/flowmind-author — Write, Validate, Fix, Promote, Measure

SkillAI & models

Authors FlowMind .flow.yaml graphs with validate, fix, promote, and measure steps. Use when user writes or audits flows, fixes violations, or advances lifecycle quality.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the /flowmind-author — Write, Validate, Fix, Promote, Measure skill

What this skill tells your AI

The instructions your AI receives, as published by lev-os/agents in skills/flowmind-author/SKILL.md and read by ahel’s review.

Intent Router

IntentRouteWhat happens
WriteAuthoring guideDecision matrix + pattern catalog + golden rule
ValidateAudit flowDeterministic probes + domain lenses + typed findings
FixMeta-prompt KB + constraint-iterativeStructured reasoning chain per violation
PromotePromotion ladderTyped PromotionDecision with gate evidence
MeasureTelemetry analysisFlowTelemetry: tokens, cost, outcome, wall time

Write a Flow

Load the authoring guide first:

# Canonical reference
cat ~/lev/docs/design/design-flowmind-authoring-guide.md

The Golden Rule

If you can express the check as bash -c "..." && echo pass || echo fail, it is a lev.validate node. Full stop.

The Two Primitives

PrimitiveOpAI involved?
Validatelev.validateNo — shell command, exit 0/1
Execlev.execYes — prompt dispatched to adapter

Everything else is topology over these two primitives.

Required Fields

name: my-flow           # unique identifier
entry: first_node       # where execution starts

policy:
  determinism: true     # enforce reproducibility

knobs:                  # parameterize the flow
  my_param:
    type: string        # MUST have type declaration
    default: value

nodes:
  first_node:
    type: action
    description: |      # instruction-shaped, not vague
      What this node does and why.
    op: lev.validate    # or lev.exec
    ...

Flow Patterns (from authoring guide)

PatternWhenShape
Linear gate chainSequential deterministic checksA→B→C→done
Implement-validate loopIterative constraint engineeringimpl→gate→review→impl (loop)
Fan-out analysisIndependent domain lensesload→{lens1,lens2,lens3}→synthesize
Adversarial reviewRed-team a constraintverify→probe→validate→generate→report

Validate a Flow

Run the audit flow against any .flow.yaml:

# Audit all flows in the repo
lev exec "audit all flows" \
  --flow ~/.lev/flows/lev-flowmind-audit.flow.yaml

# Audit a specific directory
lev exec "audit sdlc flows" \
  --flow ~/.lev/flows/lev-flowmind-audit.flow.yaml \
  --set scan_path=plugins/sdlc/flows/

# Audit a single flow
lev exec "audit this flow" \
  --flow ~/.lev/flows/lev-flowmind-audit.flow.yaml \
  --set scan_path=path/to/my.flow.yaml

What the Audit Checks

  1. Syntax: Valid YAML, required fields, reachable nodes, terminal nodes
  2. Spec parity: op: is lev.validate or lev.exec, branch keys valid, knobs typed
  3. Patterns: Golden rule enforced, iteration bounded, prompts exist
  4. Telemetry: Terminal nodes present, descriptions are instructional
  5. Parity: Duplicate flows across plugin directories flagged

Fix a Flow

When the audit finds violations, use the remediation meta-prompt pattern:

VIOLATION: {description}
FLOW: {flow_file}
CANON RULE: {gate_id}

REASONING CHAIN:
1. Read {flow_file}
2. Identify the specific violation
3. Determine fix strategy
4. Apply fix (touch ONLY the flow file)
5. Verify with deterministic command

For project-specific constraint flows (like NAAC), the meta-prompt KB lives at: .lev/flows/prompts/naac-auditr-revival/kb-remediation-metaprompts.md

General FlowMind fixes:

  • Missing entry: → add entry field pointing to first node
  • Orphaned node → wire it into a branch path or remove it
  • lev.exec for deterministic check → change to lev.validate
  • Untyped knob → add type: declaration
  • Missing prompt file → create at path referenced by prompt: field

Promote a Flow

Flows follow the same promotion ladder as other Lev artifacts:

workshop (experimental)
  → poc (proven in one project)
  → plugin (proven across projects)
  → core (canonical)

Promotion requires a typed PromotionDecision:

  • gate_pass_ratio == 1.0 for all authoritative gates
  • blocking_findings == []
  • Evidence artifacts exist
  • Transition is on an allowed edge

Measure a Flow

After execution, check FlowTelemetry:

  • total_tokens — cost signal
  • wall_time_ms — performance signal
  • outcome — green/red/partial
  • gate_pass_ratio — quality signal

Over N runs, flows build a track record. Successful patterns get promoted; failing patterns get mutated or retired.

Related

SkillRelationship
/workLifecycle router — calls flowmind-author for flow-related work
/levloopLoop scheduler — flows are the execution unit per tick
/dump/captureCapture pipeline — proposed flows should pass audit before filing
/lev-builderPOC→core placement — uses promotion ladder
/naacNAAC operator console — naac-auditr-revival is the reference implementation

Reference Documents

DocLocationWhat
Authoring guide~/lev/docs/design/design-flowmind-authoring-guide.mdDecision matrix + patterns
FlowMind spec~/lev/docs/specs/spec-flowmind.mdBehavioral contract
Lifecycle design~/lev/docs/design/design-flowmind-lifecycle.mdObservability events
Evolutionary learning~/lev/docs/_inbox/flowmind-evolutionary-learning.mdSelf-improving flow thesis
Meta-substrate~/lev/docs/design/design-meta-substrate-declaration-compiler-ir-runtime.md4-layer doctrine
NAAC reference implnaac/.lev/flows/naac-auditr-revival.flow.yamlWorking constraint audit
NAAC meta-prompt KBnaac/.lev/flows/prompts/naac-auditr-revival/kb-remediation-metaprompts.mdRemediation templates

Signals

GitHub stars
22
Forks
2
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
flowmind-author
Source
github.com/lev-os/agents