Stepmeter Decision Log

018 Oct

Problem selection

Choose the problem on usefulness and viability

Criteria: How many teams have the problem. Whether an outside tool can reach the data it needs. Whether a first version can reach real users in about four weeks. Whether the value can be measured.
Not used: Fit with my own past experience. It can break a tie, nothing more.

028 Oct

DroppedJira Service Management

Don't build an AI request triage agent

Why: The Service Triage Rovo agent already suggests request type, urgency and priority, and flags requests that may need escalation.

038 Oct

DroppedRovo

Don't build anything that needs Rovo agent data

Why: Admins clearly want an organisation-wide view of agent evaluations and usage (ROVO-721, ROVO-154). But there is no public API to list agents or read their usage (ROVO-516), and Rovo can't be called from Forge backend code. An agent governance view or an agent scope tester can't be built properly from outside today.

048 Oct

DroppedJira Service Management

Don't build a knowledge gap finder

Why: Rovo's knowledge base skill, rolled out from April 2026, finds topics without documentation and drafts Confluence articles from resolved requests for review.
Change: I had shortlisted this before checking. From here on, checking Atlassian's documentation and release notes for the exact capability is the first step for every idea.

058 Oct

Method

Scout widely, then apply three checks

What: About 20 ideas across Jira and Product Discovery, Confluence and Loom, Jira Service Management, and Rovo with organisation admin.
Checks: Evidence of demand (votes, Community threads, paid apps). Not already shipped by Atlassian. Data reachable through a public API or Forge module.
Shortlist: Automation step cost. An "I'm affected too" option on the customer portal (JSDCLOUD-1743, 491 votes). Idea-to-delivery status sync for Product Discovery. A broken-link checker for Confluence (CONFCLOUD-65867, 108 votes).
Set aside: Expiring page verification, approval reminders and sprint carryover tracking, where several apps already compete. CSAT insights and long-range SLA history, which Atlassian has signalled it is working on.

068 Oct

Chosen

Build a tool for automation step cost

Why: Per-step billing starts on 3 December 2026 for every cloud organisation that uses automation, with extra usage on by default. Cost depends on how each rule is built, and from the public documentation I couldn't find a per-rule cost view. The value can be measured in money.
Constraints: The Rule Management API takes an admin's API token, not Forge or OAuth, so the first version is a web tool. There is no public API for how often each rule runs, so frequency has to be estimated.
Risks: Atlassian may add a per-rule view. Estimates may be too rough to trust. Admins may be wary of sharing an API token.

078 Oct

Validation

Validate for three days before building

Checks: The API returns each rule's structure, including nested branches and conditions. What the product already shows per rule. Whether Jira admins already know which of their rules are expensive.
Stop if: The API can't return rule structure, or most admins interviewed already know their expensive rules. The next shortlisted idea takes over.

088 Oct

Security

Strip secrets before processing, keep no rule data

Why: The API returns rule configuration in full, including values the UI masks, such as webhook URLs and authorisation headers (AUTO-2207).
How: Values under secret-looking fields and any URL or credential string are masked before anything is saved. Tested on a sample rule with four planted secrets: all four removed.

098 Oct

Process

Build in public

Why: A public repository shows how the work develops over time, not just the end result.
Guardrails: API tokens stay in a local file that is never committed. Rule data from any real site, even redacted, stays out of the repository.

108 Oct

Naming

Call it Stepmeter for Jira

Why: "Step" is the unit Atlassian uses to measure automation usage, so the name says what the tool is about. Atlassian's brand guidelines for Marketplace vendors ask that app names avoid the word "Atlassian" and place the product name after a preposition, as in "for Jira".

118 Oct

Phase 0 findingFocus

Focus on wasted steps and fixes, not on ranking flows

Finding: Each space's Automation Usage tab already lists high-usage flows with runs and estimated steps. My earlier assumption, based on the public documentation, was wrong.
Gap: The Usage tab shows totals only. In the test site, one flow used 7 steps over 3 runs, and 4 of those steps went to two runs that stopped at a condition and did nothing. Nothing on the tab shows that, or what to change.
New focus: For each flow, show how much of its cost goes to runs that do nothing, explain which part of the flow causes it, and suggest a rewrite with the expected saving before anything is changed.
Evidence: My step model matched Atlassian's count exactly: 11 of 11 steps across 5 runs on 3 flows.

Decision log

Decisions