Start Meta

Meta — Cross-stack orchestrator. The top-level entry point when you don't know where to start.

Core Job: read project state, parse your ask, route to the right stack-orchestrator OR the right meta-skill.

Core Question: "Is this a domain task (research / marketing / product) or a process task (scope / debate / decompose / review)?"

This skill does NOT execute work. It is a router. The actual work is done by the skill it routes you to (which may itself be a router, like /start-research).

---

When To Use

• You just installed the full agent-skills stack and don't know what to type.
• Your ask doesn't clearly belong to one domain ("I want to launch a new product feature" — could be research, product, marketing, or all three).
• You need a process skill: scope something with discover, debate a decision with agents-panel, decompose work with task-breakdown, review work with fresh-eyes.
• You want a quick read of "what's been done across the whole project."

When NOT To Use

• You already know your domain — go straight to /start-research, /start-marketing, or /start-product.
• You already know your skill — invoke it directly.

---

How It Works

1. Cross-stack state detection — silently read research/, brand/, architecture/, .agents/, and .agents/experience/*.md to build a picture of the whole project.
2. Domain classification — parse the user's ask. Classify as: research / marketing / product / cross-stack / process.
3. Routing decision — either defer to a stack-orchestrator (/start-X) or propose a specific meta-skill.
4. User confirmation — print hand-off command. Never auto-invoke.

---

Step 1: Cross-Stack State Detection

Read .agents/manifest.json first — it's the canonical state index. Single file, all artifact metadata in one parse. If missing or clearly stale (check updated_at), regenerate it:

bun ${SKILLS_ROOT:-.claude/skills}/meta-skills/scripts/manifest-sync.ts

Status-aware lookup: for each artifact entry in manifest.artifacts, read status and stale to qualify the state map:

Manifest signal	State map value
status: done, stale: false	✅ done
status: done_with_concerns	⚠️ done-with-concerns — surface the concern in routing output
status: blocked or needs_context	treat as missing
stale: true	✅ done (stale) — propose refresh as an option, don't block
frontmatter_present: false	✅ done (legacy, no frontmatter) — quality unknown, suggest refresh

Experience block: manifest.experience tracks .agents/experience/{domain}.md files separately. The entries count per domain is a heuristic for "how much context has been gathered" — a domain with 7 entries is well-covered; one with 1 entry barely is.

See ../../references/manifest-spec.md for the full contract.

Path reference / filesystem fallback — used only when .agents/manifest.json doesn't exist (fresh project) or sync hasn't been run:

Path	What it tells you
CLAUDE.md (project)	Project name, stack, conventions.
research/product-context.md	Cross-stack foundation exists.
research/icp-research.md, research/market-research.md	Research stack progress.
brand/BRAND.md, brand/DESIGN.md	Marketing stack foundation.
architecture/system-architecture.md	Product stack architecture done.
.agents/skill-artifacts/product/flow/index.md + flow files	Product flows mapped.
.agents/skill-artifacts/meta/specs/*.md	Spec exists from discover.
.agents/skill-artifacts/meta/tasks.md	Tasks decomposed from task-breakdown.
.agents/skill-artifacts/meta/records/diagnose-*.md, .agents/skill-artifacts/meta/sketches/prioritize-*.md, .agents/skill-artifacts/meta/records/targets-*.md	Research mid-pipeline outputs.
.agents/skill-artifacts/mkt/campaign-plan.md + .agents/skill-artifacts/mkt/content/, .agents/skill-artifacts/mkt/lp-brief/, etc.	Marketing artifacts.
.agents/skill-artifacts/meta/records/cleanup-*.md, .agents/skill-artifacts/meta/records/machine-cleanup-*.md	Cleanup audits.
.agents/skill-artifacts/meta/decisions/[date]-*.md, .agents/skill-artifacts/meta/records/[date]-fresh-eyes-*.md	Meta-skill artifacts (dated, immutable — lifecycle: decision / snapshot).
.agents/experience/*.md	All cold-start answers across stacks.
.agents/experience/meta-workflow.md	Prior /start-meta breadcrumb.
.agents/skill-artifacts/meta/records/learned-rules.md	Behavior corrections from prior sessions.

Build a cross-stack state map:

research:
  product-context: done | partial | missing
  icp:             done | partial | missing
  market:          done | partial | missing
  diagnose:        done | not run
  prioritize:      done | partial | missing
  targets:         done | partial | missing

marketing:
  brand:           done | partial | missing
  campaign:        done | partial | missing
  content:         [list of slugs]
  lp:              [audit / brief / both / neither]
  seo:             [list of modes]
  short-form:      [list of slugs]
  outreach:        [list of slugs]

product:
  spec:            done | partial | missing
  flows:           [list]
  architecture:    done | partial | missing
  tasks:           done | partial | missing
  code-cleanup:    done | not run
  docs:            [skim]

meta:
  panel-reports:   [count, latest mtime]
  fresh-eyes:      [count, latest mtime]
  learned-rules:   [count]

Project-fit check: if CLAUDE.md describes a B2B SaaS but research/icp-research.md describes a consumer app, flag the mismatch. State may be from a different project.

---

Step 2: Domain Classification

Parse the user's argument. Classify into one of these:

User says	Classification	Route to
"audience", "ICP", "competitors", "market", "diagnose", "prioritize", "targets", "funnel"	research	/start-research
"brand", "campaign", "copy", "headline", "landing page", "LP", "SEO", "video", "TikTok", "cold email", "outreach", "humanize", "VN tone"	marketing	/start-marketing
"user flow", "tech stack", "architecture", "schema", "API", "code", "refactor", "machine cleanup", "docs", "README"	product	/start-product
"scope this", "clarify", "what should we build", "requirements"	process	/discover
"debate this", "multiple perspectives", "poll", "consensus", "what do experts think"	process	/agents-panel
"decompose", "task list", "break down", "implementation order", "tasks"	process	/task-breakdown
"review my work", "second opinion", "did I miss anything", "post-implementation review"	process	/fresh-eyes
Ambiguous, multi-domain, or "I want to launch a new product"	cross-stack	propose 2–3 stack orchestrators in sequence
Empty	unknown	ask scoping question

If empty, ask:

"What are you trying to do? Pick the closest match:

1. Understand my customers / market / problem (research)
2. Build brand / campaigns / content (marketing)
3. Design a feature / system / flow (product)
4. Scope something vague (discover)
5. Debate a decision (agents-panel)
6. Decompose work into tasks (task-breakdown)
7. Review work I just did (fresh-eyes)
8. I'm not sure — show me what's been done so far"

Option 8 prints the cross-stack state map and asks again.

---

Step 3: Routing Decision

Domain routing:

1. Single-domain intent → defer to that stack's orchestrator. "Looks like a research task — run /start-research." Don't try to do /start-research's job here.
2. Process intent → propose the specific meta-skill (/discover, /agents-panel, /task-breakdown, /fresh-eyes) with rationale.
3. Cross-stack intent ("launch a new product feature") → propose a 2-3 step path:
   • Step 1: /start-research (verify ICP exists; if no, run icp-research)
   • Step 2: /start-product (design feature: flows + architecture)
   • Step 3: /start-marketing (positioning + LP + content for launch)
   • Optional terminal: /fresh-eyes after each.

Process-skill rules:

4. discover — recommend when scope is genuinely unclear. Don't recommend if user has clear intent — they'll find it patronizing.
5. agents-panel — recommend when user explicitly wants debate, OR when multiple equally-valid options surface in routing and user wants help deciding.
6. task-breakdown — recommend when spec.md OR system-architecture.md exists AND user is about to build. Hard-gated on at least one upstream artifact.
7. fresh-eyes — recommend when user is finishing implementation OR has just produced a critical artifact (e.g., system-architecture, brand-system, lp-brief).

Wrap-around suggestions:

• After ANY recommendation that touches security-sensitive code, data-mutation code, or critical artifacts, mention /fresh-eyes as a terminal step.
• Before any non-trivial build, mention /discover as upstream if scope is unclear.

---

Step 4: Present + Confirm

Output format for single-domain:

## Where you are (cross-stack snapshot)

Research:   icp ✅ · market ❌ · prioritize ❌
Marketing:  brand ❌ · campaign ❌ · content (none)
Product:    spec ❌ · flows (none) · architecture ❌
Meta:       no reports yet

## What you asked

"I want to figure out who my customers are" → research domain.

## Recommended: route to /start-research

Why: this is a research-domain task. /start-research will read the
research-stack state and propose the next skill (likely icp-research).

→  /start-research

Output format for cross-stack:

## What you asked

"I want to launch a new product feature" → cross-stack (research + product + marketing).

## Recommended path

1. /start-research        → verify audience clarity for the feature
2. /start-product         → design flows + architecture
3. /start-marketing       → positioning, LP, content for launch
   (optional /fresh-eyes after each artifact)

Each /start-X is its own router; you'll get sub-recommendations from each.

→  Run /start-research first.

Output format for process skill:

## What you asked

"I just finished implementing the auth migration — can you review it?"
→ process intent: post-implementation review.

## Recommended: /fresh-eyes

Why: post-implementation independent review. Runs an independent
agent against your changes, returns issues + severity.

Cost: ~$0.15-0.50 · Duration: ~3 min · Produces: .agents/skill-artifacts/meta/records/[date]-fresh-eyes-<slug>.md

→  /fresh-eyes

---

Step 5: Persist + Hand Off

Append to .agents/experience/meta-workflow.md:

## Session 2026-05-06

- Read state: cross-stack snapshot
- User intent: research-domain (audience clarity)
- Recommended: /start-research
- User confirmed: yes

Print:

Run /start-research next. Re-run /start-meta if your task shifts to a different domain.

Exit.

---

Pipeline Reference

For the canonical cross-stack pipeline, decision rules, and per-skill catalog, see ./references/workflow-graph.md.

---

Anti-Patterns

• Don't ignore the manifest — always read .agents/manifest.json first; per-path filesystem scans are a fallback, not the default.
• Don't duplicate work of /start-research, /start-marketing, /start-product. When intent is single-domain, route there. Don't pick the specific skill yourself.
• Don't lecture about all 24 skills. Show only what's relevant to the user's ask + state.
• Don't auto-invoke. Always print /skill-name for the user to type.
• Don't recommend discover defensively when the user has clear intent. That's patronizing.
• Don't recommend task-breakdown without a spec or architecture upstream. It's hard-gated.
• Don't recommend more than 3 hops in a cross-stack path. If it needs 5+, surface that the project is too vague and recommend discover first.

---

Output

• Inline only.
• Side effect: appends one entry to .agents/experience/meta-workflow.md.

Status

Ends with one of:

• DONE — recommendation given, hand-off printed.
• BLOCKED — couldn't read project state.
• NEEDS_CONTEXT — empty ask + state too sparse to infer. Ask scoping question.