46 identical doc-shipping loops in two days — encode the loop, don't search for a skill

Over two days a static-site project accumulated 46 content(...) commits, and each one walked the same path by hand: research → verify → write → build → dead-link check → tests → commit → push → wait for Actions → confirm the live URL → close browser task spaces. Eight to ten tool calls per document, every time. The question was whether an installable skill could compress that.

What the registry had

#

npx skills find across the recurring tasks — web research, scraping, RSS, fact-checking, link checking, GitHub Actions, Korean writing. Results with real install counts were either generic methodology prose (a journalism fact-check workflow, less strict than the project’s own “official sources only, date every figure” rule), paid-API wrappers (Parallel deep research, TypeSafe’s typed-judgment model), or tools the repo already had (checklinks.py, a weekly source-validation workflow).

One thing did fit, and it was two shell lines:

npx -y mcporter call --stdio 'uvx duckduckgo-mcp-server' search query="Gu Thai Duluth GA address" max_results=5
npx -y mcporter call --stdio 'uvx duckduckgo-mcp-server' fetch_content url="https://..."

Research had been done by opening DuckDuckGo result pages in an automated browser and scraping them — 4–5 seconds a query plus task-space bookkeeping. The CLI returns the same results instantly, runs in parallel, and fetch_content returned the body of pages that had been 403 under curl (an Akamai-fronted city site, a JS-rendered county tax page). No skill install needed; the two commands went into the project’s instructions file.

Where the time actually went

#

Reading the writing-skills guidance made the shape obvious:

Don’t create a skill for project-specific conventions (put them in your instructions file) or mechanical constraints (if it’s enforceable with a script, automate it).

The loop split cleanly along that line.

flowchart LR
  subgraph judgment["Judgment — CLAUDE.md (loaded every session)"]
    R["Research
DDG search + fetch_content
browser only for login/click"]
    V["Verify
official page only
date every figure"]
    W["Write
front matter, sources,
no markdown in YAML notes"]
  end
  subgraph mech["Mechanical — scripts/ship.sh (one command)"]
    B[build] --> L[dead-link check] --> T[tests] --> C[commit + push]
    C --> A["wait for the Actions run
matching HEAD sha"] --> U[live URL 200] --> X[close browser spaces]
  end
  R --> V --> W --> B

• scripts/ship.sh <city> "<msg>" [/path/] — the six mechanical steps, set -e, stops at the first failure. --check runs only build → links → tests.
• CLAUDE.md — the rules that need judgment, written against the mistakes actually made that week: a program described as existing when it was only proposed, aggregator prices quoted as fact, **bold** markdown pasted into a YAML note and rendered as literal asterisks, cross-city links hand-written when the build already generates them.

The bug the first run of the script exposed

#

ship.sh waited for “the most recent Actions run” and reported its success. The first real use shipped only CLAUDE.md and the script itself — and the workflow has a paths: filter (cities/**, content/**, …), so that commit never triggered a run. The script happily reported the previous cron run’s success as if it belonged to this commit.

# wrong: whichever run is newest
RID=$(gh run list --workflow daily.yml --limit 1 --json databaseId -q '.[0].databaseId')

# right: the run whose head_sha is this commit; if none appears, say so and skip the wait
SHA=$(git rev-parse HEAD)
RID=$(gh run list --workflow daily.yml --limit 5 --json databaseId,headSha \
      -q ".[] | select(.headSha==\"$SHA\") | .databaseId" | head -1)
[ -z "$RID" ] && echo "this commit did not trigger the deploy workflow (paths filter) — skipping"

A deploy check that can pass on someone else’s run is worse than no check.

Testing an instructions file the cheap way

#

The writing-skills rule is no skill without a failing test first. For a conventions file the baseline failures were already on record (the week’s mistakes). The green check was a single-shot Haiku subagent given only CLAUDE.md and six retrieval questions — file path for a new doc, required front-matter keys, what to do with a Yelp opening-hours value, whether to hand-link another city’s site, whether markdown is allowed in a YAML detail, the ship command. Six for six, twenty seconds. Enough to know the file answers what it needs to before it gets loaded into every future session.

Two smaller fixes that came out of the same pass

#

• Cross-city links, first mention per section not per document. “Michigan” mentioned in passing in an early section had been taking the one link, leaving the tax-comparison table further down unlinked. Splitting on <h2> fixed it.
• Link to the same document on the other city’s site, not its home page. “The consulate is closer than in Michigan or Texas” now lands on those cities’ community pages — where the consulate section actually is — falling back to home only when the slug doesn’t exist there.

Takeaway

#

When a workflow repeats dozens of times, the skill registry is the wrong first stop. Sort the steps: the ones a script can enforce go in a script; the ones that need judgment go in the instructions file that loads every session; the one external tool that helped was a two-line CLI, not a skill. And test the script on a commit that doesn’t trigger your deploy — that’s where the false positive was hiding.