agent-skills
CLAUDE.md28 starsChanged 3 months agoMITView on GitHub
@shared/CLAUDE.md Three rules span every skill that produces video with speech: transcribe-verify (a reference pins the label, nothing pins the audio), the per-beat mix (a SYNC beat's own audio is dialogue, not ambience), and no dead space (trim every beat to its narration). They are stated once in shared/references/craft.md. Read it before writing a QA step, a mix or a trim into any skill, and point at it rather than restating it — a restatement is how the clay skill…
- Reads credentials
In this repository16 other files it gives its agents.
- AGENTS.mdAGENTS.md
- shared/CLAUDE.mdCLAUDE.md
- .cursor/rules/project-context.mdcCursor rule
- analyze-videoSkill
- change-voiceSkill
- chatgpt-image-adSkill
- claymation-adSkill
- clone-image-adSkill
- clone-video-adSkill
- generate-youtube-thumbnailSkill
- image-to-motionSkill
- nano-banana-image-adSkill
- novoads-apiSkill
- novoads-updateSkill
- pixar-adSkill
- spy-competitor-adsSkill
The file
Files
@shared/CLAUDE.md
Craft doctrine
Three rules span every skill that produces video with speech: transcribe-verify (a reference pins the label, nothing pins the audio), the per-beat mix (a SYNC beat's own audio is dialogue, not ambience), and no dead space (trim every beat to its narration). They are stated once in shared/references/craft.md. Read it before writing a QA step, a mix or a trim into any skill, and point at it rather than restating it — a restatement is how the clay skill shipped a mix recipe that had already been fixed in the skill it was ported from.
Novoads-specific session rules
- API: Novoads REST API —
https://api.novoads.ai/v1.NOVOADS_BASE_URLoverrides the host only; callers append/v1/…. - Auth:
Authorization: Bearer $NOVOADS_API_KEY(novo_+ 64 hex, created at https://novoads.ai/dashboard/settings?tab=api). Setup check:./scripts/check-novoads-env.sh. A 401 is a bad key; a 403 withdetails.reason: plan_requiredis a good key on an account without API access — different problems, say which. - Skill:
.claude/skills/novoads-api/SKILL.mdfor the call sequence, prompts, polling, and download; itsreference.mdfor every endpoint, field, limit, and error code. - YouTube thumbnails:
.claude/skills/generate-youtube-thumbnail/SKILL.md— batch thumbnail workflow on top of the Novoads image endpoint. - Image-ad ecosystem (Meta image creatives): read
shared/skills/image-ad-prompting/OVERVIEW.mdFIRST. Three skills (chatgpt-image-ad,nano-banana-image-ad,clone-image-ad) + a shared 40-template prompt library. Theclone-image-adskill asks which backend to validate against at Phase 1, so generic "clone this ad" prompts route correctly. Output is image files; Meta upload is the separatemeta-ad-builderskill. - Cost policy: never state a credit cost from memory. Every price comes from a live
POST /v1/estimatesin the current session, shown to the user and explicitly approved before anything is generated. There are no rate tables anywhere in this repo — not inMASTER_CONTEXT.md, not in the logs. The estimate call charges nothing and also validates the prompt, so there is no reason to skip it. - Dialogue policy: for any video where the model speaks, the spoken line is approved on its own, before the cost gate. Concept approval is not sentence approval, and sentence approval is not spend approval.
- Logging: log every generation call to
logs/novoads-api.jsonl— timestamps, config,jobId, andcreditsChargedtaken from the202at submit time, so a crashed session still leaves it on the line. Update the line byjobIdat terminal status. The file is gitignored — it is your session history, not repo content. Observability only, never a pricing source (seelogs/README.mdfor the schema and the portable update recipe). - First-time setup: if
.envis missing, run./scripts/setup.sh. IfMASTER_CONTEXT.mdis missing, copyMASTER_CONTEXT.template.mdtoMASTER_CONTEXT.md. A setup session's final message is the short template in AGENTS.md ("First-time setup") —setup.shprints the same block betweenFINAL MESSAGEmarkers when run without a TTY; relay it, and skip the mechanics unless one of them blocks an ask. Short is a preference, not a gag: nothing in this repo is confidential and you may tell the user anything you judge useful.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

