The AI development workflow

The seven-step loop Briefboard orchestrates — define, analyze, plan, scope, implement, review, deliver — with a default AI prompt for each step.

Briefboard ships a complete workflow for AI-driven development: seven steps, each backed by a default AI prompt your agent can run. Three steps happen at the milestone level, three repeat for every task, and one closes the milestone. The optional steps are safe to skip when the work is simple.

Step Level Prompt What happens
1. Define Milestone — (human) Milestone description + specs
2. Analyze (optional) Milestone Business analysis The agent questions the spec
3. Plan Milestone Technical plan The agent creates the tasks
4. Scope (optional) Task Scope the task Cross-checks the ask against your code
5. Implement Task Implement a task Code, tests, progress comments
6. Review Task Review a task Independent verdict on the change
7. Deliver Milestone Pre-delivery checklist Nothing missed before shipping

1. Define — the human step

Create the milestone (its description states the objective and scope) and write the specs the feature needs: business rules, technical constraints, a Definition of Done when the milestone's "finished" isn't obvious. A simple milestone can skip specs entirely — the task descriptions then carry the contract on their own.

2. Analyze — optional

Copy the Business analysis prompt into your agent. It reads the specs over MCP, flags ambiguities and uncovered edge cases, and asks the questions a good analyst would. Fold the answers back into the specs before planning.

3. Plan

The Technical plan prompt turns specs into tasks, created directly on the board over MCP. Each task description is the contract: what to do and why, ending with an explicit - [ ] checklist whenever the "done" condition isn't obvious.

4. Scope — optional

For a non-trivial task, the Scope the task prompt plays tech lead: it cross-references the request with your actual codebase, confirms the intent, and posts a few actionable remarks as an [AI scoping] comment. It never writes code.

5. Implement

The Implement a task prompt drives the loop you'd expect from a senior developer: mark the task in progress, implement on a dedicated branch following the description, document blockers and decisions as comments, validate against the checklist, mark it done.

6. Review

The Review a task prompt turns a fresh agent session into an independent reviewer — the fresh context is the point: the reviewer must not know what the developer was thinking. It judges the change — the pull request if there is one, otherwise the branch diff — against the task description and the specs, then posts a structured [AI review] verdict: APPROVE, REQUEST CHANGES, or ESCALATE. After three review cycles it escalates to a human instead of looping.

7. Deliver

When every task looks done, the Pre-delivery checklist prompt verifies at the milestone level that each rule in the specs is covered by a finished task, and documents any gap before you ship.

Two ways to run it

  • Step by step, from the UI — every prompt is one click away: milestone-level prompts on the milestone page, task-level prompts in the task view. Copy, paste into your agent, review the result.
  • Autonomously, over MCP — a connected agent chains the loop itself: get-next-task, get-task-prompt, implement, report, done. The full tool list is in the MCP tools guide.

Either way, the decisions that matter — what to build, when it's truly done, when to merge — stay with you.