Two weeks after the initial launch, two things shipped based
on community feedback.
The most common feedback: outputs sounded like a generic template,
not like the agency.
The fix: drop a BRAND.md
file into .claude/skills/[skill-name]/
or .claude/brand/
(global fallback). Each skill reads it before
generating and applies your tone, terminology, and forbidden phrases.
The format is simple — agency name, 5 voice example sentences,
a forbidden words list, optional client-specific overrides.
A BRAND.md.example
is in the repo root.
The ninth skill. Handles the 60-minute task most agencies
rewrite every week.
Trigger: "proposal", "write a proposal", "proposta" What it does differently: if you provide past proposals or
scope/outcome example pairs, the skill extracts your tone and
structure patterns and applies them — without copying verbatim.
Treats past proposals as example pairs (scope + outcome) rather
than templates.
Output: client-facing proposal + internal note with margin
estimate and risks. Human-gated by design — it drafts,
you sign off.
npx agencykit install
→ github.com/MarcoZorn/agencykit