New Loop Skill for Claude

new-loop — spin up a new loop

A loop (a domain) is a recurring thread of work the agent owns: a charter, a cadence, and

the artifacts it produces. This skill creates one, proves it works with a single real run, and

leaves behind a domains/<loop>/README.md that is the loop's live state.

When to use

The user wants to stand up a new workstream/beat/job (e.g. "a weekly SEO loop", "a support

triage loop", "a competitor-watch loop"). Don't use this for a one-off task — that's a backlog

line in an existing domain, or a doc/signal.

Inputs to gather (ask only what's missing)

Infer from the request; ask a short clarifying round only for what you can't:

• name — kebab-case, the loop's home folder (domains/<name>/). Keep it short.
• goal — one line: the outcome this loop drives.
• cadence — manual / daily / weekly / a cron expr. Default manual.
• what it does — what it consumes (signals? data? an inbox? a URL?) and produces (signals?

docs? a report? code changes shipped via /verify?).

• tools/data — sources or credentials it needs (point at a setup skill or .env; never

inline secrets).

If the request is already specific, infer all five and just confirm in your summary.

Procedure

1. Bootstrap the substrate (one-time; skip if already set up)

Check the knowledge-base repo root for:

• ARCHITECTURE.md and LOG.md, and
• a CLAUDE.md that has a "Knowledge base" section.

All present → the substrate exists; skip to Step 2.

Anything missing → read references/KNOWLEDGE_SETUP.md and follow it — it copies in

ARCHITECTURE.md + LOG.md, creates signals/ docs/ domains/ with their README schemas, and

injects the knowledge-base section into CLAUDE.md (or scaffolds one from

references/CLAUDE.template.md). It's idempotent: it only creates what's missing.

(Read references/ARCHITECTURE.md once if you haven't — it's the model this skill instantiates.)

2. Scaffold the loop README

Create domains/<name>/README.md from the domain template (in domains/README.md, also

quoted in references/KNOWLEDGE_SETUP.md), filled with the gathered inputs. Required sections:

frontmatter (kind: domain, domain, status: active, goal, cadence), a 2–4 line

description, ## Current focus, ## Backlog (to-dos inline — they stay in the README until they

earn a task kind), and an empty ## Timeline. Add ## Evidence & analysis / ## Metrics

placeholders if relevant.

Collision check: if domains/<name>/ already exists, stop and ask whether to update it instead

of overwriting.

3. Do ONE real test run

The point of the skill: prove the loop actually runs, not just that the folder exists.

Actually run the loop once, at small scale — do whatever it's meant to do (triage a few real

tickets, pull one real SERP, fetch the inbox, draft one comment, run one analysis query, scope

one code change…). Use real tools/data where you can; if a credential is missing, do the

furthest-reachable dry run and note the gap.

Producing an artifact is optional — a legit run may surface nothing worth filing. Only create

a signal/doc if the run genuinely produced one.

Two required outputs regardless:

• Append one dated line to the loop README's ## Timeline:

YYYY-MM-DD | test run — <what you did and found / "nothing actionable yet">.

• Append one entry to LOG.md (its grammar):

  ## YYYY-MM-DD · <loop-name> loop created + first run · #ops
  What: <one line — what the loop is and what the first run did/found>.
  Refs: domains/<name>/README.md (new)[, any artifact created].

4. Report back

Summarize: the loop's charter (the five inputs), what the test run did/found, any artifacts

created (or "none — nothing actionable this run"), missing tools/credentials to wire up, and how

to run it again (cadence + entry point). Keep it tight.

Notes

• Don't gold-plate the scaffold. A loop README is live state, not a spec — start lean; let it

accrete via its Timeline.

• One loop = one separable workstream. If what the user described is really part of an

existing loop, add it there (a backlog line + a domain: tag) instead of a near-duplicate.

• For loops that ship code, the loop's run works in an isolated git worktree and ships via

the /verify skill (proof + PR), giving each parallel agent its own isolated stack via

crabbox-setup (sibling harness skills in this plugin).

Point the README's Backlog at them.