content-config.yaml¶

Content guidance for LLM prompts — media tool usage, narrative strategy, and context assembly.

Location¶

Priority	Location
1	`{repo}/.social-hook/content-config.yaml` (per-project)
2	`~/.social-hook/content-config.yaml` (global fallback)

See examples/content-config.example.yaml for a fully commented template.

Note

This file controls content guidance (how the AI writes). For infrastructure settings (which platforms/models are enabled), see config.yaml.

media_tools¶

Per-tool content guidance for the Drafter agent. Controls when and how each tool is used. Tool availability is set in config.yaml under media_generation.tools.

Default tool slots: mermaid, nano_banana_pro, playwright, ray_so (always present even if omitted from YAML).

Field	Type	Default	Description
`enabled`	bool	—	Per-project override for tool availability. `null` = inherit from config.yaml
`use_when`	list	`[]`	Situations where this tool is appropriate
`constraints`	list	`[]`	Usage rules and limitations
`prompt_example`	string	—	Example prompt for the drafter

media_tools:
  mermaid:
    description: "Text-based diagram generation"
    use_when:
      - "Technical architecture explanations"
      - "Flow diagrams and processes"
    constraints:
      - "Don't overuse - can feel dry/boring"
      - "Best for technical audience"
    prompt_example: |
      Create a Mermaid diagram showing the flow from git commit
      to social post. Use graph LR orientation. Maximum 8-10 nodes.

  nano_banana_pro:
    description: "AI image generation (Google)"
    use_when:
      - "Marketing/announcement visuals"
      - "Polished graphics for launches"
    constraints:
      - "Always specify 'no text' unless text is essential"
      - "Request specific dimensions for platform"

  playwright:
    description: "Browser screenshot capture"
    use_when:
      - "Demonstrating actual UI/product"
      - "Showing working features"
    constraints:
      - "Only use when there's actual UI to show"
      - "Ensure no sensitive data visible"

  ray_so:
    description: "Code screenshot generation"
    use_when:
      - "Highlighting interesting code snippets"
      - "Code-focused posts"
    constraints:
      - "Best for short snippets (< 30 lines)"

Tool selection heuristics¶

The drafter uses these as soft rules:

Technical/architectural post → prefer mermaid (unless used in last 2 posts)
Announcement/milestone post → prefer nano_banana_pro
Working feature demo → prefer playwright
Code-focused post → prefer ray_so
Fallback: nano_banana_pro
Variety: avoid using the same tool 3+ consecutive times

strategy¶

Controls narrative arc management, episode variety, and posting rhythm. These are soft thresholds — the evaluator uses judgment.

Field	Type	Default	Description
`narrative_debt_threshold`	int	`3`	Force synthesis post after N reactive/opportunistic posts
`portfolio_window`	int	`10`	Number of recent posts to consider for variety
`arc_stagnation_days`	int	`14`	Flag narrative arc if no posts for N days
`strategy_moment_max_gap_days`	int	`7`	Force strategy review after N days of inactivity
`episode_preferences.favor`	list	`[]`	Episode types to prefer when applicable
`episode_preferences.avoid`	list	`[]`	Episode types to avoid unless necessary

strategy:
  narrative_debt_threshold: 3
  portfolio_window: 10
  arc_stagnation_days: 14
  strategy_moment_max_gap_days: 7
  episode_preferences:
    favor: [demo_proof, before_after]
    avoid: []

context¶

Controls how much historical context is assembled into LLM prompts. Larger values = better awareness but higher token costs.

Field	Type	Default	Description
`recent_decisions`	int	`30`	Past decisions to include in prompts
`recent_posts`	int	`15`	Past posts to include in prompts
`max_tokens`	int	`150000`	Compaction threshold (triggers at ~75% of 200K window)
`include_readme`	bool	`true`	Include project README.md in evaluator context
`include_claude_md`	bool	`true`	Include CLAUDE.md in evaluator context
`max_doc_tokens`	int	`10000`	Max tokens allocated to project docs
`project_docs`	list	`[]`	Glob patterns for priority files to always include
`pending_draft_detail`	string	`"full_content"`	`"full_content"` or `"summary"` — detail level for pending drafts in evaluator prompts
`arc_context_chars`	int	`500`	Max chars of arc context per pending draft
`pending_drafts_cap`	int	`10`	Max pending drafts included in context
`max_hold_count`	int	`5`	Max held decisions before forcing skip

context:
  recent_decisions: 30
  recent_posts: 15
  max_tokens: 150000
  pending_draft_detail: full_content
  project_docs:
    - "docs/REQUIREMENTS.md"
    - "src/core/**/*.py"

summary¶

Hints for when project summaries should be refreshed. The evaluator uses these as suggestions, not hard triggers.

Field	Type	Default	Description
`refresh_after_commits`	int	`20`	Suggest refresh after N commits
`refresh_after_days`	int	`14`	Suggest refresh after N days

summary:
  refresh_after_commits: 20
  refresh_after_days: 14