Skip to main content

EP12. Workflow Creation Skill

🏛️ The Meta-Engine and 7-Step Meta Snowflake

Recent episodes: from blank page through ASR, ADR, Spec, to incremental build-out.

That generic pipeline assumes structure isn't obvious upfront. But for slide decks or tech blog series with proven, repeatable workflows, re-running universal ASR/ADR every time is wasteful.

Enter generating-skill-creation—designing dedicated skills like document-authoring and presentation-authoring. Not the artifact—the recipe for producing it. Let's open this meta-skill's heart.


🧩 Five Prisms for Decomposing Artifacts

When asked to design a new generation skill, generating-skill-creation tears the target into five dimensions:

ComponentRoleBlog series exampleImage generation example
Assignment & ConstraintsScope, quality bar, taboos, success criteriaSeries theme, audience, episode countResolution, format, color limits
Context & RulesDomain rules, world-building, style guidePlatform norms, toneBrand guidelines, palette
EntitiesReusable building blocksEpisode bodies, outlinesIcons, logos, character assets
Space & PlacementPhysical/visual layoutEpisode order in series arcCanvas layout, object placement
Structure & FlowHierarchical content and hierarchyIntro → body → conclusionLayer order, visual hierarchy

❄️ 7-Step Meta Snowflake Pipeline

Cocrates doesn't throw a one-off prompt blob. Under Snowflake strict control—no generation until detail design is locked—it runs seven steps:

[1. Kernel] ─→ [2. Components] ─→ [3. Frame] ─→ [4. Outline] ─→ [5. Spec] ─→ [6. Skill] ─→ [7. Verification]

1. Kernel (One-sentence definition)

Identity and mission in one sentence. If it wobbles, no next step.

"This skill helps produce [artifact type] in [situation] through [reviewable stages]."

2. Components (Dimension identification)

Catalog what the skill must handle using the five prisms.

3. Frame (Workflow design)

Design the new skill's 5-step Snowflake workflow, intermediate file tree, and user approval gates.

4. Outline (Per-stage specification)

For each stage: Input, Creation activity, Completion criteria, Review questions, Approval criteria—so intermediates land on disk, not in volatile chat.

5. Spec (Assemble specification)

Lock frontmatter, when to use, procedures, dialogue rules, prohibitions for final .SKILL.md.

6. Skill (Generate skill file)

From approved spec only, build .opencode/skills/{skill-slug}/SKILL.md.

7. Verification (Harness check)

Confirm the file obeys harness rules and runs correctly.


🚫 Meta-Engine Prohibitions

While generating-skill-creation runs, the agent cannot:

  • Roughly generate final artifacts because no skill fits yet.
  • Fill a SKILL.md template without multi-dimensional analysis.
  • Design skills missing user review/approval gates.
  • Leave intermediates only in chat without file isolation.

📝 Three-Line Summary

  1. generating-skill-creation adds recipes to the system—not one-off dishes.
  2. Five dimensions—assignment, context, entities, space, structure—yield precise dedicated workflows.
  3. Dedicated skills auto-load for repeat requests and evolve with the user.

🎬 Coming Up Next

You've mastered Learning → Artifact Generation → Meta Skill authoring.

With architecture understood—what commitment should users hold, and how does the harness evolve sustainably? The penultimate episode closes the arc.

"The moment you encode a skill in the system, your know-how is etched into AI's genes."


This series introduces the Cocrates Harness framework. Cocrates is an agent harness designed for Socratic dialogue so users keep agency and grow.