This Playbook does not introduce another framework. It sequences the operating ideas already established across Signal into one method for one repeatable workflow.
One workflow. One operating method. No new doctrine.
Start with one workflow.
Do not begin by asking: Where can I add agents?
Begin with: Which repeated piece of work is stable enough to become a system?
Good candidates have recurring inputs, a recognizable output, enough repetition to justify design, failure modes you can name, a definition of good and consequences you understand.
If the work changes every time, do not automate confusion.
Define the work before AI touches it.
Write down the objective, trigger, inputs, output, constraints, definition of good and failure modes. If you cannot define the work, no agent architecture will rescue it.
Establish evidence state.
Before recommendation or action, determine what the system is allowed to believe.
| State | Meaning |
|---|---|
| Verified | Supported by current evidence. |
| Historical | True at a previous point but may not be current. |
| Unresolved | Evidence conflicts or is incomplete. |
| Hypothesis | Interpretation that still needs confirmation. |
| Approved | Reviewed state or language that should persist. |
| Expired | No longer safe to use without a fresh check. |
Do not force uncertainty into certainty for the sake of a clean output.
Assign the Action Boundary.
Use Deep Dive 003. For every action in the workflow, assign the maximum right the AI receives.
| Work item | KNOW | RECOMMEND | CHANGE | RELEASE |
|---|---|---|---|---|
| Read source material | ✓ | |||
| Draft recommendation | ✓ | ✓ | ||
| Update internal record | ✓ | ✓ | Conditional | |
| External send | ✓ | ✓ | Staged only | Human / explicit authority |
Then test each higher-right action against consequence, reversibility, uncertainty and propagation.
Turn routine work into a skill.
A skill is not just a saved prompt. It is the smallest repeatable operating unit.
Trigger
When should it run?
Required inputs
What must exist before execution?
Source rules
Which information can it trust?
Procedure
What sequence should it follow?
Constraints
What is forbidden?
Output
What exactly should it produce?
Self-check
What can it validate before handoff?
Escalation
What forces a stop?
Handoff
What must travel with the result?
Authority
Can it know, recommend, change or release?
The goal is not to preserve clever wording. The goal is to preserve how the work runs.
Build the handoff package.
Every handoff should carry the assignment, relevant source material, evidence state, decisions already made, constraints, unresolved questions, required output and next authority boundary.
Design review before scale.
Use the seven practices from Signal 021: put constraints upstream, preserve approved state, keep uncertainty visible, use deterministic checks where possible, test before scaling, escalate exceptions and measure review load.
If there is no clear answer, remove or redesign the review.
Use the Evidence-to-Action Contract at release.
| SOURCE | Can material claims be traced? |
| STATUS | What state is the work actually in? |
| SCOPE | What authority does the workflow have? |
| REVIEW | Who still owns judgment? |
| RELEASE | Has the approved action actually been committed? |
DD003 defines what the system may do. The Evidence-to-Action Contract decides whether this specific output is ready to do it.
Test small.
Run a bounded batch. Capture total items, verified items, exceptions, corrections, human interventions, rework, review time, incorrect changes and attempted boundary violations.
Then fix the system. Scale after the workflow holds.
Earn autonomy.
Autonomy can expand when output quality is stable, evidence state is reliable, exception rates are understood, failure modes are controlled, rollback works, review load is falling and the environment remains stable.
Then move one boundary. Not all of them.
A workflow may earn CHANGE rights while RELEASE remains human. That is normal.
The one-page operating sheet.
Build it for one workflow. Run it. Measure it. Then decide what deserves to become a system next.
Signal 013 — reusable operating procedure and complete handoffs.
Signal 015 — evidence state before recommendation.
Signal 019 — the Evidence-to-Action Contract.
Signal 021 — review debt and human-attention placement.
Deep Dive 003 — KNOW / RECOMMEND / CHANGE / RELEASE.