Inception
Requirements, user stories, and workflow planning turn a raw request into an approved plan. Nothing generates until that plan carries a human signature.
SpecForge · The method behind the work
SpecForge is the published method behind every engagement I run.
SpecForge fuses GitHub Spec Kit's executable specifications with AWS AI-DLC governance, then enforces that fusion with hooks instead of trust. A spec and a plan need a human signature before an agent may touch a file. Nothing ships until that signature exists.
What it fuses
Two proven systems, run as one enforced workflow
Spec Kit gives the method its artifacts. AI-DLC gives it its discipline. SpecForge is what happens when both run on every engagement, not just the ones where someone remembers to be careful.
From GitHub Spec Kit
- Executable specifications an agent can act on without guessing
- Acceptance criteria written before generation starts
- Versioned intent as the source of truth, not a chat thread
From AWS AI-DLC
- A managed lifecycle instead of an unsupervised chat session
- Phase gates that require a decision before work continues
- A governance layer built for teams, not solo experiments
The gate
Approval is enforced, not requested
Claude Code runs a PreToolUse hook on every session. Before an agent can edit a file, the hook checks whether the relevant spec and plan already carry human approval. If they do not, the edit is blocked at the tool layer, before a single line changes. That removes the failure mode where an agent runs ahead of review because nobody happened to be watching closely enough. The gate does not depend on anyone remembering to look.
The lifecycle
Three phases, one accountable owner each
Every engagement moves through the same three phases, in order, with a human decision gating the move to the next one.
Construction
The code gets written here. AIDLC, my eight-phase build engine, runs inside this phase and carries every change through eval before it ships.
Read the AIDLC deep diveOperations
Deployment, monitoring, and the audit trail carry the engagement past launch into a running system your team owns and can operate without me.
Autonomous mode
The Ralph loop stays inside the plan
Once a plan carries approval, SpecForge can run in an autonomous mode I call the Ralph loop. The agent works through the approved steps without a human confirming each one, but it never gets to invent new steps along the way. Scope stays fixed to what the plan already said yes to. Autonomy buys speed inside the boundary. It does not buy a wider boundary.
Built for real codebases
Brownfield gets read before it gets touched
Most engagements start inside an existing codebase, not a blank folder. SpecForge runs a reverse-engineering pass first, documenting the architecture and the business logic already in production, before any spec gets written against it. Every decision after that point, every approval, every generated diff, lands in a committed audit trail alongside the code. The record ships with the repository. It does not live in a chat log that disappears when the session closes.
Work with the method
Bring SpecForge to your next build.
Send a one-paragraph brief and get a written read on how SpecForge would run on it, or start smaller with a fixed-scope pilot first.