Small slices
One implementation worktree represents one understandable slice, with explicit scope and non-goals.
Agentic engineering, with a human hand on the wheel
Build small.
Resume clearly.
Ship with proof.
Sillage is a skills-first protocol: one implicit router and small Markdown-friendly skills for intent, architecture, testing, interfaces, security, decisions, evidence, review, and handoff — without privileging a language or framework.
Product decisions stay human-owned.
One task, one trailLIVE MODEL
01INTAKEBound
02INVESTIGATEKnown
03DECIDEHuman
04IMPLEMENTSmall
05VERIFYProof
06REVIEWSeparate
07HANDOFFReady
The problem
The next agent should not need your chat history.
Agent work gets brittle when context is large, the slice is vague, and completion means “the code looks right.” Sillage keeps the useful state visible, routes the next skill from ordinary language, and leaves the rest disposable.
Fresh starts
PRODUCT, DESIGN, AGENTS, domain language, and a task record make a new context useful immediately.
Evidence first
Tests, runtime checks, screenshots, and domain observations sit beside the criterion they prove.
Human gates
Scope changes, destructive actions, waivers, merges, deployments, and external writes never become assumptions.
The mental model
A trail, not a dashboard.
Every stage leaves a small artifact or a clear reason to loop back. Operational history stays with the task; durable knowledge stays in its canonical home.
What are we trying to change?
Outcome, scope, non-goals, risks.
Where are we starting?
Project language, current task, active slice.
What did the human approve?
Choice, boundaries, digest.
What is the smallest path?
One slice, one worktree, one risk owner.
What changed?
Focused code and focused tests.
What can we observe?
Deterministic checks and fresh evidence.
Is the claim actually sound?
Independent findings, no hidden blockers.
Can someone resume?
Outcome, limits, next safe action.
Designed for both sides
Calm for humans. Concrete for agents.
Sillage is not an autonomous project manager. It is a shared protocol that makes recommendations legible and consequential actions deliberate.
For the human
See what is known, what is missing, and where your decision is required.
- Approve intent and scope.
- Accept or reject evidence waivers.
- Own merge, release, deployment, and communication.
For the agent
Get enough context to take one safe next action without inventing project history.
- Read the project contract and domain language.
- Keep one slice active at a time.
- Never call code inspection “done” without proof.