UltraFlow

AskDB-customized agentic workflow skills for high-rigor software work. UltraFlow keeps skills on, uses the ultraflow: namespace, and bakes in AskDB's trust-layer doctrine: value check, reachability, runtime probes, invocation-surface verification, TDD, two-stage review, adversarial pressure, and evidence-gated completion.

What Changed In This Fork

• Operator-facing names now match the skills the operator types.
• building was split into test-driven-development and subagent-driven-development.
• requesting-code-review is a first-class skill.
• council is v2 with mandatory reachability and frontend-wiring gates.
• adversarial-testing keeps v3's six cross-cutting hunters, subsystem dispatch map, SPEC/PLAN mode, and frontend-coherence hunter.
• audit adds the 4+1 read-only audit workflow.
• UFSD is canonical at docs/ultraflow/specs/.
• Git flow no longer auto-creates worktrees or branches.
• Token claims are conservative; rigor is prioritized over the old savings target.

Skills

Core Flow

1. using-ultraflow - conversation spine, model routing, reasoning-effort header, value check, skill routing, and AskDB doctrine.
2. brainstorming - entry point, happy path, failure-state UX, acceptance criteria, and out-of-scope before planning.
3. council - council v2 decision review with Gate A Reachability/Proof-of-Life and Gate B Frontend-Wiring/Coherence.
4. writing-plans - implementation plans with runtime probe, invocation-surface check, operator-flow acceptance, rollback/risk, and comprehension gates.
5. test-driven-development - RED/GREEN/REFACTOR, AskDB preflight, trust-layer branch, and no live LLM/API tests.
6. subagent-driven-development - plan execution with fresh implementers and mandatory spec-compliance review before code-quality review.
7. requesting-code-review - reviewer dispatch template with Critical/Important/Minor triage.
8. adversarial-testing - v3 defensive pressure in CODE, SPEC/PLAN, or TRUST mode, sharing council's two gates.
9. verification-before-completion - evidence before completion claims, with UI smoke boundary.
10. finishing-a-development-branch - branch-in-place finish flow; merge, push/PR, keep, or discard.

Supporting

• systematic-debugging - root-cause-first debugging with reachability, runtime probe, invocation-surface checks, TDD bug fix, skeptic, and journal.
• dispatching-parallel-agents - focused parallel agents for independent tasks and audit areas.
• receiving-review - verifies review feedback before implementing it.
• audit - 4+1 read-only audit every 4-5 commits or before risky continuation.

Workflow

using-ultraflow
  -> brainstorming
  -> council
  -> writing-plans
  -> test-driven-development / subagent-driven-development
  -> adversarial-testing
  -> verification-before-completion
  -> finishing-a-development-branch

Supporting:
  systematic-debugging
  dispatching-parallel-agents
  requesting-code-review
  receiving-review
  audit

UI/UX work routes through taste, impeccable, ui-ux-pro-max, or frontend-design first, then returns to UltraFlow for planning, execution, review, and verification gates.

Derived Rename Mapping

Previous UltraFlow skill	New operator-facing skill
discovery	brainstorming
planning	writing-plans
building	test-driven-development + subagent-driven-development
debugging	systematic-debugging
verification	verification-before-completion
git-flow	finishing-a-development-branch
parallel-dispatch	dispatching-parallel-agents
missing	requesting-code-review
missing	audit

Council v2 Gates

GATE A - Reachability/Proof-of-Life: a persona names the live entry path by grep/trace to the targeted code; dead code, non-existent APIs, or unreachable inputs cannot be CONFIRMED.

GATE B - Frontend-Wiring/Coherence: a persona names the frontend seam that consumes the backend change, confirms payload-shape match, and accounts for loading, error, empty, stale, and agent-editability states.

Council and adversarial testing use identical gate wording.

Adversarial v3

Every run dispatches six cross-cutting hunters: silent-wrong-answer, silent-noop/fail-open, tenant/schema isolation, concurrency/TOCTOU, payload leak, and determinism/dead-code-green. A dispatch map adds only the subsystem clusters the surface touches, plus frontend-coherence in UI CODE mode and always in SPEC/PLAN mode.

Cross-Skill State

All skills read/write UFSD at:

docs/ultraflow/specs/UFSD-YYYY-MM-DD-<feature>.md

Do not reference legacy non-specs UFSD locations or filename conventions.

Long audit outputs go to:

docs/ultraflow/audits/

Deferred bug tracking remains:

docs/superpowers/triage/deferred-bugs.md

Project Structure

ultraflow/
  .claude-plugin/
    plugin.json
    marketplace.json
  docs/
    token-budget.md
  hooks/
    hooks.json
    session-start
  skills/
    adversarial-testing/
    audit/
    brainstorming/
    council/
    dispatching-parallel-agents/
    finishing-a-development-branch/
    receiving-review/
    requesting-code-review/
    subagent-driven-development/
    systematic-debugging/
    test-driven-development/
    using-ultraflow/
    verification-before-completion/
    writing-plans/
    ufsd-template.md

Installation

git clone https://github.com/sidkush/ultraflow.git
cd ultraflow
claude plugin add .

Manual skill copy:

cp -r ultraflow/skills/* ~/.claude/skills/

Token Discipline

UltraFlow now optimizes for correctness first. Full workflows typically load more than the old savings claim when council v2, audit, adversarial v3, or trust-layer gates are active. See docs/token-budget.md for current estimates.

Contributing

Use UltraFlow's own workflow. Keep skills concise, evidence-bound, and resistant to rationalization.

License

MIT