Six things we build, and where each one ran.
Each one has real work behind it, and each comes with what it did. We usually suggest starting with two or three.
Connect
Claude, working inside your JobAdder.
A connector lets Claude do the work in your database instead of stopping at copy and paste. Your consultants ask in plain words. Claude searches, reads, writes the note and moves the application along, attributed to the person who asked.
We built our own, NativeMCP, and run it as a managed service. Every consultant signs in to JobAdder as themselves. Your workspace is pinned to your JobAdder account at the server. Deleting records is off unless you ask for it.
Under the hood: MCP, the standard way to plug tools into Claude, talking to JobAdder's official API from our own servers in Australia.
Where it ran
- 174 toolsCandidates, jobs, pipeline, clients, contacts, notes, tasks and reporting.
- 8,036 requests in a weekSeven of seven people in the office using it.
- Nine fixes from live useIn the first fortnight, from what the team ran into.
Automate
Routines and skills for the work you do every day.
When someone does the same thing by hand two weeks running, it's time to save it. A skill teaches Claude how your agency does a job: which candidates make a call list, how a call note is written, what a LinkedIn approach sounds like. A routine goes further and runs the heavy part on the server.
Rules are saved for each consultant, and they change them by asking Claude. No list of names to keep, no spreadsheet to paste.
Under the hood: each consultant's saved rules, and a weekly folder in JobAdder so nobody comes up twice.
Where it ran
- About 1,000 lookups to 1 requestA consultant's daily call list, moved onto the server.
- Four weeks without repeatsEach list is saved to a weekly folder, and the next run skips them.
- Two saved skillsThe daily call list and logging calls.
Train
A handbook your team will actually open.
Training fades. A handbook stays. One written in your agency's words, about your real work, is still there on Monday. Ours start with signing in and a first real task, cover asking well and working in projects, and end with what to do when something looks wrong.
Every connector tool is listed, filterable by area and by whether it reads or changes data, so nobody has to wonder what Claude is allowed to touch.
Under the hood: simple pages in your brand, with a search that understands your team's own words for things.
Where it ran
- 16 articles in four partsStart, Claude, the JobAdder connector and help.
- A reference to all 174 toolsEach one marked as reading or changing data.
- A training deckWith presenter notes built in.
Measure
A weekly report that tells you the truth.
Every request through the connector is logged, so you can see who is using it, for what, and where they get stuck. Each week the owners get a short written report: the headline, each person's numbers, what changed and what we're doing about it.
When the news is bad, the report says so first. A team version goes to everyone, with totals, wins and one thing to try, and nobody ranked.
Under the hood: the connector's record of every request, read by a person each week. Not a dashboard nobody opens.
Where it ran
- Six weekly reportsIn a row, and still going.
- Two people flagged, two came goodWeek one flagged one person fading and one not started. Both were regulars by week two.
- 93% finding and readingThe finding that led to the team report and its weekly ‘one thing to try’.
Clean
Data you can trust before you automate it.
AI on a messy database finds the mess faster. So we sample yours first, through the connector, from the newest records, the middle and the oldest, and put real numbers on what's wrong.
Then rules, then a clean-up. Rules so Active means something again. A client clean-up for duplicates, head office and branch links, and consistent locations. And a small website where old candidates update their own details, with their CV read by Claude and checked before anything goes into JobAdder. At the agency on our Work page, these are scoped and quoted; the analysis is done.
Under the hood: a look through the connector, written rules, and Claude reading CVs, with every change checked before it lands.
Where it ran
- 33,820 candidates sampled91% marked Active, one created in 1999 and still Active.
- 4,347 client records analysedDuplicates, head office and branch links, and locations, record by record.
- Rules before clean-upScoped and quoted. The full candidate clean-up waits for agreed rules.
Publish
Websites rebuilt from research.
We build recruitment websites too, through SMB Labs, our web studio. The useful part is the research: who you're up against, what their sites promise, and what employers and candidates actually say. The site is then built, or rebuilt, around it, from your own brand.
Where it ran
- specpro.com.auBuilt in a day, researched the next, then rebuilt around the findings.
- 54 agencies, 15 websites pulled apart, 86 quotesFrom Australia and overseas, and from employers and candidates.
- Automated accessibility checks on every pageWCAG 2.2 AA, at phone and desktop widths, no issues found.
The workbench behind every job.
The same tools and checks on every engagement, so every job gets the same checks.
| Stage | What we use | What it catches |
|---|---|---|
| Look | Live sampling through the connector. Sessions on the desk with the people who do the work. | The status that stopped meaning anything. The workaround nobody wrote down. |
| Research | Research into your competitors and what their customers say about them. | What the market already promises, and what customers actually complain about. |
| Build | Claude, with our own checks for design, writing and safe changes. Our tooling runs on our own servers in Australia. | Generic, template work. Changes that break something nearby. |
| Check | Automated accessibility audits on every page, screenshots at phone, tablet and desktop widths, a search review, and a fresh reviewer that hasn't seen the build. | Contrast, pages you can't use without a mouse, buttons too small to tap, layouts that break at the widths people actually use. |
| Measure | The connector's own log of every request, read each week by a person. | Adoption that looks broad but is one person. Faults that look like answers. |