Chain
Named skills, in order.
The chain tells the agent what to load first, what each stage produces, and where the next skill picks up the work.
← Open Skills overview
Open Skills runbooks.
A skill makes one kind of agent behavior reliable. A runbook makes a whole workflow reliable. It names the chain, the handoffs, and the point where a human still needs to approve, choose, or judge.
This is where Open Skills starts to look less like a library and more like leverage: a voice memo becomes a page, a release becomes a briefing, a rough recording becomes a finished video, and the discoveries from one run make the next run faster.
The primitive is the unit. The runbook is the production line you build from those units.
10 runbooks
Same primitives, recombined into useful outcomes.
Runbooks show why small skills matter. You can swap, refine, and reuse the same primitives across different workflows instead of rebuilding the whole operation every time something changes.
Handoff
Humans keep the judgment.
Runbooks should move mechanical work forward, but approvals, taste calls, secrets, accounts, and final judgment still belong to the person running the system.
Payoff
The result is concrete.
A good runbook ends in an artifact you can inspect: a page, a briefing, a timeline, a verified deploy, or a better operating map.
Runbook 01 · Talk to Published
Media Transcription → Brain Dump Processor → Personal Voice → HTML Artifact Builder → Personal Site Publisher
You record a voice memo on a walk. Transcription turns it into clean text; the Brain Dump Processor separates and evaluates the ideas in it; you pick the one worth writing; the Voice skill drafts the piece as you'd write it; the Artifact Builder lays it out; the Site Publisher ships it to a clean URL with a proper link preview.
The payoff: A voice memo becomes a published page.
Runbook 02 · Release Day
Current-Information Search → New Release Briefing → Branded Image Prompting → Image Generation Gateway → Personal Site Publisher → Stakeholder Update Email
Something big ships in your field at 10am and you want an accurate, on-brand briefing live by noon. Search gathers primary-source facts with dates (this step is what keeps you from publishing training-data hallucinations); the Briefing skill packages them into your standard format; the image skills produce a matching branded thumbnail; the Publisher ships it; the Update Email tells your list or team it's live.
The payoff: An accurate, on-brand briefing published the same day with speed that never costs correctness.
Runbook 03 · The Video Production Line
Media Transcription → Radio Edit → B-Roll Pipeline → AI Editing Assistant → Stakeholder Update Email
Raw talking-head footage in, finished video with motion graphics out. Transcription produces the timestamped foundation everything else reads. Radio Edit fixes the spoken narrative and hands you a paper edit to approve — the editorial decisions happen here, on paper, where they're cheap to change. The B-Roll Pipeline scouts the approved cut for graphic-worthy moments and generates consistent animated overlays. The NLE Assistant assembles it in your editor. The Update Email tells your editor or client it's ready for review.
The payoff: A raw video becomes a finished, graphics-laden edit with the editorial work front-loaded and cheap to change.
Runbook 04 · Ship a Page You Can Trust
Frontend Taste System → Personal Site Publisher → Browser Automation QA → Testing Runbook Creator
The difference between shipping a page and shipping a page you'd bet on. The Taste System builds it well; the Publisher takes it live; Browser QA then verifies the live page with instruments rather than vibes — screenshots across breakpoints, Core Web Vitals, console and network checks — and everything QA learned about testing this page lands in the repo's runbook, so the next deploy verifies in minutes.
The payoff: A personal site with a regression-test habit — every page shipped with verified quality.
Runbook 05 · The Research Engine
Heavy File Ingestion → Current-Information Search → Assumption Checker → Meeting Synthesis → HTML Artifact Builder → Reading Pack Builder
For real research questions with messy inputs: a folder of PDFs, some meeting recordings, and a claim you're not sure you believe. Ingestion converts the heavy sources into clean artifacts first (this ordering is the whole trick — analysis over converted text is faster, cheaper, and reusable). Search fills the gaps with current information. The Assumption Checker runs adversarially against the emerging conclusions — a separate skill with a skeptic's posture, not the same conversation grading its own homework. The output ships as a styled artifact, and when the material needs human review, the Reading Pack presents it in order.
The payoff: Research with a chain of custody: every claim traceable to an artifact, every conclusion stress-tested.
Runbook 06 · Delegate and Verify
Session Operating Map → Goal Prompt Generator → Visible Delegation → Self-Authored PR Merge → Stakeholder Update Email
How one person runs parallel engineering lanes without becoming the bottleneck. The Operating Map records what each lane owns, so no session needs you to explain the project. The Goal Prompt Generator packages a task with a definition of done and verification gates; Visible Delegation runs it in a watchable session; when the delegate finishes, its work is verified against the gates it was given — the goal prompt is also the acceptance test. The PR Merge skill reviews and lands it; the Update Email closes the loop with whoever's waiting.
The payoff: Parallel engineering lanes with you only touching the two decisions that need you: what 'done' means, and whether the diff is good.
Runbook 07 · The Flywheel
Session-to-Skill Extractor → Testing Runbook Creator → Page Testing Memory → Session Operating Map
This one is different: it's not a pipeline you run, it's a posture that runs under every other runbook. The Extractor watches your sessions for patterns worth keeping and drafts new skills from them. The Runbook Creator banks every testing discovery in the repo it belongs to. Page Testing Memory keeps the global/local boundary clean as both libraries grow. The Operating Map preserves coordination state across sessions.
The payoff: No useful discovery dies in chat — the mechanism by which a skill library compounds.
Runbook 08 · Claim Appeal Packet
PDF / Document Ingestion → Document Chunking and Tagging → Case Data Normalization → SQLite Case Store → Deterministic Retrieval Map → Citation Guard → Packet Export → Human Gate
A denied claim becomes a grounded appeal packet. Ingestion converts plan docs and the denial into citeable artifacts. Chunking and tagging makes the plan sections addressable. Normalization extracts dates, denial reason, claim lines, and deadline, then reconciles each claim's fields across the denial letter and EOB: codes, dates, and amounts; mismatches become named review questions in the packet. The case store keeps the evidence queryable. Deterministic retrieval maps denial type to plan language before the agent drafts. Citation Guard rejects unsupported claims and blocks export until every failure is fixed or converted to a named review question. Packet Export produces the review folder. Human Gate stops before filing or sending.
The payoff: A denial letter becomes an editable, cited appeal packet a human can review and send.
Runbook 09 · Tax Prep Packet
PDF / Document Ingestion → Document Chunking and Tagging → Case Data Normalization → SQLite Case Store → Open Brain Case Store (optional · OB1 path) → Deterministic Retrieval Map → Citation Guard → Packet Export → Human Gate
A pile of tax documents becomes a CPA-ready review packet. Ingestion converts forms, receipts, and CSVs. Chunking and tagging makes the form boxes, receipt lines, and statement rows addressable evidence. Normalization merges receipt and bank evidence so one transaction is one ledger row, cross-checks every W-2 and 1099 against deposits, and turns unmatched payers into named missing-document items. SQLite is the beginner store; Open Brain is the OB1 path for people with durable context already running. Deterministic retrieval maps income, expense, and missing-document types to the right review rules. Citation Guard keeps tax-rule claims grounded and blocks export while any claim fails. Packet Export creates summaries, ledgers, questions, and PDF. Human Gate stops before filing.
The payoff: A messy tax folder becomes a structured prep packet with evidence, questions, and clean handoff artifacts.
Runbook 10 · Email Follow-Up Packet
PDF / Document Ingestion → Document Chunking and Tagging → Case Data Normalization → SQLite Case Store → Open Brain Case Store (optional · OB1 path) → Deterministic Retrieval Map → Citation Guard → Packet Export → Human Gate
A mailbox export becomes a commitments ledger and a folder of cited drafts. Ingestion converts an mbox export, Sent folder included, into citeable messages anchored by message-id. Chunking strips quoted history, signatures, and disclaimers so evidence anchors to the message where a sentence first appeared. Normalization reconstructs threads from headers, resolves sender identities, and extracts commitments and waiting-on rows with owners and due dates, then reconciles restated promises into single rows and closes any loop the Sent mail already answered. Deterministic retrieval maps thread state, dropped commitment, waiting-on-them, or dispute, to the evidence a draft needs. Citation Guard rejects any draft claim that lacks an anchoring message. Packet Export produces drafts with full headers, every one marked pending. Human Gate is the send boundary: fixture runs hold no mail credentials, the live loop ingests through a read-and-draft connector with no send verb, and an ignored draft means no.
The payoff: A neglected inbox becomes an urgency-ordered ledger and ready-to-send cited drafts, and nothing sends itself.
Back to the Skills directory.