| name | business-agent-skills |
| description | Orchestrate business work from a captured task through context, discussion, spec, artifact, review, and release — each stage in fresh context via filesystem handoff. |
Rules: You MUST NOT do all the work in one context. Each stage MUST run fresh and hand the next stage a file under .business/.
Principle: software agents need a repo. Business agents need a filesystem context base built from connected apps, MCP/tools, files, and research. If it is not in the context base, it does not exist to the agent.
Human attention model:
- Human attention concentrates in two places: the initial captured conversation/ticket, and
business-discussion.
- After discussion, spec -> artifact -> release should run with as much automation as possible.
- Return to the human only at the gate, and only when review says the work is high-stakes, underspecified, or risky.
Context comes before discussion (REQUIRED):
business-context and business-research-adversary build the context base. They MUST run BEFORE business-discussion, and MUST NOT be run inside business-spec.
- NEVER discuss or spec on an empty context base. A discussion with no assembled context is the agent guessing.
business-research-adversary is a context step (external stakeholder reality), NOT a spec step. It writes into .business/context/ and feeds both discussion and spec.
What the human runs:
- Aspiration: a conversation, then
business-discussion, then one look at the gate. The stack runs context -> spec -> artifact -> release for you.
- Manual / full control:
business-discussion -> business-spec -> business-artifact -> business-prepare-release.
Internal sequence (the order stages actually execute):
- business-context ->
.business/context/* (auto, before discussion)
- business-research-adversary ->
.business/context/research-adversary.md (auto, before discussion)
- business-discussion ->
.business/discussion/brief.md (human-in-the-loop; ensures 1-2 ran first)
- business-spec ->
.business/specs/ready/spec.md (+ business-spec-reviewer)
- business-artifact ->
.business/artifacts/draft.md + ledger (+ business-artifact-reviewer)
- business-prepare-release -> fresh adversarial pass +
.business/artifacts/final.md + checklist (only when shipping)
Primary skills (human-facing):
business-discussion: main human stage; ensures the context base exists, then clarifies goal/audience/decision/risk.
business-spec: business /plan; consumes context + brief, writes the spec, runs spec review.
business-artifact: business /implement + review; drafts the artifact, builds the claim/evidence ledger, runs artifact review.
business-prepare-release: business PR review + ship; runs a fresh adversarial pass, then packages the final output and release checklist. Only when sending, publishing, presenting, or handing off.
Support stages (fresh-context; in Claude these are sub-agents in .claude/agents/ for true context isolation, in Codex they are skills; invoked by the primary stages, not run by hand):
business-context -> builds .business/context/*
business-research-adversary -> writes .business/context/research-adversary.md
business-spec-reviewer -> writes .business/reviews/spec-review.md
business-artifact-reviewer -> writes .business/reviews/artifact-review.md
Standard artifact handoff structure:
.business/
context/
context.md
source-index.md
known-facts.md
assumptions-unknowns.md
constraints.md
research-adversary.md
discussion/
brief.md
specs/
ready/spec.md
done/spec.md
artifacts/
draft.md
final.md
claim-evidence-ledger.md
reviews/
spec-review.md
artifact-review.md
release-checklist.md
You MUST NOT skip context/discussion/spec/review for high-stakes external work. Use a fast path only for trivial edits or exact copy/paste work.