IndyDevDan — The Library Meta-Skill: How I Distribute PRIVATE Skills, Agents and Prompts
Why this is in the vault
26-minute Dan walkthrough of the library meta-skill — a SKILL.md + YAML reference file + cookbook of agent-executed commands (/library add, /library use, /library push, /library list, /library search, /library sync) that turns a developer’s collection of private skills/agents/prompts (“agentics”) into a synced library across devices, team members, and agents. The vault keeps it for three reasons. (1) It is the missing distribution layer for RDCO’s own ~/.claude/skills/ ecology — Ray currently has 50+ skills on the Mac Mini, no equivalent on Ben’s MacBook, and no clean way to share or sync between them; this video describes the exact pattern Ray should adopt. (2) It is a load-bearing example of pure-agentic applications — no compiled code, no CLI binary, just SKILL.md + YAML + agent reasoning. Dan’s prediction (“we’re going to see a lot of pure agentic solutions as we move through this year”) is materially relevant to skill-design decisions in 2026 — the bar for “this should be a Python script” vs “this should be a SKILL.md” is shifting. (3) It strengthens CA-013 (R&D framework: Reduce and Delegate) with a concrete reduce-the-CLAUDE.md-load mechanism — references over copies = less context bloat per device — and adds an operational answer for the multi-device synchronization sub-problem that’s been latent in RDCO’s channels-agent stack since the Mac Mini was provisioned.
Core argument
- The problem is structural, not annoying. When you operate across 10+ codebases × multiple agents × multiple team members, copy-pasting skills between repositories is technical debt that compounds invisibly. Skills go out of sync, duplicates accumulate, and there’s no shared truth.
- Solution: a meta-skill that references rather than copies. The library skill stores pointers (private GitHub URLs, public repos, local file paths) in a single YAML file — never the skill content itself. Like package.json or pyproject.toml for prompts/agents/skills.
- Three primitives, not one. Skills (raw capabilities) ≠ Agents (scale and parallelism) ≠ Prompts (single-file orchestration). Most engineers conflate them and “overkill skills doing everything inside the skill.” Dan’s distribution surface treats all three as first-class citizens.
- The four-stage workflow: Build → Catalog → Distribute → Use. Build natively in your value-generating repo; catalog with
/library add; distribute via git/sync; use via/library use <name> install [locally|globally]. - Pure-agentic application. The library skill has zero compiled code. SKILL.md describes intent; library.yaml stores references; the cookbook directory contains markdown for each command (one .md per workflow). The agent reads these and executes git operations directly. “There is no code associated with this… that gives us a lot of very powerful capabilities.”
- Privacy-first. “Your top-notch skills, agents, and prompts should absolutely be private. I know a lot of vibe coders, you think that everything is just out there in the public. Trust me, it’s not.” The library is designed for private GitHub distribution as the default; public repos are an exception, not the rule.
- Local + global namespacing per consumer device. Same library can install items into
<repo>/.claude/skills/(project-scoped) OR~/.claude/skills/(global). The consuming agent decides scope per/library useinvocation, NOT the library itself. Some skills want to be global on all devices (meta-agentics); others belong only in the project that needs them. - Push-back-to-source is a first-class command.
/library push <name>finds local edits and commits them back to the referenced repo. Bidirectional sync, not just one-way distribution. This is what makes the library a working source of truth, not just a deployment surface. - Trust scaling is the 2026 theme. “The best way to increase trust is to know exactly what they’re doing and know exactly what they’re running. And that means down to the lines inside of your skills, your prompts, and your agents.” The library forces every shared piece through a private-controlled distribution surface — no
cloud code plugins, install from whereverfor production work. - Orchestrator agents need this. As the agentic path scales (base → better → more → custom → orchestrator), the orchestrator can’t operate effectively without a single coherent capability surface across devices and team members. The library is the prerequisite for serious orchestrator work.
Notable details
- [00:01:00] Threshold for adoption: “If you’re an engineer working on one to two repos, what I’m about to show you doesn’t matter at all.” Explicitly scoped to multi-device, multi-team, multi-agent operations.
- [00:02:30] Stripe precedent: Dan name-checks Stripe’s “tool shed” as the prior art for meta-tooling. The library extends this from tools-only to skills+agents+prompts.
- [00:03:30] The library file is a YAML reference document. Three top-level sections:
skills,agents,prompts. Each entry is a name + source reference (GitHub URL, local file path). - [00:04:30] Pure-agentic prediction: “I’m predicting we’re going to see a lot of pure agentic solutions as we move through this year.” Significant claim — implies the bar for compiled-code solutions is shifting upward.
- [00:06:30] The 6-command API:
add,use,push,list,search,sync. Sync pulls the actual code bases, not just refreshes the catalog index. - [00:09:30] Two libraries demonstrated:
idubdan-library-engineer(the actual reusable agentics — multiple of these per role/domain like support, sales, engineering) ANDlibrary(the personal reference YAML pointing to the engineer libraries). Important separation — the consumer library is small and personal; the producer libraries are bigger and role-scoped. - [00:12:30] PI coding agent demo: Dan uses his customized PI coding agent (covered in a separate video) with damage-control hooks like “skip rm-rf” baked in. The library skill works identically with Claude Code OR PI — agent-harness-agnostic.
- [00:13:30] Agent-application-as-skill paradigm: “There’s a very interesting paradigm that’s emerging where you can encode entire applications into a single prompt into a single skill as anthropic likes to call it.”
- [00:16:30] Bug confession on-stream: Dan notes the default directory for prompts should be
commands(Claude Code’s actual storage location), notprompts, and commits to fixing it. Honest engineering moment. - [00:21:00] Why prompts deserve a library: Dan’s meta-prompt is “like 800 lines long” — too important to drift across devices. “The prompt this important needs to be the exact same across all my team members, all of my work on all my code bases and on all my devices, both human operated and agent operated.”
- [00:22:00] Push demo with model swap: Dan deliberately swaps from Quinn 3.5 Flashlight to Sonnet 4.6 (“nothing more stable than Claude Sonnet noise models”) for the push operation — model selection per task, not per session. Trust scaling principle in practice.
- [00:24:00] Cookbook structure for skills: Dan’s recommended skill organization is
SKILL.md+cookbook/directory with one .md per individual workflow that the SKILL.md describes when to invoke. This separation is itself an R&D move — the cookbook entries don’t load until referenced.
Mapping against Ray Data Co
- The library is the missing distribution layer for RDCO’s ~/.claude/skills/ ecology. Right now skills exist on the Mac Mini in one set, on Ben’s laptop in a different set, and there’s no synchronized truth between them. The library pattern says: maintain ONE private GitHub repo per skill family (rdco-skills-channels, rdco-skills-content, rdco-skills-vault, etc.), maintain ONE library.yaml on each device pointing at those repos, and use
/library syncto keep them aligned. Probably 4 hours to scaffold the meta-skill, 2 hours to migrate the first batch of skills, and an ongoing discipline thereafter. High-priority operational improvement. - The “pure-agentic application” pattern is a critical design heuristic for new skills. RDCO has multiple skills with backing scripts (
graph-ingest.py,audit-newsletter-outputs.py,imessage-send-file.sh) — sometimes the script is genuinely needed (binary protocols, system calls), sometimes it’s a leftover from before SKILL.md + agent reasoning could do the job. Worth a one-pass audit: for each script-backed skill, ask “could this be pure SKILL.md + cookbook + agent reasoning?” Net result: smaller surface area, fewer dependencies, easier review. - The library/library-engineer split maps directly to RDCO’s split between operational skills and content skills.
idubdan-library-engineer(producer-side, role-specific) ↔library(consumer-side, personal reference). RDCO’s analog:rdco-skills-engineering(operational —/check-board,/process-newsletter,/process-youtube, etc.) ↔ a personal library YAML on each device that picks WHICH of those operational skills are activated for that device’s role. Mac Mini activates everything cron-driven; Ben’s laptop activates only the human-in-loop skills. Clean separation of distribution from activation. - The “agent-harness-agnostic” property is a hedge against vendor risk. Dan’s library skill works identically with Claude Code AND PI coding agent. RDCO is currently 100% Claude Code dependent — if Anthropic shifts pricing, deprecates a feature, or pushes an undesirable harness change, every skill stops working. A pure-agentic library that runs in any harness is a vendor-portability moat. Worth the design constraint: every new RDCO skill should pass a “would this run in PI / Cursor / Cline?” smell test even if Claude Code is the default. The library skill itself is the proof that this is achievable.
/library pushis the missing bidirectional sync RDCO needs for cross-device skill iteration. Today, when Ray edits a skill on the Mac Mini in the middle of a cycle, that edit is stranded on the Mac Mini until manual git commit + push. The library push command bakes that into the skill ergonomics. Concrete proposal: when/improvemakes a skill change, the cycle should end with/library push <skill>so the change propagates to all consumers automatically. Eliminates a class of “Ray’s Mac Mini has the latest version, Ben’s laptop doesn’t” failure modes.- The “trust scaling is 2026’s big theme” framing is the right operational philosophy for RDCO. SOUL.md already gestures at this (“Vault metadata discipline”, “audit-newsletter-outputs as Jepsen-style proof”), but Dan articulates it more sharply: trust scales with the engineer’s ability to know-line-by-line what every skill does. Implication: every external skill installation (Anthropic plugins, marketplace skills, etc.) is presumptive technical debt. RDCO’s skills should be a closed garden — if it’s not in our private library, it doesn’t run on our devices.
- CA-013 (Reduce and Delegate context discipline) gains a concrete reduction mechanism. The library pattern reduces per-device context load: each device only has the skills currently activated, not the full producer-library catalog. Maps the R&D framework to a file-system-level reduction — different device profiles activate different subsets of the global skill universe. This is a 5th source for CA-013 if we count library-skill-as-distribution-mechanism as a distinct articulation of the reduce principle.
- The orchestrator-prerequisite framing maps to where RDCO is heading. Dan’s path (base → better → more → custom → orchestrator) is exactly the trajectory of RDCO’s autonomous loop. The current
/check-boardcycle is approaching orchestrator behavior. The library is the prerequisite Dan names for orchestrator work to be sustainable. Good evidence the library investment is timely, not premature.
Open follow-ups
- Build the RDCO library skill. ~/.claude/skills/library/SKILL.md + library.yaml + cookbook/ for the 6 commands. Migrate the first 5 skills (
/check-board,/process-newsletter,/process-youtube,/sync-contacts,/finance-pulse) into private GitHub repos and reference them from the library file. ~6 hours total. Single biggest leverage move from this video. - Audit script-backed skills for pure-agentic conversion candidates. Walk every SKILL.md in ~/.claude/skills/ and flag those with backing scripts. For each, decide: keep (genuine system-call need) or convert (SKILL.md + cookbook can do it). Likely 30% are convertible. Reduces dependency surface, easier review, agent-harness-portable. ~2 hours audit, then per-skill conversion as opportunity arises.
- Add
/library pushto the end of/improve. When/improvefinishes editing a skill, the cycle ends with a push to the source repo. Eliminates cross-device staleness. ~1 hour to scaffold the integration. - Define device profiles.
mac-mini.library.yamlactivates cron skills + autonomous loop skills;ben-laptop.library.yamlactivates founder-in-loop skills (review, planning, research). Document the rationale in SOUL.md so future device additions know which profile to start from. ~1 hour. - Sanity Check angle: “Pure-agentic applications are the new bar.” Lead with the visceral image of an entire application encoded in a SKILL.md + YAML + cookbook with zero compiled code. Pivot to the data-engineering audience: most ETL pipelines, data-quality checks, and ad-hoc analysis tools are still scripts because the LLM-reasoning-as-execution mode is unfamiliar. Dan’s library is the proof point that the bar has moved. Land on the design heuristic: every new tool you build should fail a “does this need code?” test before you write the first line of Python. ~1500 words.
- Add CA-020 candidate: pure-agentic-application as a distinct architectural pattern. Sources so far: this video (library skill = pure-agentic), 2026-04-20-indydevdan-agent-experts-self-improving (expertise.yml + agent reasoning, no compiled code). Need 1 more source for the 3+ ripeness bar — likely candidate is the next IndyDevDan video that articulates the pattern even more explicitly, or a Cobus Greyling piece on prompt-as-application.
Sponsorship
The video has no paid placement. Dan’s standard close (“you know where to find me every single Monday — stay focused and keep building”) is brand-building, not sponsorship. The video does promote Dan’s own GitHub template (disler/the-library) and his previously-released PI coding agent video, but these are self-references to his own ecosystem, not paid advertising.
Related
- ~/rdco-vault/06-reference/transcripts/2026-04-20-indydevdan-library-meta-skill-transcript.md — full transcript
- ~/rdco-vault/06-reference/2026-04-20-indydevdan-agent-experts-self-improving — the meta-agentics that the library distributes are the agent-expert pattern operating at scale
- ~/rdco-vault/06-reference/2026-04-20-indydevdan-agent-threads-boris-cherny — the orchestrator-agent destination this video frames the library as prerequisite for
- ~/rdco-vault/06-reference/2026-04-20-indydevdan-one-agent-to-rule-them-all — adjacent piece on orchestrator design constraints
- ~/rdco-vault/06-reference/2026-04-19-indydevdan-cracked-claude-agent-skills — skills as the static building blocks that the library distributes
- ~/rdco-vault/06-reference/2026-04-15-thariq-claude-code-session-management-1m-context — Anthropic’s case for context discipline that the library answers at the file-system layer
- ~/rdco-vault/06-reference/concepts/CANDIDATES.md — strengthens CA-013 (R&D context discipline) with a concrete reduce-by-distribution mechanism; potential new candidate CA-020 (pure-agentic-application architectural pattern)