Safe AI Dlc

SAFe x AI-DLC Skill

TEMPLATE: This skill uses {{PLACEHOLDER}} tokens. Replace with your project values before use.

Purpose

Encode the SAFe x AI-DLC fusion so agents structure multi-issue programs consistently: SAFe supplies the hierarchy and guardrails, AI-DLC supplies the cadence and the human checkpoint. Inside a program that adopts this fusion, sprints give way to Bolts, because agent teams elaborate, build, and verify in hours rather than weeks. Adopting the fusion is a per-program choice; the standard sprint path stays valid, and single isolated tickets route to safe-workflow instead.

When This Skill Applies

• Turning an audit, epic, or initiative into an executable program
• Structuring work in the tracker as initiative, projects, milestones, issues, and sub-issues
• Breaking an initiative into Units of Work and sequencing them into Bolts
• Wiring dependencies so gates land before the fixes they verify
• Running a Bolt: elaborate, validate with the human, construct, verify gates are real, evidence, merge

Do NOT use it for a single isolated ticket with no program context — use safe-workflow for that.

The Fusion Model

SAFe gives hierarchy, WSJF, roles, DoD, and dependencies. AI-DLC gives cadence (Bolts) and the human-in-the-loop.

The loop: AI plans, AI asks clarifying questions, the HUMAN validates business context, then AI executes. Never skip the human validation step.

How to Structure the Program

Build top-down:

1. Initiative = the program. One per program.
2. Projects = Units of Work. One coherent outcome each.
3. Milestones = the three AI-DLC phases (Inception, Construction, Operations). Create all three per project.
4. Issues = Stories. Each carries the issue template below. Label with bolt:N and agent:* (lead role).
5. Sub-issues = Mob Elaboration tasks. Create during Inception, not before.
6. Dependencies = blocks / blockedBy edges forming the dependency DAG.

Program streams map to sub-initiatives where the tracker plan supports them; where it does not, encode the stream as project priority instead (highest-risk stream = Urgent/High).

Use linear-sop for tracker operations, label prerequisites, and evidence templates.

Issue Template

Every issue description block contains:

Header:      <ID> — <title>  [{{TICKET_PREFIX}}-XXX]
WSJF:        <score>   MoSCoW: <Must|Should|Could|Won't>
Finding:     <source refs, e.g. audit finding IDs>
Phase:       <Inception | Construction | Operations>
Bolt:        <bolt:N>
Role:        <agent:* lead>
AC:          - acceptance criterion (testable)
DoD:         - merged to {{MAIN_BRANCH}}; relevant gate REAL and green; evidence in tracker;
               no silent suppression
Deps:        blocks: [...]   blockedBy: [...]
Sub-issues:  (Mob Elaboration tasks, added during Inception)

Dependency-Wiring Rules

Wire blocks / blockedBy so the enforcement lands before the thing it enforces:

• Fix the gate before the fix it verifies. Harden CI or un-mask a check before landing the fix that check is supposed to catch.
• Rotate before scrub. Rotate a leaked credential before scrubbing it from history.
• Capture before alert. Error capture or health check blocks the alerter and the poller built on top of it.
• Upgrades before re-baseline. Framework and SDK upgrades block audit re-baselining, which blocks the scheduled recurring audit.

Dependency chains are usually rooted at CI hardening — one keystone issue tends to unblock a whole stream. Identify it and sequence its Bolt first.

Running a Bolt

1. Open and elaborate every issue in the Bolt (Mob Elaboration: sub-issues, unknowns, questions).
2. Validate with the human on all business, security, and policy decisions before Construction.
3. Construct in dependency order on {{TICKET_PREFIX}}- branches with SAFe commits, rebase-first.
4. Verify gates are real — a green check that still masks a failure is not an exit.
5. Post evidence to the tracker (dev/staging/done) per linear-sop.
6. Merge — RTE prepares the PR; the human (HITL) merges. Exit = merged + real gates green + evidence posted.

Human-Gate Checklist

Route to the human, with options and a recommendation, whenever a decision is:

• A secret or credential action (what to rotate, when, blast radius)
• A security policy: CSP allow-list, auth behavior, headers
• A CI or branch-protection decision: which checks become required
• A severity or risk policy: dependency-audit thresholds, fix-vs-suppress budget
• An SOP exception (e.g. a migration deviating from the documented process)
• Signing the Definition of Done

If a decision changes business, security, or risk posture, it is the human's — surface, recommend, stop.

Anti-Patterns (Do NOT use)

Forcing an ambiguous epic into one Bolt    (run a spike instead — elaboration is the hard part)
Creating sub-issues before Inception       (Mob Elaboration produces them, not the other way round)
Bolts as calendar sprints                  (a Bolt exits on evidence, not on a date)
Skipping human validation to move faster   (the loop is the method; without it this is just SAFe)
Treating a green check as an exit          (verify the gate enforces, not that it merely passed)

Reference

• docs/guides/SAFE-AI-DLC-METHODOLOGY.md -- Methodology guide (rationale, worked example)
• specs_templates/program_template.md -- Program document scaffolding
• AI-DLC source: https://aws.amazon.com/blogs/devops/ai-driven-development-life-cycle/

Routes To

• linear-sop — tracker operations, program structure, evidence templates
• safe-workflow — branch, commit, and PR conventions; HITL merge
• orchestration-patterns — multi-agent swarm coordination for a Bolt
• pattern-discovery — find existing code and doc patterns before constructing