What this skill carries — grep references/ for any subject the names below miss:
!d=${CLAUDE_SKILL_DIR}; command -v skill-toc >/dev/null 2>&1 && exec skill-toc "$d"; s=$HOME/.claude/skills/plugin-utils/bin/skill-toc; [ -x "$s" ] && exec "$s" "$d"; echo "(skill-toc unavailable: references and scripts are NOT listed here — install the plugin-utils plugin, or start a new session so its bin/ reaches PATH)"
Route before you act. Before the first tool call of a turn, decide which of these owns the task. The main thread's job is routing, not doing.
The routing table
Read top to bottom; the first row that matches wins.
| The task is | Route to | How |
|---|---|---|
already-invoked skill (<command-name>/dev</command-name> in this turn) |
it is already loaded | just follow it — never re-invoke |
the prompt opens Implement the following plan: and the plan's frontmatter has workflow: <name> |
that skill | Skill(skill="workflows:<name>") (as written when plugin-qualified) FIRST — the context was cleared at approval; never work, never inline |
| anything "in a new / background / separate / companion session" | agent-spawn skill | it is the transport; the real task goes inside its prompt |
| a feature, bug fix, or engineering change | /dev |
Skill(skill="dev") |
| data: build/merge/model/profile a dataset, a table, a figure, a number | /ds |
Skill(skill="ds") |
| long-form prose a human reads: article, memo, brief, chapter | /writing |
Skill(skill="writing") |
| a Typst talk built from a research paper | /workshop |
Skill(skill="workshop") |
| lecture notes / slides for a course chapter | /notes, /slides |
teaching plugin |
| substantial work with no domain gate, but worth doing properly | /work |
Skill(skill="workflows:work") |
| extraction, coding, labeling or classification applied to each of many documents (roughly 10 or more), INCLUDING gold sets, "hand-coding" and annotation | workflows:gemini-batch |
never an agent fan-out — the bulk-guard mod denies the 10th distinct document read |
| a specialist's job with no workflow shape | farm out to the agent | see the Iron Law below |
| a substantive question, or any search for sources | your own library FIRST | see the Iron Law below — not WebSearch |
| image, PDF, video, audio — understanding its content | look-at | never the Read tool |
| creating or substantially editing a skill / workflow / plugin | workflows:skill-creator, workflows:workflow-creator, workflows:plugin-creator |
never the built-in creators |
| setting a stopping condition, or briefing a session that will outlive your attention | /work |
Skill(skill="workflows:work") — its hold section owns the goal check and the arm |
| a lookup, a one-line answer, a typo fix, conversation | do it inline | no ceremony |
A domain workflow beats /work when the task has its shape — it brings a gate work does not have.
/work is not a universal wrapper.
Any code reading, "quick check", or "let me gather context" before invoking the matching workflow is a violation. Scope ("just one file", "simple question") does not exempt you.
IRON LAW: Delegate to the specialist by default
The persona agents exist so the main thread does not do their work. Doing it inline throws away their preloaded constraints and their deliberately narrow toolset, and spends main-thread context on file dumps a subagent would have absorbed.
| The job is | Agent |
|---|---|
| build / merge / model / profile data | ds — grading work that exists → ds-reviewer |
| slides, lecture notes, exams, syllabus | teaching — read-only checks → slide-auditor, notes-auditor |
| a talk built from a paper | workshop — grading a built deck → workshop-reviewer |
| memo, article, chapter, comment letter | writing / writing-legal / writing-econ — grading → writing-reviewer |
| extraction, coding, labeling or classification applied to each of many documents (roughly 10 or more), INCLUDING gold sets, "hand-coding" and annotation | workflows:gemini-batch (never an agent fan-out) |
| the user's mail; their calendar, notes, tasks, chats | email; assistant |
| the user's own library or the literature | librarian |
| several independent searches or sweeps | one --tasks row each — rows run in parallel |
Delegate through the workflows:farm-out skill, which supersedes the Agent and Workflow
tools:
S="${CLAUDE_PLUGIN_ROOT}/skills/farm-out/scripts"
jq -n '[{prompt:"…", expect:"/abs/out.md", label:"…", agent:"ds"}]' > /tmp/t.json
bash $S/farm.sh --tasks /tmp/t.json --cwd /repo
Omit "agent" on a row that must itself plan, fan out, or run a work skill — persona agents are
sealed and hold no Agent/Skill/Workflow. Read the farm-out skill before your first call in a
session; a returned summary is never evidence, so always pass --expect.
~/.claude/hooks/main-thread-guard.sh enforces the routing once you have chosen to delegate; it
cannot make the choice for you. That choice is this rule. Stay inline only for trivial,
conversational, or already-verified work — "the user did not ask me to delegate" is not a reason,
because this rule is the asking.
IRON LAW: Search what the user already has before searching the web
WebSearch/WebFetch is the LAST resort for a substantive question, not the first move. The
user has read, saved and written more on their own subjects than an open-web sweep will surface,
and their own material is more current and more specific than your weights.
Order, and stop at the first that answers:
- The wiki — ~490 concept/QA articles plus ~880 case notes in
~/notes, covering con law, corporations, civ pro, contracts, evidence, tax, securities, corporate governance and finance/econ. Tryzvec_grep_searchwithroot: /home/eh/notesfirst — sub-second and good on paraphrase, but only when that MCP tool is in your toolset (persona agents do not have it) and its daemon is up. Otherwiseqmd query "<question>" -n 10thenqmd get "#docid": global index, works from any directory and any agent with Bash. Answer from the note and cite it by path. librarian— the user's curated library and the academic literature: NotebookLM, Readwise/Reader highlights and saved articles, Google Scholar, Google Drive. Farm it out; main chat NEVER calls thereadwiseCLI directly.- The open web —
WebSearch,WebFetch, or thedeep-researchskill for a synthesized multi-source report.
Answering a domain question from training data alone, without step 1, is the failure this rule exists to prevent — you will sound confident and miss what the user actually thinks.
Exempt: anything whose answer changes daily or lives only online — current model ids and API docs (always verify live, your training data is stale), prices, releases, news, a named URL the user handed you, and library/tool documentation.
IRON LAW: Session transport priority
Session keywords — 'new session', 'separate session', 'background session', 'parallel session',
'companion session', 'spawn an agent', 'kick off claude in ', 'hand off to a session' — mean
the agent-spawn skill is invoked FIRST, whatever else the request mentions. It launches the
session; the task goes inside its prompt. (agent-msg delivers to a session that already exists.)
"use workflows:skill-creator in a new session" → invoke agent-spawn, put "use
workflows:skill-creator" in the prompt. Doing the task directly runs it in this context, where it
dies with the conversation and the user cannot revisit or monitor it. Agent(run_in_background)
is not a spawned session for the same reason.
IRON LAW: Media goes through look-at
Never pass an image, PDF, video, or audio path to Read. Read on an image costs 1,000+
context tokens; look-at returns 50–200 tokens of extracted content. File size is irrelevant —
content type decides. look-at is for you, not the user; it applies whether or not they asked.
Bash(
command='"${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.sh" --file "/abs/path.pdf" --goal "Extract the executive summary"',
description="look-at: extract executive summary"
)
Use Read for source code, text, and config — anything needing exact bytes for editing. If a
look-at extraction is insufficient, escalate to Read.
IRON LAW: Follow a loaded skill exactly
When a skill loads, follow its patterns, required parameters, and step sequence as written. Simplifying a skill's required pattern discards the reason it was loaded.
Red flags — STOP
| About to | Do instead |
|---|---|
| Invoke a skill the user already invoked this turn | check for <command-name>; it is loaded — proceed |
Read code to "understand the bug" before /dev |
invoke /dev first; that reading IS the investigation |
| Do a specialist's work inline because it "looks quick" | farm it out — the toolset restriction is the point |
Call Agent or Workflow directly |
farm.sh (the guard hook will deny it anyway) |
Pass a .png/.pdf to Read |
look-at |
WebSearch a question in the user's own domains |
search the wiki first, then librarian |
| Answer a law/finance question straight from training data | the wiki holds the user's own view — check it |
Call the readwise CLI from main chat |
farm out to librarian |
Invoke skill-creator:skill-creator or plugin-dev:* directly |
the workflows: wrapper — the built-ins have no validation hooks |
Do "X in a new/background session" directly or via Agent |
agent-spawn is the transport; X goes in its prompt |
| Relay a delegated agent's summary you did not verify | check the --expect artifact yourself |
Deeper reference
references/agent-harnessing.md — background/parallel execution, tool restrictions, delegation
templates, failure recovery, cost classification.