Shipping a Branded Essay Series

Takes a series brief or spec and produces the full set of brand-templated, hard-rule-compliant, publishable HTML essays — with lane enforcement, reuse detection, protected-material gating, and adversarial verification before delivery.

TL;DR

Read ground truth first. Draft the content as a reviewable manuscript. Template into standalone HTML from ONE reference file via parallel subagents. Scan for forbidden terms before delivery. Gate anything that touches a protected lane or unconfirmed fact — never guess it clean.

The moat is not the writing. It's the enforcement: employer never surfaces, no fabricated numbers, memoir material stays cut, reused assets get renumbered instead of rewritten.

---

Reference Map

The body covers a normal run. Pull these only for the depth they name.

You need	Read
Exact template CSS/structure per brand (Load-Bearing + FlashFusion)	references/brand-template-contract.md
The full hard-rule / lane-separation table + provenance rules	references/hard-rules-and-lanes.md
Voice rules, openers/closers, banned words	references/voice-spec.md
Subagent templating pattern + the verification scan in detail	references/templating-and-verification.md
Which brand, reference-essay path, project store, act names	assets/config.md
The acceptance gate before shipping	assets/success-criteria.md
Runnable pre-delivery verifier	scripts/verify-series.sh

---

Decision Tree

Do I have ground truth (project instructions, voice spec, existing essays)?
  ├─ No  → project_read/project_search FIRST. Read the voice files and one
  │        shipped essay for the exact template. Do not draft from the brief alone.
  └─ Yes → continue.

Does the brief conflict with ground truth?
  └─ Flag it explicitly. Do NOT average. Ground truth wins; say so.

Which brand is this series?
  └─ Read assets/config.md. If active brand unset → ask (config pattern).
     Load the matching block in references/brand-template-contract.md.

For each planned post — material status?
  ├─ Fully sourced, no protected lane        → draft it
  ├─ Reuses an existing shipped essay         → RENUMBER (badge edit), do NOT rewrite
  ├─ Touches memoir / clinical / [PERSON]      → CUT-FORM only, or gate for Kyle's call
  ├─ Touches the day-job / employer           → GATE. No anonymization salvages it.
  └─ Depends on an unconfirmed fact/date      → GATE. Draft around needing it.

Output stage?
  ├─ First pass  → content-first MANUSCRIPT (one .md, all posts, footers).
  │                Let Kyle correct the voice before HTML locks.
  └─ On "ship"   → template each into HTML from ONE reference file, in parallel.

Before delivery → run scripts/verify-series.sh. Resolve every hit's context
(substring false positives are common). Only then SendUserFile + project_write.

---

Usage Examples

Example 1 — Spec first

User: "Synthesize the series spec for the first published blog series (14 posts)."
→ Read ground truth + voice files. Propose 3 spines, pick one, justify.
→ Map each post to a real inventory item or flag [NEEDS MATERIAL].
→ Detect reusable existing essays; flag protected-lane posts as GATED.
→ Deliver the spec with CLAIMS/GAPS footers. Do NOT draft essays yet.

Example 2 — Build the rest

User: "Build out all the rest."
→ Draft every buildable post in ONE manuscript .md (content-first, reviewable).
→ Reused posts = renumber note, not a rewrite. Gated posts = held with reason.
→ Per-post CLAIMS/GAPS. Flag any tonal-sameness risk across similar posts.

Example 3 — Ship

User: "ship"
→ Template each drafted post into standalone HTML from ONE reference file, in
   parallel (one subagent per post, character-for-character prose).
→ Renumber reused essays via badge edit. Run scripts/verify-series.sh.
→ Resolve scan hits → SendUserFile + project_write.

---

The 7-Step Pipeline

1. Ground truth before drafting. project_read/project_search the project instructions, voice files, and one shipped essay. Flag brief-vs-truth conflicts — don't reconcile silently. Material can be split across project stores; a "missing" file may live in the other store.

2. Enforce hard rules / lane separation. These are FAIL conditions — one violation kills the post. Full table + provenance rules: references/hard-rules-and-lanes.md.

3. Detect reuse — renumber, don't rewrite. If an existing shipped essay already covers a slot, change only the badge + footer meta. Project docs are full-replace only; reproduce from a fresh read. See references/templating-and-verification.md.

4. Gate protected / unconfirmed material. Do not guess a post clean. Hold it, state the exact confirmation needed, and note the count impact. Draft the rest around not needing it.

5. Content-first manuscript, then template. Draft all buildable posts as ONE markdown manuscript so Kyle can correct voice before HTML locks. On "ship", template in parallel — one subagent per post. Pattern + brand template: references/templating-and-verification.md, references/brand-template-contract.md. Voice: references/voice-spec.md.

6. Adversarial verification before delivery. Run scripts/verify-series.sh over the outputs. RESOLVE each hit's context — substring false positives are common (spect→"inspection"; percentile inside a "no percentile" line). Confirm template integrity: title, badge number, no editorial note leaked.

7. Deliver with footers. SendUserFile the set (attach for many files) and project_write each under claude/. End every deliverable with CLAIMS / GAPS. Name the real publish blocker (platform / live URL), not just "content done".

---

Gotchas

1. Material is split across project stores. A file the brief calls "missing" may live in the Abstrekkt project while the search ran against ChaosCollective/Notion. Read the actual store before declaring a gap.

2. Memoir material must be CUT, not anonymized. The cognitive percentiles and "Construction, Not Recovery" sit inside a clinical record. A "bare fact, no origin" version still fingerprints. Default: cut. Flag as Kyle's call — it touches an active custody matter; don't auto-execute either way.

3. Reuse = renumber, never rewrite. Regenerating a shipped essay to change a badge is a rebuild where a two-string edit was correct.

4. The forbidden-term scan produces false positives. Always print context and confirm before treating a hit as a violation — and never treat a clean scan as proof of voice quality; it catches terms, not tone.

5. Editorial notes leak into published copy. "Cut-form note", [NEEDS MATERIAL], manuscript scaffolding — strip them when templating. Tell subagents explicitly to omit them.

6. Content-first is not optional for a first-ever public series. Ten HTML files locked from unreviewed voice is a gamble. Manuscript first.

7. Watch tonal sameness across similar posts. Posts sharing a rhetorical move (name a number/credential, then undercut it) read fine alone and samey in a run. Flag which cluster to differentiate.

8. Terse command mid-series = execute. "ship" / "continue" / "build the rest" after a spec exists means run the pipeline with documented defaults, not ask more questions. State the defaults, proceed.

9. Overlaps producing-ai-newsletter — Load-Bearing IS the newsletter. This skill owns the multi-post series → shipped HTML pipeline. A single regular newsletter issue belongs to producing-ai-newsletter. Chain, don't collide.

---

Self-Verification Checklist

0. Ground truth (project instructions + voice files + a shipped essay) read before drafting.
1. Every brief-vs-ground-truth conflict flagged, not averaged.
2. Zero hard-rule violations (per references/hard-rules-and-lanes.md).
3. Reused essays renumbered (badge only), not rewritten.
4. Gated posts held with an explicit confirmation ask + count impact — none guessed clean.
5. Manuscript delivered before HTML (unless Kyle said ship outright).
6. HTML templated from ONE reference file; prose reproduced character-for-character.
7. scripts/verify-series.sh run; every hit's context resolved; no editorial note leaked.
8. Template integrity: correct title, badge number, footer meta per file.
9. [SOURCED] / [GENERATED] provenance; CLAIMS / GAPS footers on the deliverable.
10. Real publish blocker named (platform / live URL), not just "content done".

---

Trigger Evals

Positive (must activate):

• "synthesize the series spec for the 14-post series" → activates
• "build out all the rest of the posts" → activates
• "ship" / "template these into standalone pages" after a spec exists → activates

Negative (must stay silent):

• "write me one LinkedIn post about AI adoption" → generating-social-posts, not this
• "draft this week's Load-Bearing newsletter issue" → producing-ai-newsletter, not this

---

CLAIMS

• [SOURCED] The Load-Bearing template and hard rules — shipped essays lb-essay-04, lb-essay-07, and Abstrekkt project instructions (Jul 7 2026).
• [SOURCED] FlashFusion color tokens (void-dark #0B0B14, violet #A855F7) — project design-token artifacts.
• [GENERATED] Parallel-subagent templating from one reference file is the lowest-drift way to produce many consistent pages — held across 10 files with no drift this session.

COUNTEREXAMPLE

• For 1–2 fully-sourced posts with no protected-lane exposure, the manuscript step is overhead — draft straight to HTML.
• A clean forbidden-term scan is necessary, not sufficient — it misses a memoir leak phrased in novel words, or a flat, off-voice closer.

CONTRADICTIONS

! "Ship fast on a terse command" vs. "content-first manuscript" — resolved by the command word: "ship" templates directly; "build the rest" delivers the manuscript first. ! "Reuse, don't rewrite" vs. a clean arc-numbered file set — resolved by renumbering the reused essay into a correctly-numbered copy with the badge changed and content otherwise identical. ! "Cut the memoir material" (safe default) vs. its differentiation value — defaults to cut but must surface the tradeoff for Kyle's decision; it is his risk to set, not the skill's.