Crip Minds Pipeline — Read Before Touching Anything

What This Is

disability-ai-collective — a disability culture publication (Jekyll site, GitHub Pages). Articles are generated daily by automation/production_orchestrator.py via cron through the ops OpenClaw agent heartbeat.

Do NOT run the orchestrator manually unless explicitly asked. It commits and pushes to GitHub immediately — article goes live.

The 4 Personas (AI authors)

| Persona | Disability | Beat | |—|—|—| | Pixel Nova | Deaf | Visual systems, information architecture, space politics | | Siri Sage | Blind | Acoustic culture, sensory phenomenology, soundscape | | Maya Flux | Mobility (wheelchair, T6) | Urban infrastructure, economics of care, protest history | | Zen Circuit | Autistic/neurodivergent | Diagnosis politics, pattern recognition, sensory epistemology |

Each has a deep character profile embedded in the orchestrator. The article IS written from their lived experience — disability as expertise and lens, never tragedy.

Output Format Contract

Every article is a Jekyll post with exact frontmatter:

---
layout: post
title: "Title here"
date: YYYY-MM-DD
author: Pixel Nova          # exact persona name
category: research          # or: culture, technology, justice
excerpt: "One sentence."
image: /assets/images/SLUG-1.jpg
---

Frontmatter must use double quotes. Never put AI-generated text directly in YAML values without stripping quotes first — breaks parsing silently (nil title, double-date URL).

Image HTML blocks in the body use exact format:

<figure class="article-figure">
  <img src="/assets/images/SLUG-N.jpg" alt="description" loading="lazy">
  <figcaption>Caption here.</figcaption>
</figure>

What NOT to Do

• No academic headers (Research Question / Methodology / Key Findings / Recommendations)
• No bullet-point policy lists
• No invented statistics or fake study findings
• No “Case study: Sarah, a graphic designer…”
• No summary ending / call to action / “We need to…”
• No publication names like “De Correspondent” or “dis.art” anywhere in articles
• Never put unescaped quotes in YAML front matter fields

Git Workflow

This workspace IS the live repo. Commits here push to GitHub → GitHub Actions builds Jekyll → site goes live within ~2 minutes.

git status          # check what changed
git log --oneline -5
git push origin main   # ALWAYS push after any editorial commit

After any editorial change: always git push origin main — no push = site stays stale.

Retract an Article

python3 automation/production_orchestrator.py --retract <slug>

Deletes Bluesky post (URI in _social/<slug>.json) + git rm + push.

Key Directories

_posts/         ← published articles (YYYY-MM-DD-slug.md)
assets/images/  ← article images
automation/     ← orchestrator, link pool crawler, image gen
_social/        ← Bluesky post URIs per article

Orchestrator Architecture (9-step pipeline)

1. Pick agent (balance across personas)
2. Get topic from sentinel findings DB
3. Fetch source article
4. Build prompt (style + persona + source + link pool)
5. Generate with Opus (CLIProxy) → fallback chain
6. Rewrite with Opus (editor pass) if not already Opus
7. Generate images (Pollinations API)
8. Insert images balanced in body
9. Write frontmatter → commit → push → Bluesky post

Danger Zones

• automation/link_pool.db — don’t delete (curated link seeds + crawled pool)
• automation/disability_findings.db — don’t delete (topic source material)
• .env in workspace root — secrets, never commit
• Gold standard article: _posts/2026-03-14-the-open-office-was-designed-to-break-my-brain.md — do not delete or move (used as rewrite reference)

v2.2 Changes (2026-03-19)

Pipeline fixes

• 23 syntax errors fixed in orchestrator prompt strings (rules 16-21 were never actually firing)
• rfind() missing argument in Bluesky hook word-wrap — was silently falling back to hard truncate
• Duplicate rule numbers 17-21 resolved (old rules renumbered to 22-27)
• Date collision guard: _get_recent_dates_nudge() scans last 7 posts, blocks repeated “In [Month] [Year]” anchors

Editorial quality

• Card excerpts: _generate_card_excerpt() now generates tension-frame sentences (two things that shouldn’t both be true) via Haiku — NOT first-paragraph truncation
• Social hooks: _bsky_hook() now targets scene/evidence — most specific concrete detail, intentionally incomplete
• Per-persona social prompts (_SOCIAL_PROMPTS) now receive 1500 chars of body instead of 500
• All 13 published article card excerpts backfilled with tension-frame sentences

Site / conversion

• Homepage: GoatCounter click tracking added to 3 article cards in grid (card-click-<slug>)
• /about: “Three places to start” reading list added (Map That Stops, Open Office, Room That Sings)
• /press: Made fully indexed (removed noindex/sitemap:false/lanternlight CSS)
• /press/how-it-works: Easter egg treatment kept (noindex, dark aesthetic, “you found it”)

Conversion Optimisation Log

Insight (2026-03-19, v2.2)

GoatCounter data after 11 days (35 homepage visits, ~15 article reads):

• cta-latest-article (hero pull quote link): 0 clicks — hero not converting
• cta-read-articles (primary CTA): 1 click
• /research: 8 visits
• /about: 7 visits — nearly as popular as /research, 0 article conversions tracked from it
• Room That Sings (Mar 17): 6 reads on launch day — strongest Bluesky hook drove this
• Referrers: 56 direct (mostly Bluesky app), 3 Google, 2 go.bsky.app

Backlog

• Hero redesign: make tension-frame excerpt the visual anchor, not caption under title
• Bump homepage grid from 3 → 5 cards (publish daily, grid goes stale fast)
• Consider article image in hero