Add seller-copilot: a consultant front door for broad and multi-part seller questions

[fiharet34]

Summary

Adds one skill, seller-copilot, with 15 references/ files.

The catalogue's 17 existing skills each do one job well. This one covers the two request
shapes they don't:

1. Vague or multi-part questions — "me ajuda a crescer", "por onde começo", "analisa
   minha operação" — where the seller doesn't yet know which analysis they need. Today these
   have no entry point.
2. Deeper analyses with no focused twin here — real margin and Mercado Livre fees,
   pricing and price wars, whether a product is worth selling, finding new products, vetting
   a supplier catalogue, import-vs-domestic, diagnosing a listing that doesn't sell,
   parameter-level competitor benchmarking, assortment and price gaps, market structure, and
   seasonality.

It classifies the request, asks at most one or two clarifying questions, runs the minimal set
of procedures from references/, and returns one consolidated verdict-first answer rather
than a pile of tables.

It routes to your skills rather than competing with them

This is deliberate and load-bearing:

• SKILL.md opens with "When to use another skill instead", naming eight existing skills
• Every routing table hands off by name where one already does the job — a single category
  snapshot to category-opportunity-index, buy-box comparison to my-product-vs-catalog,
  growth ranking to growing-leaf-category-tracker, trending terms to
  top-keywords-in-my-category
• Three reference files (category-evaluation, keyword-intel, competitor-profile) point
  at their focused twin first and explain when the reference procedure fits better instead
• The frontmatter description deliberately omits trigger phrases for capabilities that
  already have a focused skill, so it doesn't intercept their traffic

Verified in testing: of nine single-purpose questions, this skill correctly stayed out of the
way on four and the focused skill took them.

Structure

skills/seller-copilot/
├── SKILL.md              (293 lines; description 983/1024 chars)
└── references/           (15 files, ~1,200 lines)

Only the frontmatter sits in the always-loaded metadata budget. The body loads on activation;
a reference file is read only when that specific analysis runs.

This is the first skill here to use a references/ subdirectory, and CONTRIBUTING.md
describes one folder containing a single SKILL.md. Flagging it as the main thing to accept or
reject. The reason for the deviation: bundling 15 analyses costs one metadata entry and one
README row instead of 15, and 15 separate skills couldn't reference each other once each is
zipped individually for the desktop install path. If you'd rather have the flat shape,
splitting is mechanical — no external contract depends on the layout.

Evidence

Run through a real agent CLI against the live JoomPulse connector, scored by an LLM judge plus
deterministic disclosure checks, against two comparators on identical questions and criteria:
the internal predecessor this skill is derived from, and a bare prompted agent.

Nine multi-part scenarios:

	bare agent	internal predecessor	this skill
avg judge (1–5)	3.61	4.06	4.39
disclosure checks passed	—	6/9	7/9

Eight further scenarios written for the procedures the existing criteria never exercise
averaged 4.00 — or 4.67 excluding two items that measure something other than this
skill (below).

Across both rounds the scenarios surfaced eight real defects in these procedures — provenance
mislabels in both directions, a ranking padded past the data it returned, dates recalled from
the model's own knowledge presented as marketplace data, and a fallback that could not produce
the output its own rule demanded. All eight are fixed here. For the six with a measurable output
signature, the affected scenario was re-run to confirm the behaviour changed and not merely
the score; the remaining two are structural and were not separately re-measured. The commit
messages carry the detail.

The honest summary of the whole exercise: answer quality was close to the internal
predecessor's, and the repeatable difference was disclosure discipline — labelling estimates
as estimates, naming what data is unavailable instead of filling the gap, and carrying caveats
that are easy to drop. So that's what's made structural here: every reference file ends with a
mandatory Boundaries section, the estimate disclaimer is required once per answer, — is
required for missing values, and "never fabricate a number" is a hard rule.

What this evidence does not prove

• Two of the eight scenarios don't measure this skill. One scored 1/5 because the question
  routed to an existing focused skill lacking a caveat this skill's references require —
  routing is decided by frontmatter matching before this skill loads, so that score reflects
  where the question went rather than anything in these files. The other ran under a
  harness flag telling the agent no user was available to answer, which contradicts this skill's
  ask-first rule; interactively the same question behaves as intended.
• The strict tool-evidence check fails on every item and is excluded as inapplicable — the
  existing criteria expect the predecessor's tool and data-model naming.
• Three capabilities are unmeasured: supplier vetting against a real supplier list, listing
  diagnosis end-to-end, and keyword intelligence.
• Samples are small (9 and 8 items). Every score was read against the answer text rather than
  taken at face value — that's how two of the six defects above were told apart from checker
  artifacts.

Notes for reviewers

• Pure addition. No existing skill, asset or row is modified. The only change to an
  existing file is one appended README table row.
• Works standalone. Where a step names another skill, SKILL.md says to use it if
  installed and otherwise perform the equivalent analysis directly — so the skill works on its
  own via per-skill .zip install as well as alongside the full catalogue.
• Mercado Livre (Brasil) only, read-only, no real-time claims.
• Three commits, kept unsquashed on purpose: the initial skill, the six evaluation-driven
  fixes, and the handoff-caveat change. The fix history is the part that shows the skill was
  measured rather than just written.

If anything here isn't what you want, say so and I'll change it — the references/ layout,
the scope of the description, which requests route here versus to an existing skill, or any
wording in the reference files. Splitting it into flat skills, trimming it, or narrowing what it
claims are all straightforward, and I'd rather adjust to fit the catalogue's conventions than
have it land as an exception. Same goes for anything the review turns up that needs following up
elsewhere.

Public-Safety Checklist

• No credentials, tokens, keys, or secrets.
• No internal filesystem paths.
• No private Slack, Jira, Notion, admin, or service URLs.
• No private service names or implementation details.
• No customer data or private product data.
• Skill directory name matches SKILL.md frontmatter name.

scripts/check-public-safety.sh passes. No data-model or field names, tool or resource
identifiers, internal service names, paths, or customer data appear anywhere in these files —
they describe procedures and judgement only.