agencykit — I turned my agency's 8 most-repeated workflows into Claude skills (free, MIT)
Hey IH,
I work at a creative agency in Rome and kept running into the same problem:
every week, the team was rewriting the same 8 things from scratch.
Project briefs. Client approval emails. Monthly reports. Quotes.
Shooting day schedules. Social content calendars. Post-mortems.
Each one taking 30-90 minutes. Each one slightly different every time.
No consistency, no knowledge transfer, no way to onboard new people fast.
So I built agencykit — a pack of Claude Agent Skills that encode these workflows.
You install it once, then trigger them with plain English:
- "new project" → full project brief with task list, phases, and risk register
- "client approval" → approval request with 48h SLA, escalation at 72h,
and automatic revision counter (flags revision 3+ as billable extra) - "monthly report" → narrative client report, not a data dump
- "quote" → line-item quote with margin check (warns below 40%)
- "shooting" → ODG, shot list, equipment checklist, crew WhatsApp message
- "content calendar" → 4-week calendar with first-week captions written out
- "post-mortem" → structured retrospective appended to LESSONS.md
- "status update" → traffic-light per project + morning briefing message
Works with Claude Code, Codex CLI, Cursor, Gemini CLI.
All bilingual EN + IT. MIT license.
Install: npx agencykit install
GitHub: https://github.com/MarcoZorn/agencykit
Landing: https://marcozorn.github.io/agencykit/
The technically interesting part: the client-approval skill maintains
revision count state across the conversation — entirely prompt-side,
no backend. It generates a different output at revision 3+, flagging it
as a billable extra. Most agency contracts have this rule, but nobody
actually counts. Now it counts automatically.
Happy to answer questions. Especially curious if anyone's building
Claude skills for other non-developer domains — feels like the ecosystem
is 90% dev tooling and 10% everything else.
Full disclosure: I'm Avery Lin (avrlin). I've been packaging a Paid Skill Fulfill Engine ($29 — Stripe/x402 verify → email/return the zip) with AI assistance, so take this as adjacent interest, not neutral advice.
Shipping the skill pack is the fun part — you did that. The leak I keep seeing once people put a price on the same zip: buyer pays, then someone still hand-DMs the download. A small fulfill layer (Stripe webhook sig verify or x402 settle → idempotent email/return bytes) closes payment→file-sent without babysitting every order.
Curious — for paid Claude/OpenClaw-style skill packs, is fulfillment still manual on your side, or is payment→delivery already wired end to end?
This is the use of AI I actually find interesting, and it's the opposite of the "AI-generated content" fight everyone keeps having. You didn't turn it into a text faucet — you handed it real jobs your agency already does: the quote skill flagging sub-40% margins, the approval skill counting revisions and marking 3+ as a billable extra. That's AI as an operator doing the work a person would, end to end, not generating filler. The revision-count-held-in-prompt-state trick is the part I'd steal; most people reach for a backend the second they need state. One genuine question: when a skill makes a judgment call — the margin warning, "this one's now billable" — do you keep a human on that gate, or does it just fire on its own?
Good question, and honestly the one I went back and forth on most. Right now it's human-gated by design — the skill flags the margin warning or the billable-revision line, but it doesn't send anything or lock in a price on its own. It surfaces the judgment call with the number/reasoning attached, and a person signs off before it goes to the client. I didn't want a false positive (e.g. miscounted revision) turning into an awkward invoice conversation. Might loosen that once I trust the counting logic more, but for now the human stays on the gate.
The real shift here isn’t workflow automation—it’s turning repeated agency knowledge into executable structure.
Most agencies don’t lose time because they lack tools, but because institutional knowledge stays implicit. Encoding recurring workflows as triggerable “skills” turns tacit operational behavior into something consistent, reusable, and transferable across teams without re-explaining it every time.
Exactly this — that's basically the thesis. The 8 workflows weren't slow because of typing, they were slow because whoever wrote the brief/report/quote had to reconstruct "how we do this here" from memory or old files every time. Turning that into a skill means the tacit knowledge only has to be right once, not re-derived by whoever's doing it this week.
Interesting.
Your reply made me think less about the skills themselves and more about what happens to an agency once its operating knowledge stops living inside people's heads.
I don't think I can do that line of reasoning justice in a thread without flattening it.
If you're open to it, what's the best email to reach you on?
Sure, marco@aconite.dev
Thanks! I’ve just sent it over.
Looking forward to hearing your thoughts whenever you have a chance.
The revision counter on the client approval skill is clever. Most agencies I know lose real money on revision 4+ because nobody tracks it at the contract level. On your question about non-dev domains: I have seen people building Claude skills for sales outreach sequencing and customer support triage. The pattern of encode-a-workflow translates well to any domain with repeatable processes. Are you planning to add a skill for proposal writing? That is another 60-minute task most agencies rewrite from scratch every time.
Yeah, revision creep is the one nobody bills for even though everyone complains about it informally. Proposal writing is a good call actually — it's on my list. Haven't built it yet because the inputs are messier (scope, past-project references, pricing logic all mixed together) vs. something like the quote skill which is more mechanical. If I do it, it'll probably need a lightweight "past proposals" reference doc to draw tone/structure from rather than generating cold. Appreciate the nudge, will bump it up.
Reference doc approach makes sense for proposals — the tone/structure variance is real. One thing that helped us was treating past proposals as example pairs (scope + outcome) rather than templates. Gives the skill more to draw from without over-indexing on a single format.
That's a better frame than templates — scope/outcome pairs give the
skill something to calibrate against without locking it into a
single structure. Going to borrow that when I build the proposal skill.