Build Sprint, backlog for a one-off run
A prioritised, self-contained backlog for an autonomous build session. Fire it as a one-off Routine (see the prompt at the bottom) whenever there is work to do and nobody watching.
Ground rules for whoever runs this:
- Commit and push after each item, not at the end. A run that dies halfway must leave the finished items live, not lose everything.
- Never break the build.
python3 tools/predictions.py check && python3 tools/build_site.pyplus the link checker must pass before every push. - Verify visually. Chromium and Playwright are installed. Screenshot what you change, in at least two themes, and actually look at it. A green build is not a design review.
- Work top-down. The list is in priority order. Finishing three items well beats starting six.
- Do not touch
reports/daily/orreports/weekly/for dates a scheduled run owns. Backfilling an explicitly listed missing issue is fine; competing with the 6 AM daily is not. - Report what you did, what you skipped, and why.
1. Decks retired (done)
01-top-20-deck.html and 05-finance-deck.html were the last things on the site that looked like they came from somewhere else, and the plan here was to fold them into the design system. They are gone instead.
The reason: Top 20 existed twice, once as a list and once as slides, with the same content in both. Two presentations of one thing means two places to keep right, and the deck was the one nobody was reading. The list stayed.
2. Backfill the first weekly issue
No weekly exists yet; the section is empty and the new cadence is invisible.
Done when: reports/weekly/2026-08-09.md exists, covering 3–9 August, following §2 of RUNBOOK.md in full. The Week, What Changed, Deep Read, Finance, Watch & Listen, Editorial, Winners & Losers, Threads, Ledger, Next Week, Sources.
Byline: ISO week number mod 3 (week 32 → index 2 → Nour Haddad).
Source it from the existing daily briefs plus fresh research. Log at least three new predictions. A weekly with no new falsifiable calls has failed at its main job.
3. Charts across the rest of Vol. 1
Only the finance report has charts. The executive summary, Top 20 and full report are still prose around numbers that want to be seen.
Done when: each has 2–4 charts where they genuinely help. Candidates: the enterprise adoption gap (80% embedded vs 31% deployed vs 23% reporting ROI, a bar), Anthropic's revenue ramp $1B → $9B → $47B (a line), custom vs merchant growth (a split), AI-cited share of US job cuts 0.6% → 4.5% → 13% (a line).
Do not decorate. A chart that repeats a sentence without adding shape is noise. See tools/charts.py for the four types and the fenced-JSON syntax.
4. Site search
Long-form archive with no way to find anything.
Done when: a client-side search works over every report, build a JSON index at build time (title, date, kind, headings, body text), ship a small script, no dependencies, no external service. Keyboard shortcut / to focus. Must degrade gracefully with JS off.
5. Entity pages
The roadmap's Phase 2, and the thing that makes accumulation visible instead of theoretical.
Done when: /entities/<slug>.html exists for the ~15 most-mentioned companies (Nvidia, OpenAI, Anthropic, Google, Microsoft, Meta, CoreWeave, Palantir, TSMC, Moonshot, DeepSeek…), each listing every mention across every issue in reverse-chronological order with links and dates, plus any open predictions naming that entity.
Detect mentions by scanning the Markdown for a curated alias list in data/entities.json, no NLP, no dependencies. Link entity names in article text to their page on first mention.
6. If there is still budget
- Reading position memory: return to where you left a long report.
- An RSS/Atom feed at
/feed.xml. - Print stylesheet worth the name; these get printed.
prefers-reduced-motionhonoured on the progress bar and tile hovers.- A sixth theme. One dict entry in
tools/theme.py. Try a high-contrast accessibility theme, or something with real personality, the palette is not the limit.
The one-off Routine prompt
Create a routine at claude.ai/code/routines, repository builtwithclaude4313/the-ai-read, environment with Network access: Full, no connectors, trigger Schedule → one-off (or just Run now):
Run a BUILD SPRINT on The AI Read.
Repository: builtwithclaude4313/the-ai-read (work on main)
1. Read docs/SPRINT.md and work the backlog top-down. It carries the ground rules,
the priority order and the acceptance criteria for each item.
Also read docs/RUNBOOK.md (procedure), docs/METHODOLOGY.md (the filter) and
docs/PERSONAS.md (the bylines) before writing any content.
2. Establish today's real date with `date -u +%F`. Never assume it from training data.
3. Commit and push after EACH completed item, not at the end. Before every push:
python3 tools/predictions.py check && python3 tools/predictions.py render
python3 tools/build_site.py
and confirm no internal links are broken. If CI goes red, fix it in the same run.
4. Verify visually. Chromium and Playwright are installed at /opt/pw-browsers.
Screenshot what you change in at least two themes and actually look at it.
A green build is not a design review.
5. Do not create or edit reports/daily/ entries, a scheduled run owns those.
Backfilling reports/weekly/2026-08-09.md is explicitly in scope.
6. Work until the backlog is done or you run out of budget. Then reply with what you
finished, what you skipped and why, and links to the commits.
Standing rules: never fabricate a URL; never average contested figures, show the
disagreement; tier every number [A]/[B]/[C]/[?]; dated issues are immutable, corrections
append; personas change the argument, never the facts.