knowledge-delta-skills

Not knowledge for beginners. Missing expertise for already-capable AI.

Scope: Meta and Google paid media (no TikTok/Microsoft), affiliate trackers (Keitaro deep, Binom partial), plus frontend/engineering/security skills. The media-buying set has a clean lane (buy mechanics, feeds, measurement) and an opt-in grey lane (*-grey-ops, senior-buyer-ops) that teaches account-survival tactics for aggressive verticals — install the grey lane only if that is your business; an agent with it in context will propose those tactics as normal steps.

baseline what the model already does  →  research only the gaps  →  distil  →  ship the delta

Skills for models that already know

Claude, GPT, Gemini, Grok, DeepSeek, Qwen, Kimi, GLM — the 2026 generation already holds most of the public internet. Restating documentation spends context teaching what the model would produce unprompted; topically-adjacent-but-useless content degrades its reasoning. The unit here isn't the topic — it's the delta: what the model gets wrong, doesn't know, skips under pressure, or solves weaker than a stronger option that exists.

Not tutorials. Not documentation summaries. Not generic best practice restated in Markdown.

Not our idea alone

• Anthropic, skill-authoring guidance — "Only add context Claude doesn't already have."
• Microsoft, Waldek Mastykarz — "How do you know what the model knows? You don't, unless you measure." And: "models don't need a textbook. They needed a cheat sheet."
• ETH Zurich, Evaluating AGENTS.md — repository context files did not generally improve task success rates, and raised inference cost by more than 20%.

Three deltas, verbatim

A competent outsider wouldn't reliably know these — nor would a model with no skill loaded.

iOS 18 does NOT update window.innerHeight when address bar expands; 100vh always equals lvh on iOS Safari.

— skills/responsive-adapter/references/platform-quirks.md

Defense: resolve DNS, check IP, disable redirects, re-check on every socket connect. Production: dedicated egress proxy (Smokescreen, ssrfproxy) with connect-time IP validation — application-layer checks are racy.

— skills/secure-coding/ssrf.md

Daily CPL (for BUYING) = click-date spend (account-tz day) ÷ that same click-date cohort's payout count. ... pairing click-date spend with conversion-date conversions is the classic apples-to-oranges CPL.

— skills/tracker-ops/references/03-metrics-and-math.md

How they're built

Run real tasks with no skill → research only the confirmed gaps → keep only rules traced to a prevented failure → rerun the draft and cut what didn't fix anything. Full method: docs/METHODOLOGY.md.

Featured

Every skill, by domain: CATALOG.md.

Install

Copy the skill directories you want — nothing else to configure.

git clone https://github.com/sergeyizmailov/knowledge-delta-skills.git
mkdir -p ~/.claude/skills
cd knowledge-delta-skills/skills
cp -R meta-ads ~/.claude/skills/                                                    # one skill
cp -R meta-ads google-ads google-feed-ops tracker-ops measurement-experimentation-ops ~/.claude/skills/   # clean lane
cp -R meta-grey-ops google-grey-ops senior-buyer-ops ~/.claude/skills/              # + grey lane (opt-in)

Personal-scope directory by runtime, verified against each vendor's docs on 2026-08-29. Most also read a project-local equivalent.

Media-buying lanes (skills cross-reference each other by name — install a whole lane):

Pointers from clean skills into grey files degrade to "not installed"; nothing breaks. Playbook numbers (CPL, kill thresholds, tags) are one team's priors — replace with your contract before an agent acts on them.

Nothing here is permanent

Base models absorb more of the internet every release, so a skill earns its place only while its gap stays open — a re-baseline that finds a section already correct cuts that section, and skills shrink or retire as their delta closes: docs/METHODOLOGY.md.

Contributing

Quality bar, skill anatomy, local checks CI runs: CONTRIBUTING.md.

License

MIT