agentic-content-pipeline

An agentic content pipeline: one topic in, up to four medium renditions out — blog post, LinkedIn post, email, Bluesky thread. A 12-stage orchestration runs two review loops — 12 reviewers grading the outline, 15 grading the draft, roster swapped per medium, drawn from a 28-agent roster — that fan out, synthesize, and edit findings into apply-ready fixes; an assertion audit (extract every checkable claim, web-check named entities and attributions) that runs before the review loops and again on the assembled metadata; a deterministic AI-tell tone gate (aiScore < 15 on a 0-100 scale, plus a hard permaban phrase list); and a generation-side learner, per medium, that feeds confirmed fixes back into future drafts. Ships as two companion skills — content-pipeline (the orchestrator) and human-tone (the tone gate) — neither runs without the other.

Quick start

Run inside your own content repo:

npx agentic-content-pipeline setup --harness claude   # or --harness codex

This:

• installs content-pipeline + human-tone into .claude/skills/
• registers the agents natively for your harness (skip --harness and they still dispatch from reference files, fully functional)
• interviews you into content-pipeline.config.json, validating your hero-image provider along the way
• seeds the working ledgers and voice/editorial-rules templates
• installs the skills' own runtime deps

Safe to re-run — every write is no-clobber. setup --help lists every config flag, for a non-interactive run.

Then invoke the content-pipeline skill — "write a new post" — the topic-approval gate is the only human checkpoint.

Alternative install — npx skills add buildaloud/agentic-content-pipeline places both skills under .claude/skills/ directly. Then run npx tsx skills/content-pipeline/scripts/setup.ts from inside the installed skill for the same config/seed/deps steps (skills-only; agents still dispatch from reference files).

Mediums

One topic can ship as up to four renditions, gated by config.mediums (blog on by default; linkedin / email / bluesky opt-in). The content unit is a folder, not a file:

yyyy-mm-dd-postname/
├── blogpost.md   # full frontmatter — title, description, pubDate, tags, heroImage...
├── linkedin.md   # frontmatter: postDate, status: draft, hook
├── email.md      # frontmatter: subject (≤50 chars), preheader (≤90 chars), status: draft
└── bluesky.md    # frontmatter: postDate, status: draft, thread; ≤300 chars/segment

Research and brief run once per topic, shared across mediums. Each enabled medium then gets its own outline, outline-review loop, draft, and draft-review loop, built from that medium's own formula doc (references/post-formulas.md, linkedin-formulas.md, email-formulas.md, bluesky-formulas.md) — and swaps in its own fit-reviewer gate in place of seo-reviewer (linkedin-reviewer, email-reviewer, bluesky-reviewer). The tone gate and core prose reviewers grade every medium unchanged.

Publishing, honestly. blog publishes the way it always has — commit, and your site's own pubDate filter plus its daily rebuild does the drip. linkedin / email / bluesky have no platform API in v1: the status: draft file is the deliverable, and publishing it is yours.

How the review loops work

Two fixpoint loops, one per grain (outline, then draft). Each round is two-fold: a mechanical tone lint first, then a 12-15-reviewer fan-out (the army never burns dispatches on tells a regex already caught). synthesis dedups and ranks the findings into apply-ready edits (quote → replacement), one editor agent applies them, and the loop re-reviews — up to 5 rounds.

The tone gate is deterministic code, not a model's opinion: it runs before every round and once more after the loop exits, so the loop's own last edit pass is never the one thing left unmeasured.

What you provide

• a voice file and an editorial-rules file (both optional — fill-in-the-blank templates ship if you don't have one; the skill can interview you into one)
• bannedTerms — anything that must never ship (internal tool names, codenames)
• source-material directories to scan for topic ideas (optional — skip it and the skill just asks you directly)
• a hero-image provider + key (codex, openai-api, screenshot, or none)
• a tone corpus (optional — recalibrates the AI-tell baseline to your own writing) and a stats file (optional — per-post performance data; shape at skills/content-pipeline/references/stats-schema.md)

tone subcommand

npx agentic-content-pipeline tone <file> scores one file's aiScore/permaban-phrase hits against the ship gate, outside the full pipeline.

Cost

This is not cheap to run. At defaults — two review loops, up to 5 rounds each, roughly 12-15 reviewers per round — expect on the order of 100-160 agent dispatches and several million tokens per post. Run one post and check your usage dashboard before queueing many. The fan-out also bursts up to 15 agents at once per round; mind your provider's rate limits. The payoff is that a lot of what would otherwise be manual editing passes happens unattended.

Development

Contributing to this package itself (not consuming it): npm install && npm test runs the full suite — 215 tests via vitest.

Provenance

Extracted and generalized from a production build-in-public content pipeline (July 2026); this package is a point-in-time fork, not a live mirror.

License

MIT — see LICENSE.