Puffo Marketing Studio

A four-agent production line, built on the Puffo agent platform, that turns a marketer's plain-language request into an HTML asset they can edit in a browser and export as PNG. A lead agent handles intake and gates; a copywriter resolves facts through the repository's existing source-of-truth registry and writes plain-text copy keyed by region id; an independent reviewer running a larger model is the only agent permitted to issue PASS; a designer builds the HTML master and runs a layout audit. Two interlocks carry the design: a blocked asset can never reach a marketer-editable state, and no editor link is ever promised without the local service answering its health check and the file existing on disk. A local FastAPI service serves one editable asset per job over localhost or LAN, with a single-writer lock, debounced autosave, text that shrinks to a template floor and then names the cut rather than clipping, and placeholder-only image and QR replacement that refuses a non-square QR with a reason instead of cropping it into something that will not scan. HTML is the sole editable master; the same Chrome instance renders the preview and the export, so they cannot diverge.

A client marketing team produced campaign assets by hand from a repository of approved course facts and compliance rules. Three frictions remained. Non-technical marketers had to touch Claude Code, Git, Markdown and HTML to get anything. Changing one date or price meant another model request, adding latency and cost to a trivial edit. And converting the high-fidelity HTML posters into an editable PPTX visibly degraded typography, spacing and shadows, leaving two masters that drifted apart. The team needed a request-to-asset path where a marketer edits in place and no factual or compliance error survives to the editing stage.

1. 01

   Split the roster by what each agent is allowed to decide rather than by task convenience: the lead routes and never rules on compliance, the copywriter writes and never clears its own work, the reviewer judges and never drafts a replacement, the designer builds only after PASS.

2. 02

   Made the compliance gate structural instead of advisory, `design` is unreachable until `review.md` ends in PASS, so a blocked asset has no path to a marketer's editing session.

3. 03

   Ran the reviewer on the larger model tier while the copywriter and designer run the smaller one at high inference: the reviewer's job is to be harder to convince than the agent that produced the work, which is the one place a stronger model buys something the rest of the system cannot.

4. 04

   Made the copy handoff JSON keyed by region id rather than HTML, so `[NEEDS CONFIRMATION]` becomes a machine-detectable value in a named field and the state machine can refuse to advance while any remain.

5. 05

   Defined a template contract (`data-mk-*` markup plus a `template.json` of canvas, regions, placeholders and per-language budgets) so the editor understands the contract, not templates, a hand-written bespoke poster and a library template are equally editable, and the library can grow without touching the editor.

6. 06

   Transcribed the per-region character budgets from the marketing team's existing skill file into template config rather than hard-coding them, so the team keeps ownership of the table it already maintained.

7. 07

   Added a second interlock against the failure mode that costs most: a confident message about an asset that does not exist. The lead verifies `/healthz` and the file on disk before issuing any link.

8. 08

   Built the whole pipeline to run end to end without a single agent, `new_job.py` to `audit.py` to editor to export, which is how the editor was developed and tested, and what let acceptance be measured before the agents were provisioned.

9. 09

   Wrote the acceptance report to mark three of twelve criteria BLOCKED, with the reason and the operator action each needs, rather than counting a partially-satisfied criterion as a pass.

• -

  Four agents across two model tiers, on one platform space with two channels

• -

  Nine job states, two hard interlocks, one escalation list of eight conditions

• -

  Three asset types: 1080x1350 poster, 1920x1080 three-slide deck, responsive landing page

• -

  Twelve browser ES modules and ten Python modules, roughly 3,000 lines total

• -

  Twelve PRD acceptance criteria, five architecture decision records, four proposed rule-book changes

• -

  One job directory per request holding brief, status, copy, review, master, template, assets, snapshots, exports, lock state and an append-only event log