Weekly Signal Diff — AI

Problem

A wall of AI news does not tell the user what structurally changed in the AI ecosystem. Most weekly roundups over-index on model launches and benchmark screenshots, underweight economics, dependency shifts, and regulation, and ignore what the user actually cares about. This skill turns a noisy week in AI into a small set of structural changes.

Scope: frontier labs, open models, dev tooling and agents, cloud AI platforms, data and model infrastructure, enterprise AI buyers, creative media, robotics and embodied AI, and AI regulation.

When to Use

• Weekly AI market review or Sunday/Friday ritual
• "Run my weekly AI signal diff"
• "Run my weekly signal diff for AI"
• "What changed in AI this week that matters to me?"
• "Track the AI market and tell me the structural shifts"
• "Turn this pile of AI news into a decision-grade diff"
• "What moved in frontier labs / open models / AI tooling this week"
• Ongoing automation for a weekly AI digest

Required Context

Gather as much as the environment allows:

• the user's active projects, bets, and recurring interests
• prior weekly digests or summaries the user has shared
• the desired freshness window (default: last 7 days)
• any preferred outlets, banned sources, or explicit watchlist entities

If the user has not provided categories or companies, read references/starter-universe.md and use it as a bootstrap layer only.

If live web access is available and the user wants current coverage, read references/live-search-upgrade.md and use the strongest search mode the environment supports.

Process

1. Establish the frame.
   • Confirm the topic space, freshness window, and whether the goal is personal awareness, operator strategy, investor tracking, or content prep.
   • If the user says nothing, default to a 7-day operator-style review.
2. Build the watchlist.
   • Start from the suggested 10-category / 30-company starter universe if the user has not defined a watchlist.
   • Treat the starter list as a scaffold, not a contract.
   • Re-rank or replace items based on what the user has shared:
     • promote companies, categories, or themes the user mentions often
     • demote low-signal items
     • add personal-priority entities even if they are outside the starter set
   • Preserve some baseline discovery. Personalization should shape the scan, not collapse it into only known favorites.
3. Gather the week's evidence.
   • Prefer fresh, source-backed information with links or citations.
   • If live search is available, perform a broad sweep first, then targeted follow-ups on the top candidate shifts.
   • If live search is not available, work from the user's provided sources and say that the diff is source-bounded.
   • Ignore pure announcement theater unless it changes economics, distribution, regulation, dependency, geography, or buyer behavior.
4. Ask the structural questions on every candidate signal.
   • What constraint shifted?
   • Who gained or lost leverage?
   • What got cheaper, harder, faster, or more defensible?
   • What dependency got exposed?
   • What business model or pricing assumption weakened?
   • What changed in regulation, geography, or distribution?
   • Why does this matter for the user's actual projects, workflows, or market view?
5. Score before writing.
   • Keep only the few signals that represent real change.
   • A good weekly diff usually has 3-7 structural shifts.
   • Merge duplicates, drop weak stories, and explicitly label speculation as speculation.
   • Use references/structural-scorecard.md when candidate signals are noisy or the scan risks becoming a digest.

6. Produce the weekly diff in chat.

   • Output the full diff directly in the conversation. Do NOT create a file or document. The diff is the deliverable.

   Use this default structure:

   • Coverage note — what was scanned, how it was personalized, and the date window
   • Structural shifts — 3-7 items, each with:
     • what changed
     • why it matters in general
     • why it matters to this user
     • supporting evidence or citations
   • What changed from last week — new, rising, fading, or resolved themes
   • Watch next — entities, constraints, or questions to monitor
   • Actions — optional follow-ups, only if the evidence supports them

Output

When this skill works correctly, the user gets:

• a concise weekly structural diff delivered directly in chat (no file created)
• a clear explanation of why the shifts matter to them specifically
• citations or source links when live search is available

Guardrails

• The goal is diff, not digest.
• Do not force all 30 suggested companies into the final output. They are there to prevent blank-page syndrome, not to create fake coverage.
• Do not mistake product launches, benchmark screenshots, or funding headlines for structural change unless they move a real constraint.
• Keep general market analysis separate from personalized implications.
• If evidence is thin, say the week was thin.
• If the environment lacks live search, be explicit about the freshness limitation.
• If the user's interests are unclear, use the starter universe and explain that it is a bootstrap pass.
• Never create a file or document as the output. Chat only.

Notes for Other Clients

• This skill is portable across Codex, Claude Code, Cursor, and similar clients because the core behavior is procedural.
• For scheduled runs, pair the skill with the user's automation system and keep the same structure every week so diffs stay comparable.
• If OpenRouter is available, prefer a Perplexity Sonar web-search model for the retrieval pass, then use the local AI client or model to do the actual synthesis if that split is more ergonomic.

References

• references/starter-universe.md — bootstrap categories and entities.
• references/live-search-upgrade.md — optional live retrieval upgrade.
• references/structural-scorecard.md — candidate scoring table for separating structural shifts from ordinary news.