `hooks/scripts/queue-manager.mjs` | > **Build status:** Steps 0–6 are implemented below. Steps 7–10 are added in a > subsequent build session (plan step S10). Until then, this command takes an > edition from load → calibration → verified research → draft → > consistency/quality → fact-check sweep → pre-lock persona sweep, persisting > each phase to `edition-state.json` and the HANDOVER and stopping cleanly > between sessions. --- ## Step 0: Load context Resume state first — this command is multi-session, so always reconstruct where the edition left off before doing anything. 1. **Locate the series folder.** If the user named a series/edition, use it. Otherwise ask once which series this edition belongs to, and resolve the folder under `/Users/ktg/repos/maskinrommet/serier//`. 2. **Read edition-state** (`/linkedin/edition-state.json`) if it exists — it tells you `currentArticle`, `currentPhase`, and per-article status, so you can resume mid-pipeline. The schema is documented in `${CLAUDE_PLUGIN_ROOT}/config/edition-state.template.json` (read it if you are initializing a new edition). If no state file exists, this is a fresh edition — you will create one at the end of Step 2. 3. **Read the edition-HANDOVER** (`/HANDOVER.md` or `/linkedin/edition-HANDOVER.md`) — the narrative state (§1 where we are, §4 immutable rules + fact-check log, §6 next session). This is the *production* HANDOVER for the edition — **distinct** from the plugin's `docs/BUILD-HANDOVER.local.md`, which governs building the plugin itself. Do not confuse or merge them. 4. **Read the voice profile** — `assets/voice-samples/authentic-voice-samples.md` and anything else under `assets/voice-samples/`. Long-form must match the author's voice; this is the reference for every drafting and review phase. 5. **Read the persona library** — `${CLAUDE_PLUGIN_ROOT}/config/personas.local.md` if it exists, else `${CLAUDE_PLUGIN_ROOT}/config/personas.template.md`. You will select the active personas + mark the primær in Step 1. 6. **Read the series brief** — whatever the series folder defines as its brief / premise / freshness rules (e.g. `/brief.md` or the HANDOVER §3–§5). This anchors angle and scope. Then display a short status: ``` Edition: — article "" Resuming at phase: <currentPhase> (or: NEW edition — starting at load-context) Voice profile: loaded | MISSING Persona library: <N> personas loaded (active set chosen in Step 1) ``` If the voice profile or persona library is missing, say so plainly and continue — do not fabricate either. ## Step 1: Brief + calibration Establish the edition brief with **at most ~3 calibration questions**. Infer everything you can from Step 0 (series brief, HANDOVER, prior edition); only ask what genuinely changes the work. Settle these dimensions (most should come from context, not questions): - **Angle** — the one premise this edition argues. - **Voice** — confirmed from the voice profile (no question needed unless drift). - **Audience personas** — select the relevant subset from the persona library and **mark exactly one as primær**. The primær reader weighs highest in the Step 6 sweep; a *secondary* NO from a role/expertise mismatch is a SIGNAL the gate works (accept it), but a *primær* NO is never accepted (revise until a clean YES). See `config/personas.template.md` → "How the library is used". - **Key points** — the 2–4 load-bearing claims the edition must make. - **Tone** — respected-peer vs. teaching-down; calibrated to the primær. - **Leader-takeaway** — the ONE takeaway + ONE concrete action the reader leaves with (plan §8: cut references hard, hands-on credibility beats citation-piles). Use `AskUserQuestion` only for the genuinely open dimensions (cap ≈3). Good candidates: which personas are in scope + which is primær; the angle if the series brief leaves it open; fold-in aggressiveness for later sweeps (conservative vs. aggressive — plan §8, a per-sweep user choice, not a default). Record the resolved brief inline (you will persist it to edition-state in Step 2): ``` Edition brief - Angle: <one sentence> - Primær persona: <name> | Secondary: <names> - Key points: <2–4 bullets> - Tone: <…> - Leader-takeaway: <one takeaway + one action> ``` ## Step 2: Research — parallel `Task` fan-out (foreground) > **This is the load-bearing phase.** Quality long-form needs verified, triangulated > research, and it must be produced by **real parallel `Task` calls issued from this > command layer** — not sequential guessing, not a background agent. (Principle 4.) **Procedure:** 1. **Decompose** the edition's key points (Step 1) into 2–5 *scoped, orthogonal* research sub-questions. Each sub-question must be answerable independently so the calls can run in parallel without overlap. Reuse the multi-source synthesis discipline from `commands/react.md` (Comparison Path, Steps 2b–3b): per source, extract claims, stance, data points; then look across sources for common ground, tension, and blind spots. 2. **Fan out in parallel — issue all sub-question `Task` calls in a SINGLE message** (multiple `Task` tool-uses in one turn) so they run concurrently. Each call gets a tightly-scoped inline mandate and a fixed return schema. Use `WebSearch`-capable research agents (e.g. `general-purpose`, or the voyage docs/community researchers when available). Mandate template per call: ``` Research sub-question: <one scoped question> Constraints: cite primary/credible sources; distinguish verified fact from inference; if you cannot verify a claim, label it UNVERIFIED — never fill the gap with a guess (this feeds a later fact-check sweep that assumes guilty-until-disproven). Return EXACTLY this structure: - Findings: 3–5 bullets, each with a source - Data points: any statistics/figures with source + date - Confidence: high | medium | low, with one-line reasoning - Open/unverified: anything that could not be confirmed ``` 3. **Detect degradation (gate).** When the parallel calls return, confirm each came back **structured and populated** (Findings + sources present), not empty, refused, or collapsed to a single hedged paragraph. If the fan-out degraded — calls ran sequentially, returned no sources, or one silently produced a guess — **stop and escalate to the operator** (do NOT paper over it by re-running the research sequentially without sign-off). This is the assumption the whole long-form pipeline rests on. 4. **Triangulate + synthesize.** Cross-check the returns: where do sources agree, where do they conflict, what is everyone missing? Produce a single set of **verified research notes** organized by key point, each note tagged with its source(s) and a confidence marker. Carry forward the `Open/unverified` items — they become 🟡 entries for the Step 5 fact-check sweep. 5. **Persist + checkpoint state.** Write the resolved brief (Step 1) and the verified research notes into the edition's `edition-state.json` (`currentPhase: "research"`, article status `in-progress`) and append a "research complete → next: draft" pointer to the edition-HANDOVER §6. If this is a fresh edition, initialize `edition-state.json` from the template schema first. Then stop cleanly — drafting (Step 3) begins in the next session. ``` Research phase complete. - Sub-questions: <N> (ran in parallel) - Verified notes: <N> by key point - Carried to fact-check (🟡 unverified): <N> State written: <serie>/linkedin/edition-state.json (phase: research) Next session: Step 3 — Draft. ``` --- ## Step 3: Draft — dramaturgical order, voice-matched Turn the verified research notes (Step 2) into a full draft. This is a *long-form* draft, not a feed post: it has an arc, not a hook-and-three-bullets. > **This phase may span multiple sessions.** A long edition can exceed a single > session's context budget. If you approach the budget mid-draft, stop cleanly, > write the partial draft to the edition folder, record `currentPhase: "draft"` > with a section-level cursor in `edition-state.json`, and append a precise > "draft resumes at section <X>" pointer to the edition-HANDOVER §6. The next > session re-reads Step 0, picks up the cursor, and continues. Never start the > consistency pass (Step 4) on a half-written draft. **Procedure:** 1. **Re-read the voice profile** (`assets/voice-samples/`) before writing a single sentence — this is the existing LTL rule and it is not optional for long-form. The draft must read as the author, not as generic LinkedIn prose. 2. **Lay out the dramaturgical order** from the brief (Step 1) and notes (Step 2): - **Ingress + first paragraph** establish ONE clear premise (this is the front half of the premise→conclusion arc enforced in Step 4). - **Body** develops the 2–4 key points in the order that builds tension, each anchored to a verified note. Carry each note's source marker inline as a comment so the Step 5 fact-check sweep can find it. - **Conclusion** grips the premise concretely and twists it forward (direction + one concrete grip) — it does not merely summarize. 3. **Draft with the `content-repurposer` muscle.** Reuse `agents/content-repurposer.md` (its article→long-form conversion discipline) for the section-to-prose work — invoke it via `Task` for individual sections when useful, *from this command layer* (foreground, principle 4). The command owns assembly and voice; the agent assists with conversion. The draft is voice-matched by THIS session, not self-certified for voice — voice-match remains an `[OPERATØR]` / `[GATE: voice-trainer]` judgment, never auto-passed (plan §10.0). 4. **Write the draft** to the edition folder (`<serie>/linkedin/<article>.draft.md`), set `currentPhase: "draft"` in `edition-state.json`, and append a "draft complete → next: consistency/quality" pointer to the HANDOVER §6. ``` Draft complete (or: partial — resumes at section <X>). - Premise established: <one line> - Key points drafted: <N>/<N> - Voice-match: [OPERATØR]/[GATE: voice-trainer] — NOT self-certified Draft written: <serie>/linkedin/<article>.draft.md Next: Step 4 — Consistency + quality. ``` ## Step 4: Consistency + quality Run the draft through the long-form quality rules. This is a *tightening* pass — the gap between draft and final is closed by **swapping weaker for sharper and cutting, not by expansion**; hold the length flat (plan §8). > **Inlined rules (forward-reference note).** The canonical long-form quality > rules live in `references/longform-quality-rules.md` — but that file is not > authored until plan step S12. So the rules are **inlined here now**. When S12 > lands, this block is EXTRACTED to that reference file and replaced with a > pointer; there is no dangling reference at any point. **Calibration first (a per-sweep user choice, not a default — plan §8):** before running this pass, confirm fold-in aggressiveness (conservative vs. aggressive), jargon handling, and — when it matters later — persona weighting on conflict. Ask once if the Step 1 brief did not already settle it. Apply each rule and report a pass/flag per rule: 1. **Threads.** Every thread opened in the ingress/body resolves by the conclusion — no dropped setups, no orphaned promises. 2. **Premise→conclusion arc.** The premise set in the ingress (Step 3) is the same premise the conclusion grips and twists forward. If the draft drifted to a different premise, realign the conclusion or the ingress — do not leave two premises. 3. **Leader-takeaway.** The edition lands ONE clear takeaway + ONE concrete action. Cut references hard; hands-on credibility beats a citation-pile. If the reader cannot state the takeaway in one sentence, tighten until they can. 4. **AI-slop removal (forbidden phrases — strip on sight).** These are the Seres ban-list; they read as machine-written and must not appear: - "her må jeg være ærlig" / "for å være ærlig" - "ikke bare X, men Y" - gratuitous three-item listing (rule-of-three as a tic) - "i en stadig mer kompleks verden" - tacked-on summary sentences that restate what was just said 5. **General, not org-/person-specific.** No personal agency anecdotes; present opportunities, not provocations. At most ONE structural anchoring reference per text — never repeated criticism of a named person. 6. **Formatting dose (minimal).** Bold = at most one point per section. Short lists (2–4 items) only where the text already enumerates — never turn load-bearing reasoning into bullets. Tables sparingly. *"No article should look like a PowerPoint printout."* 7. **Close gaps by tightening, not expanding.** Swap weaker passages for sharper ones; keep the length flat. After the pass, set `currentPhase: "quality"` in `edition-state.json` and append a "quality pass complete → next: fact-check sweep" pointer to the HANDOVER §6. ``` Consistency + quality pass complete. - Threads resolved: <yes/flags> - Premise→conclusion arc: <intact/realigned> - Leader-takeaway: <one line> - AI-slop phrases removed: <N> - Formatting dose: <within bounds/trimmed> - Length delta vs. draft: <flat/±N words> (target: flat) Next: Step 5 — Fact-check sweep (guilty-until-disproven, BEFORE lock). ``` ## Step 5: Fact-check sweep — guilty-until-disproven (BEFORE lock) Every factual claim in the consistency-passed draft (Step 4) is now treated as **guilty until proven** — it is not true until a primary or credible source confirms it. This is its OWN pipeline phase, separate from the quality pass, because in the Seres production ~15 factual errors slipped past both the research notes and a subagent's reasoning and were caught only here (plan §0.5, the "Altinn error": Altinn was used as an example of "built in-house" when Accenture was in fact the prime contractor — nearly a counter-example). Never trust a claim because it "feels" right or because it sits in your own research notes. > **This sweep runs BEFORE lock.** No edition is locked (Step 8) with an > unresolved 🔴 claim. Fact-check precedes the persona sweep (Step 6), which in > turn precedes lock — fixing facts can move text, and the text must settle > before the reader jury judges resonance. **Procedure:** 1. **Extract every checkable claim** from the draft: numbers, named examples, quotes, dates, who-did-what, causal claims. Skip opinions and predictions — they are not claims to verify. Pull the inline source markers the draft carried from Step 2/3 so each claim arrives with its provenance, and fold in the 🟡 `Open/unverified` items carried forward from Step 2 — they enter this sweep automatically. 2. **Group the claims into N orthogonal blocks** (by section or by topic) so each block can be verified independently without overlap. 3. **Fan out in parallel — issue all N `fact-checker` calls in a SINGLE message** (multiple `Task` tool-uses in one turn, `subagent_type: fact-checker`) so they run concurrently, from THIS command layer in the foreground (principle 4, plan §3). Each call gets one claim-block and returns the agent's standard verification log + risk-sort (🔴/🟡/🟢) + gate decision (PASS/REWORK/BLOCK). Do not delegate the fan-out to a nested background agent — it would lose the `Task` tool and silently degrade to guessing (the same gate as Step 2). 4. **Detect degradation (gate).** Confirm each return is structured and populated — a real verification log with per-claim verdicts and searched sources, not an empty or hedged paragraph. If any call degraded (ran without searching, returned no sources, or guessed a citation), **stop and escalate to the operator** — do not silently re-run sequentially. `[GATE]` 5. **Merge into one verification log + risk sort.** Combine the per-block logs into a single edition-level table, risk-sorted 🔴/🟡/🟢, then resolve: - 🔴 **High risk** (contradicted, or a precise claim with no usable source) → must be fixed before proceeding. Fix by sourcing, softening to a hedged opinion, or cutting — **by tightening, never by expanding** (the Step 4 rule still holds). Re-run the relevant `fact-checker` call on the fix. - 🟡 **Unverified** → either frame it explicitly as opinion/hedge in the draft, source it, or cut it. A 🟡 left asserted as bare fact is a REWORK. - 🟢 **Verified** → keep; record its source in the verification log. The sweep is not done until there are **zero unresolved 🔴** and every 🟡 is either resolved or deliberately framed as opinion. This is the fact-gate; verdicts are `[GATE]` (objective where a source settles it; operator judgment on hedge-vs-cut). Flag any time-sensitive figure (e.g. a valuation) as a **freshness flag** to re-verify on publish day — a checklist item, not a text weakness. 6. **Persist + checkpoint state.** Write the merged verification log into the edition-HANDOVER §4 (immutable rules + fact-check log — the durable record of what was checked), set `currentPhase: "factcheck"` in `edition-state.json`, and append a "fact-check complete → next: persona sweep (BEFORE lock)" pointer to the HANDOVER §6. ``` Fact-check sweep complete. - Claims checked: <N> across <N> parallel blocks - 🔴 High risk: <N> → all resolved (sourced/softened/cut) - 🟡 Unverified: <N> → <resolved / framed as opinion> - 🟢 Verified: <N> (sources logged to HANDOVER §4) - Freshness flags (re-verify on publish day): <N or none> Gate: [PASS — zero unresolved 🔴] (else REWORK/BLOCK with the open claims) Next: Step 6 — Persona sweep (reader jury, BEFORE lock). ``` ## Step 6: Persona sweep — reader jury, BEFORE lock The fact-checked draft now faces the **reader jury**: the personas selected in Step 1 read it read-only and judge whether it *lands* — not whether it is correct (that was Step 5), not whether it is original. This is the **single most important ordering rule in the whole pipeline.** In the Seres production this sweep was originally run *after* the texts were locked (Step 8), which forced reopening locked texts — the biggest single process error of the series (plan §0.4). **It runs here, FØR lås / before lock, without exception.** > **Order assertion (enforced).** This persona sweep (Step 6) precedes lock > (Step 8). The pipeline may NOT proceed to Step 8 until the primær persona > returns a clean JA from this sweep. A wiring that locked first and reviewed > after would reproduce the exact Seres failure — do not do it. `[GATE]` **Procedure:** 1. **Load the active personas** chosen in Step 1, with exactly one marked **primær**. Each persona's five fields (rolle, avkobler, overbeviser, ekspertise, sjargong) come from `config/personas.local.md` (or the template). 2. **Fan out one `persona-reviewer` call per persona, in parallel** — issue them in a SINGLE message (multiple `Task` tool-uses, `subagent_type: persona-reviewer`), from THIS command layer in the foreground (principle 4). Pass each call its persona name and **`mode: resonans`** (the before-lock mode — all six axes, ≤5 flags as direction). This is NOT conversion mode, which is the post-lock hook-gate in Step 9. One persona per run — never mix two. 3. **Collect verdicts and gate.** Each call returns per-axis flags (LØST/DELVIS/IKKE), ≤5 direction-only flags, a per-persona verdict (JA/NEI), and a gate decision. Aggregate per the agent's rule: - **primær JA** + no real (non-ceiling) sekundær IKKE → PASS, ready to lock. - **primær NEI**, or a fixable DELVIS/IKKE → REWORK. - **primær NEI on Krok or Leder-takeaway** → BLOCK (the reader never starts, or leaves with nothing to do) — must be reworked before lock. A *sekundær* NEI from a role mismatch or expertise ceiling is a SIGNAL the gate works — accept it; do not distort the text to chase it (plan §0.5, "primær trumfer"). The jury returns **direction only** — the editor (this session) holds the pen; never paste a persona's rewritten copy. `[GATE]` 4. **Convergence loop.** If the gate is REWORK/BLOCK, fold the flags into the draft **by tightening** (the Step 4 rule holds — close the gap, hold the length flat), then **re-run the same `persona-reviewer` calls** against the revised draft. Each round re-judges every prior flag as LØST / DELVIS / IKKE. Loop until the primær returns a clean JA (in Seres this took 2 rounds). Re-run only the personas whose verdicts are still open. 5. **Persist + checkpoint state.** Record the final per-persona verdicts and the resolved flags in the edition-HANDOVER §5 (method / persona calibration), set `currentPhase: "persona"` in `edition-state.json`, and append a "persona sweep PASS (primær JA) → next: lock/delivery" pointer to the HANDOVER §6. ``` Persona sweep complete (BEFORE lock). - Personas run: <N> (primær: <name>) - Convergence rounds: <N> - primær verdict: JA (else: still NEI — loop open, NOT ready to lock) - Accepted sekundær ceiling-NOs (signal, not failure): <N or none> Gate: [PASS — primær JA, ready to lock] (else REWORK/BLOCK) Next: Step 7 — Annotation (optional), then Step 8 — LOCK → delivery. ``` --- ## Steps 7–10 (added in a subsequent build session) Optional annotation (render annotatable review HTML), lock/delivery (POST.html "all in one place"), the post-lock hook/conversion gate (`persona-reviewer` in konverter-modus), and scheduling are implemented in plan step S10. Each will append its phase here, reading the phase contract from `config/edition-state.template.json` and (once extracted in S12) the long-form quality rules from `references/longform-quality-rules.md`. --- ## Reference Files - `${CLAUDE_PLUGIN_ROOT}/config/edition-state.template.json` — edition-state schema (11 phases) - `${CLAUDE_PLUGIN_ROOT}/config/personas.template.md` — reusable reader personas + "primær trumfer" rule - `${CLAUDE_PLUGIN_ROOT}/agents/fact-checker.md` — Step 5 fact-check agent (risk-sorted, guilty-until-disproven) - `${CLAUDE_PLUGIN_ROOT}/agents/persona-reviewer.md` — Step 6/9 reader jury (resonance + conversion modes) - `${CLAUDE_PLUGIN_ROOT}/commands/react.md` — multi-source synthesis discipline (reused in Step 2) - `${CLAUDE_PLUGIN_ROOT}/assets/voice-samples/authentic-voice-samples.md` — voice matching - `${CLAUDE_PLUGIN_ROOT}/render/build-linkedin.mjs` — POST.html delivery (Step 8) - `${CLAUDE_PLUGIN_ROOT}/render/build-html.mjs` — annotatable review renderer (Step 7)