What it is

The problem

Ticmint's blog attracted people searching for events to attend, not the event organisers who buy ticketing software. Most SaaS content drifts toward the wrong reader. The agent encodes the audience into every decision: keyword clusters, tone, calls to action and what counts as a win are all wired to organiser conversion, not raw traffic.

How it works

1. Wake upGitHub Actions runs the schedule: a daily pulse on weekdays, a full audit every Monday and on the first of the month. A topic typed into Notion starts the content pipeline.
2. RememberThe memory agent reads what worked, what did not and what competitors are doing, so every later agent starts informed.
3. ResearchThe audit agent crawls the site through DataForSEO and Search Console. The competitor agent finds keyword gaps. The keyword agent builds clusters and skips terms that already rank, to avoid cannibalisation.
4. Brief, then waitThe brief agent drafts the structure, tone and internal links and posts it to Notion. Nothing is written until a human approves it.
5. Write and optimiseThe writing agent drafts the post. The on-page agent writes the meta, schema and focus keyword straight into the SEO plugin by API.
6. Draft, then wait againThe WordPress agent creates a draft and pings Slack. A human reviews and publishes. The agent never publishes on its own.
7. Report and learnThe report agent logs to Notion and Sheets and sends a Slack summary. The memory agent writes the learnings back.

Architecture

Legend

• Context passes on
• Human gate

Trigger

GitHub Actions scheduleSeven modes

Manual topic in Notion

Orchestrator

Orchestrator

• Mode routing
• Error isolation
• Shared context object

Research

MemoryReads before the run

AuditSite crawl and Search Console

CompetitorRankings and gaps

Keyword researchClusters, no cannibalisation

Create

BriefHuman approval first

WritingTone set by audience

On-pageMeta, schema, Open Graph

Publish and learn

WordPress draftHuman review before publish

ReportNotion, Sheets, Slack

Memory written backRead at the start of the next run

Impact

It is in early beta. 94% of the 625 changed URLs were indexed after the manual runs. The image agent and the first scheduled production run are still on the build list.

Built with

Decisions that mattered

• Code first, not a no-code canvas. It is portable, Claude-native and has no vendor lock-in.
• Two human gates and no way around them: brief approval and draft review.
• Credentials live in environment files and GitHub Secrets and are never committed.
• Every Ticmint-specific value sits in client-config.json, so the logic never names a client.

What is next

• Finish the image agent.
• Add GitHub Secrets and run the first scheduled production run.