agentsclimarketplace

Master orchestrator

Skill asong56/skills/03-build/orchestration/master-orchestrator

268 AI coding assistant skills, organized across 12 workflow layers. Sources include Anthropic official, FRM, SKC, LRN, SKA, and other mainstream AI coding frameworks.

Install
npx -y skills add asong56/skills --skill master-orchestrator

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 18 days oldThe repository was created 18 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Unified engineering orchestrator. Route any build intent — bootstrapping an MVP, adding a feature, changing behavior, fixing a defect, or refining structure — to the correct internal Mode, then drive the shared Research → Plan → TDD → Deliver pipeline to a gated commit. Incorporates former: orch-build-mvp, orch-add-feature, orch-change-feature, orch-fix-defect, orch-refine-code.

SKILL.md

7.8 KB, ~2.0k tokens by cl100k_base, as published. Nobody here has run it

master-orchestrator

Why This Skill Exists

The five former orch-* skills (orch-build-mvp, orch-add-feature, orch-change-feature, orch-fix-defect, orch-refine-code) shared an identical underlying pipeline — Research → Plan → TDD → Deliver — and differed only in three settings: size floor, phase mask, and TDD first move. Per the Three-Principle Framework in the refactor spec:

  • Each skill's unique input (user intent) was the sole output of an implicit intent-classifier with no other consumer → sequential workflowMERGE.
  • Combined token budget stays well under 5 K, so no Needle-in-a-Haystack risk.

session-state, ck, and handoff remain independent (cross-cutting, shared-state, never-merge zone).


Intent Router (State Machine)

Read the user's request and classify it through this decision tree before doing anything else.

User request
    │
    ▼
Does a spec/SDD/PRD document exist as input?
    ├── YES → Mode: BUILD_MVP
    └── NO
         │
         ▼
    Does the capability exist in the codebase yet?
         ├── NO  → Mode: ADD_FEATURE
         └── YES
              │
              ▼
         Is it currently broken / producing wrong output?
              ├── YES → Mode: FIX_DEFECT
              └── NO
                   │
                   ▼
              Does the desired behavior change?
                   ├── YES → Mode: CHANGE_FEATURE
                   └── NO  → Mode: REFINE_CODE

Signals by mode:

ModeTrigger phrases
BUILD_MVP"build from this SDD/PRD/spec", "bootstrap the project", "start from this doc"
ADD_FEATURE"add", "implement", "build X" (X doesn't exist)
FIX_DEFECT"broken", "crash", "wrong output", "regression", "error", "bug"
CHANGE_FEATURE"change X to Y", "make it do Y instead", "adjust", "alter"
REFINE_CODE"refactor", "clean up", "extract", "simplify", "remove duplication"

If the intent is ambiguous, ask one clarifying question before proceeding.


Mode Settings

Once the mode is identified, apply these settings to the shared pipeline below:

SettingBUILD_MVPADD_FEATUREFIX_DEFECTCHANGE_FEATUREREFINE_CODE
Size floorlargestandardsmallsmallstandard
Phase mask0→1→2→3→4→5→60→1→2→4→5→60→(2?)→4→5→60→(1?)→2→4→5→60→2→4→5→6
TDD first moveGAN framework from specWrite new failing testsWrite regression test reproducing the bugUpdate existing tests to new spec, then implementConfirm existing tests green; add characterization tests if coverage thin
Commit prefixfeat: (per slice)feat:fix:feat:refactor:
Phase 3 (Scaffold)?✅ required❌ skip❌ skip❌ skip❌ skip
Research (Phase 1)?✅ always✅ default on❌ usually skipconditional❌ skip

Phases at a glance:

0 → Ingest (read spec / explore code)
1 → Research (external docs, libs, prior art)
2 → Plan (task_list / restructure plan)  ← GATE 1: user approves
3 → Scaffold (MVP only: stand up first vertical slice)
4 → TDD (red → green, per mode's first move)
5 → Code Review (+ security-reviewer if security trigger)
6 → Commit  ← GATE 2: user confirms

Shared Pipeline Execution

Phase 0 — Ingest

  • BUILD_MVP: Read the spec document at the given path. Extract: scope, locked decisions, feature list. Order into thin vertical slices (one end-to-end path first, not all-models-then-all-views).
  • All other modes: Run code-explorer to map the relevant module/file set.

Phase 1 — Research (conditional, see table)

  • Delegate to documentation-lookup or deep-research agent.
  • Skip if mode is FIX_DEFECT (root cause is in the code, not docs) or REFINE_CODE (no new behavior).

Phase 2 — Plan

Produce a task_list (ordered subtasks, each implementable in one TDD cycle). For BUILD_MVP, phase 2 is heavy: translate the ingested SDD into gan-FRM/spec.md + gan-FRM/eval-rubric.md.

GATE 1: Present the plan to the user. Do not proceed until approved. For FIX_DEFECT where root cause is obvious, Gate 1 may be skipped.

Phase 3 — Scaffold (BUILD_MVP only)

Stand up the first vertical slice using the GAN framework:

/gan-build "<one-line brief>" --skip-planner

Defaults: --max-iterations 15, --pass-threshold 7.0, --eval-mode playwright (use --eval-mode code-only for non-UI slices). Commit the scaffold as a separate feat: commit before iterating.

Phase 4 — TDD

Apply the mode-specific first move (see Mode Settings table), then run the RED → GREEN → REFACTOR cycle per task in the task_list. Each task is one atomic TDD cycle.

Phase 5 — Code Review

Run code-review agent on the diff. Auto-add security-reviewer if any of these triggers appear: auth, permissions, cryptography, external data ingestion, payment processing, PII.

Phase 6 — Commit

GATE 2: Show the user: diff summary, test results, review findings. Commit only after explicit confirmation.

git add -p   # stage selectively
git commit -m "<prefix>: <concise description>"

Size Classifier

When the size floor is standard or large, classify the request first. Apply phase collapse for smaller work:

SizeRule of thumbPhase behaviour
trivial< 5 lines, single functionSkip 1 and 2; go straight 0→4→5→6
smallsingle file or isolated moduleFloor for FIX_DEFECT and CHANGE_FEATURE
standardmulti-file, clear boundaryDefault floor; full 2-phase plan
largecross-cutting or multi-sliceFloor for BUILD_MVP; heavy phase 2

Examples

# BUILD_MVP
master-orchestrator: civicpulse/docs/SDD-v0.6.md
→ read SDD → slice list (vertical) → scaffold slice 1  [GATE 1]
→ /gan-build --skip-planner (generator → evaluator loop) → review
→ commit feat:  [GATE 2] → next slice

# ADD_FEATURE
master-orchestrator: add OAuth2 login to nws-poller
→ research auth libs → task_list  [GATE 1]
→ new failing tests → implement to green → code-review (+ security-reviewer)
→ commit feat:  [GATE 2]

# FIX_DEFECT
master-orchestrator: poller crashes on empty NWS response
→ regression test reproducing crash → fix to green
→ code-review → commit fix:  [GATE 2]

# CHANGE_FEATURE
master-orchestrator: alert at 2 warnings instead of 3
→ update threshold tests to new spec → change impl to green
→ code-review → commit feat:  [GATE 2]

# REFINE_CODE
master-orchestrator: extract the NWS HTTP client out of poller.py
→ confirm tests green → restructure plan  [GATE 1]
→ move in small steps (tests green throughout) → code-review
→ commit refactor:  [GATE 2]

Migration Notes

Deprecated skillReplaced byMode
orch-build-mvpmaster-orchestratorBUILD_MVP
orch-add-featuremaster-orchestratorADD_FEATURE
orch-fix-defectmaster-orchestratorFIX_DEFECT
orch-change-featuremaster-orchestratorCHANGE_FEATURE
orch-refine-codemaster-orchestratorREFINE_CODE

The underlying orch-pipeline engine is unchanged; only the five thin-wrapper files are retired.

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.