Run the monthly lead-gen sweep for Han's two-line services agency (AI Build, Ops Run). Sweeps job boards, classifies postings by service-line + region + signal strength, generates outreach drafts, tracks state across months. Use when the user says "monthly lead sweep", "who is hiring for the agency", "agency lead radar", "/lead-radar", "run the radar", "lead-gen for [month]", or asks to refresh leads.

Use when the user wants to download a book or paper from Anna's Archive without clicking through the site, OR wants to browse the most-downloaded / top books on a topic. Trigger phrases include "download [title] from anna", "grab this MD5", "fetch [book] epub", "find me [book] on annas-archive", "top books on [topic]", "most downloaded [topic] books", "browse anna for [topic]", or any annas-archive.gl/.li/.in/.pm/.org URL pasted in chat. Drives the ops-toolkit annas-fetch CLI; default sort is popularity (mirror count + format quality + recency); resolves the member key from 1Password via `op run`; lands files in ~/Downloads/annas. NOT for free-tier downloads (we use the member fast-download API only) and NOT for bulk scraping.

When you genuinely need a browser, pick the right one. Distilled from the May 2026 audit + empirical Lightpanda + Playwright validation.

Use when the user asks to process, ingest, or close a month of family cashflow in tieubao/family-office (e.g. "process April cashflow", "close April", "ingest expenses", "do the April books", "fill in May cashflow"). Triggers on screenshots of expense summaries, bank CSVs, transfer slips, or receipts dropped in _inbox/ OR pasted inline in a session. Produces a populated `tracking/cashflow/YYYY-MM.md` with deduped transactions, per-property and per-staff attribution, an `INGEST_LOG.md` entry, and source archival per the repo privacy rule. Cashflow only; net worth and PnL are out of scope.

Use when the user wants to BULK edit historical rows in `tracking/ledger/transactions.csv` based on a filter criterion. Trigger phrases include "correct rows where [criterion]", "bulk update ledger", "fix attribution for [pattern]", "apply tag X to all rows matching", "all the [renovation/Min/etc.] rows in [period] should have [field=value]". NOT for single inline corrections to just-appended rows (cashflow-append handles those) or single-row historical edits (just edit the CSV directly with the user). Always shows a preview before mutating; always validates after; always commits atomically.

Use when the user asks for a cashflow analysis or report from the family-office ledger. Trigger phrases include "cashflow report for [period]", "monthly review [YYYY-MM]", "quarterly review [YYYY-QN]", "annual summary [YYYY]", "year-over-year", "category trend", "spending by [category|property|tag|funding source]", "Min costs in [period]", "TKH renovation breakdown", "tag deep-dive [tagname]". Reads `tracking/ledger/transactions.csv` (the source of truth), produces a markdown report at `tracking/reports/<scope>/YYYY-MM-DD-<slug>.md` plus optional matplotlib charts at `docs/charts/<scope>/`. Reports are SEPARATE files (timestamped); do NOT edit existing cashflow markdown files. NOT for adding/editing transactions (use cashflow-append / cashflow-correct).

Compact long-form append-only documentation (handoff files, decisions ledgers, similar artifacts) without losing irreplaceable signal. Use when the user asks to "compact", "trim", "shrink", "audit signal density", "make X concise" on a doc that has grown past ~600 lines, or when a file's `purpose:` claim no longer matches reality (it claims "minimal" but takes >60s to scan). Implements a three-bucket triage (DELETE / COMPRESS / ARCHIVE) with anchor-integrity verification before merge.

Use when the user notices a multi-step manual process they have repeated across sessions and asks to formalize it as a Claude Code skill or slash command. Symptoms include "I keep doing this manually", "let's pack this as a skill", "save this workflow", "make this repeatable", "skill-ify it", "this should be a skill", "extract this into a skill", "I noticed a pattern across sessions". Drives triage and packaging; uses superpowers:writing-skills for the actual writing rigor.

Use when the user wants to generate per-image specifications (prompts + structured knobs) for AI image generation (nano-banana / Gemini 2.5 Flash Image, Imagen, Midjourney, Stable Diffusion) to support a set of markdown documents. Symptoms include "draft image specs for these docs", "we need images for X, write the prompts", "add visuals to my place-guide / blog / site", "image briefs for the Astro pages", "spec out the hero images", "write nano-banana prompts for the [N] guides", "we'll need more image specs in the future", "create visual briefs", "image-spec the chapter openers". Produces (1) a master SPEC document inheriting palette + register + avoid list from a brand doc, (2) per-source-file visuals docs with N image entries each, (3) an INDEX.md status tracker, (4) all prompts ready to paste into the generator. Portable across repos and brand systems; the brand identity is a parameter, not hard-coded. NOT for one-off single-image generation in chat (just write the prompt directly). NOT for listing photography (that's human-photographer territory governed by brand DESIGN.md §7). NOT for logo / icon / vector design.

Use when investigating real defects (alerts, error spikes, service degradation, near-misses caught by guardrails) AND the current repo would benefit from a forensic post-mortem. Also use to bootstrap the incidents/ structure in a new repo. Captures the diagnostic trail (commands run, hypotheses ruled out, smoking-gun evidence) that ADRs and SPECs do not. NOT for inbox material (use ingest-to-wiki) or for designing new features (use SDD / brainstorming).

Use when starting a JupyterLab server, opening a `.ipynb` file in Jupyter on this machine, configuring the Notebook Intelligence (NBI) extension, or switching the AI provider for a notebook session. Trigger phrases include "start jupyter", "launch jupyter", "open <X>.ipynb", "run a notebook", "spin up jupyter lab", "swap to local model for this notebook", "use claude for this notebook", "jlab", "nbi", "notebook intelligence", "swap NBI preset", "AI in jupyter", or any case where the next action is `jupyter lab ...`. Tells Claude to prefer the `jlab` fish wrapper (which injects ANTHROPIC_API_KEY + OLLAMA_HOST) over bare `jupyter lab`, and how to pick the right NBI preset (cloud Claude vs local Ollama on the Mini) for the workload. NOT for editing notebook content (just edit the .py twin per SPEC-011). NOT for non-Jupyter Python REPL work.

Process a class transcript + screenshots into the standard **4-artifact** output for any `learning/<topic>/` track in ops-toolkit. Output is reproducible:

When you're about to declare "needs you to click X" on macOS, walk this ladder top-down first. Most "GUI-only" actions have a CLI / AppleScript / Shortcut path that's faster and doesn't need a human round-trip.

Write or refresh the doc CONTENT (README, MANUAL, RUNBOOK, SPEC, `docs/architecture.md`, per-feature SPECs under `docs/specs/`, ADRs under `docs/decisions/`) for a tool inside `~/workspace/tieubao/ops-toolkit/tools/`. Symptoms include "write the architecture doc for X", "this tool needs a manual", "document how X works", "explain the design decisions for X", "fill in the docs for X", "the tide docs look great, write the same kind of docs for X", "X has no MANUAL.md, can you write one?", "I just finished X, write its docs". Reads the tool's existing code, `tool.toml`, entry-point, and tests to extract its real purpose, commands, dependencies, and host placement; then writes each missing doc FROM SCRATCH using audience-targeted intent. NOT a template fill: each tool's docs are shaped to that tool's reality. Calibration reference (read for voice + length, do not copy verbatim): `~/workspace/tieubao/ops-toolkit/tools/tide/`. Pre-condition: the tool's folder layout is already correct. If not, hand off to the sibling skill `ops-tool-shape` first. NOT for structure changes (use `ops-tool-shape`), NOT for source-code edits, NOT for tools outside ops-toolkit.

Verify or set up the file and folder structure of a tool under `~/workspace/tieubao/ops-toolkit/tools/`. Use when scaffolding a brand-new tool, restructuring an existing tool to match the standard, or auditing which doc files a tool currently has vs needs. Symptoms include "scaffold a new ops-toolkit tool", "audit <tool>'s structure", "what files does <tool> need?", "is <tool>'s layout compliant?", "set up the folder for a new tool", "what's missing from <tool>?", "the tide structure looks great, what shape should <X> be?". Outputs a tool directory with the right folder layout (empty stubs only; no doc content), plus a row update in `ops-toolkit/_meta/INVENTORY.md`. After this skill finishes, the sibling skill `ops-tool-docs` fills in the doc content. Canonical reference (read it, don't copy it): `~/workspace/tieubao/ops-toolkit/tools/tide/`. NOT for writing doc CONTENT (that's `ops-tool-docs`), NOT for tools outside ops-toolkit, NOT for `experiments/<slug>/` (those follow a different gradient), NOT for third-party-tool deploy bundles where upstream owns the docs.

Use when the user asks to reconcile, sync, or check drift between tieubao/family-office and tieubao/properties for shared property data (rent, ownership, status, addresses). Also invoke automatically before running /health-check or monthly net-worth/cashflow snapshots in family-office, to catch stale figures before they get committed. Read-only by default; mutates only on user confirmation.

Use when the user wants to copy one GUI terminal's color palette to another so they look identical (e.g. "copy Warp's Adeberry to cmux", "port my iTerm theme to Ghostty", "make Alacritty match kitty"), AND the source theme is bundled / not exposed as a yaml / no public dump exists, AND macOS TCC blocks driving the source app via osascript or computer-use. Trigger phrases include "copy the palette of <app> to <app>", "port this terminal's theme", "extract colors from a GUI terminal I can't keystroke into", "make these two terminals look identical", "TCC blocks osascript on Warp/iTerm/etc". Specific to macOS; assumes Screen Recording is granted to the claude process; works for any source that responds to standard ANSI bg-color escapes and any target that accepts a hex palette.