For leadership ยท How our AI engineering team works

a team of specialist bots. Two leads.
The owner approves every gate.

Every feature request or bug report enters through the Captain. Specialist bots work in parallel where it is safe, share one cache instead of re-reading raw documents, and stop at six human approval gates. Click any bot to see what it does.

Feature flowBug flowHuman gateFix loopShared cache
โœ‹ Humans own the decisionsBots stop at six gates: open questions, plan and ticket drafts, repo state, review fixes, QA verdict and posting to the tracker.
๐Ÿ”’ No bot commits or pushesThe owner does all git work. Git Bot only advises on branches, conflicts and how to group files per task.
โšก Read once, reuse everywhereAnalysts read the tracker, BRS, design and code once and write summaries to the shared cache. Every later bot reads the cache.
๐Ÿ“ฃ Nothing external without a yesTracker tickets, comments and status changes are drafted first and only posted after the owner approves.
๐Ÿšฆ Analyzer gateAnalyze clean and related tests green before review. fvm commands only.
๐Ÿงช Light TDDPlanner lists test cases. Devs write code, then tests, plus tests for any extra function.
โœ… Quality firstPlanner writes the must-haves. Devs self-check before handoff, so review mostly confirms.
๐Ÿ’ฐ Token savingRead summary files and related files only. Report only failures.
๐Ÿง  Heavy modelOnly Gap Finder and Dev Planner. Everyone else on auto.
๐Ÿ“„ QA reportOne template with section toggles, including failed API requests in the success flow.

Handoff chainWho does what, then what happens next

#AgentDoesWritesNext
1CaptainReads the input (task key, BRS, design, request) and classifies itTask folderA "?" gets an answer only. Otherwise Intake starts
2Intake analystsBRS, tracker, design and code read once, in parallelbrief, tracker, design, code .mdThe owner approves, then Gap Finder starts
3Gap FinderFinds gaps and blocker questionsgap.mdThe owner answers, then Dev Planner starts
4Dev Planner + Task DrafterPlan, test cases and ticket draftsplan.md, tracker.mdThe owner approves, or answers go back to the planner
5Git BotRepo-ready check (advice only)NoteThe owner readies the branch, then devs start
6Flutter Devs + QA Test WriterCode plus tests, and QA casesCode, qa.md casesAnalyzer gate
7Analyzer gatefvm flutter analyze and related testsResultPass goes to Review. Fail goes back to devs
8Review botsPR Reviewer, Code Quality, Security and Localization, in parallelreview.md, quality.mdThe owner picks fixes. Comments go back to devs (loop to 6)
9QA Lead + testersMulti-language tests on devicesqa/ evidenceFail means rework (loop to 6). Pass goes to the report
10QA ReporterFills the report template (HTML or PDF)Final reportThe owner accepts the verdict
11Tracker UpdaterPosts the comment and statusTrackerThe owner says yes, then tracker Done
12Cache KeeperSaves repeat issues and syncs rulesrules, must-havesNext task starts smarter

Two sidesGrok bots vs Cursor sub-agents

๐Ÿค– Grok bots

  • Chat teammates with their own memory, in group chats
  • Run on the owner's command, and Captain orchestrates
  • Skills: the shared library (flutter-captain, flutter-clean-feature, flutter-legacy-edit, flutter-responsive-ui, flutter-reusable-widget, flutter-feature-review, <money-feature-module>, coworker-engineer)
  • Store: bot memory (preferences, rules summary) and task files in <AI-ROOT>/<App-Name>/tasks/<TASK-KEY>/

๐Ÿ–ฑ๏ธ Cursor sub-agents

  • Files: <HOME>/.cursor/agents/<prefix>-*.md
  • Started with /<prefix>-captain <TASK-KEY> inside Cursor on the repo
  • Skills: the same skills, copied to <HOME>/.cursor/skills and <REPO>/.cursor/skills
  • Rules: <REPO>/.cursor/rules/*.mdc
  • Store: the same task files

๐Ÿงฉ Which agents, skills and rules are used

  1. The project's own come first: <REPO>/.cursor/agents, skills, rules and AGENTS.md
  2. Ours fill the gaps, combined with the project's
  3. If the project has none, ours are used (global)
  4. On a conflict the project wins, except safety rules (secrets, git by the owner, approvals, test accounts), which always stay ours
FileWhatWhy
PROJECT_CONTEXT.mdApp overviewRead once
project/rules.mdTeam rulesSingle source
project/MEMORY.mdDated decisionsNo repeat discussions
project/quality-rules.mdCode quality rulesLocal quality gate
project/quality-must-haves.mdPre-dev checklistRight on the first pass
project/devices.mdQA devicesSame devices every run
project/app-internals.mdApp internalsNo re-deriving
project/test-accounts.mdTest loginsLocal only, never shared
tasks/<KEY>/ brief, gap, plan, quality, review, qa .mdStage summariesEach stage reads a short file to save tokens
qa/Screenshots, videos, reportsQA evidence
_shared/qa-kit/Tools, report templateReused by every app

๐Ÿ”„ Cache Keeper keeps both sides in sync, so Grok and Cursor follow the same rules and files.

Fintech vs TelcoWhat changes in the flow

๐Ÿ’ธ Fintech (money features)

Transfers, wallet, payments, history

  • Stricter Security: PIN, OTP, amounts, no PII in logs, idempotent requests
  • Extra Gap checks: limits, fees, balances, failure and reversal, duplicate tap
  • QA error flows: timeout mid-payment, insufficient balance, one wrong PIN
  • Test accounts local only
  • Skill: <money-feature-module>
  • Stricter code quality critical issues
  • A failed API in the success flow is a blocker

๐Ÿ“ถ Telco

Plans, bundles, balance, profile, remote config

  • Focus on backend config and remote config data
  • TTL caching of plans and bundles
  • Local language text
  • Sizer and legacy UI
  • Analytics events
  • Offline and empty states

Shared by both: approval gates, fvm only, light TDD, code quality grade B+ with 80% coverage on changed code, pixel QA against the design, and git done only by the owner.

๐Ÿ“ Save paths

  • <AI-ROOT>/<App-Name>/PROJECT_CONTEXT.md
  • <AI-ROOT>/<App-Name>/project/ rules.md, MEMORY.md, test accounts (local only)
  • <AI-ROOT>/<App-Name>/tasks/<TASK-KEY>/ stage .md files
  • <AI-ROOT>/_shared/qa-kit/report-template/
  • <REPO>/.cursor/ and ~/.cursor/agents

Based on the team's own playbooks (team-builder skill and pipeline docs). Optional bots such as the Crash, Analytics and Localization checks can be switched off per project. This page shows how work flows. It contains no performance figures.