Engineering a workflow where verification outranks confidence.

ai-squad is a structured workflow for AI-assisted development. Its current V3 build phase uses one implementer, executable verification, a fresh-eyes review, and two human checkpoints, with hooks enforcing file scope and git-write rules.

“AI can write code quickly. The harder problem is preserving intent, proving the result, and keeping the human in control of consequential decisions.”

workflow engineering · current V3 architecture

decisions.md (ai-squad)

# challenge, architectural choices, and technical outcomes

focus V3 · scope

context: scope

Turn vague requests into reviewable contracts.

I designed two connected workflows: Discovery frames uncertain ideas, while SDD turns a clear pitch into a specification, plan, scoped tasks, and implementation. Claims about an existing codebase carry source evidence, so decisions start from the repository rather than assumptions.

focus V3 · build

context: build

One implementer, one task, one executable check.

The current build phase uses a single implementer in the active session. It maps reuse first, presents Checkpoint A, implements one bounded task, and runs that task's exact Verify command until it passes. A fresh-eyes review informs Checkpoint B before a commit is allowed.

focus V3 · guards

context: guards

Move critical rules out of prompts.

Python hooks restrict writes to declared task files and block git writes until the approval seal exists. Artifact schemas reject incomplete workflow documents. These controls narrow common failure modes; they do not sandbox arbitrary shell access or guarantee correct software.

selected architecture · 1 deep dive

role / challenge / choices / outcomes

The current path from a fuzzy idea to a scoped, verified change.

My contribution: Product framing, workflow architecture, schemas, hooks, CLI packaging, and migration from the earlier dispatch model.

Editorial workflow diagram

Workflow and real hook execution

technical outcomes

connected workflows

implementer per task

human checkpoints

scoped

mechanical write guards

retrospective.md

An earlier version used a fan-out of implementation agents and reviewers. The project evidence showed substantially higher cost without a measured quality gain, so the local V3 work removed that pipeline. The public repository currently documents the earlier release; this case includes the newer local V3 evolution. The durable idea is smaller: write down intent, constrain scope mechanically, trust executable verification over a model's confidence, and make the final diff a human decision.

github ↗

Engineering a workflow where verification outranks .css-h6mtlj{color:var(--chakra-colors-brand-accent);font-weight:600;}confidence.

Engineering a workflow where verification outranks confidence.