| name | ideon |
| description | Use this skill when users need to generate multi-format professional content (articles, blog posts, newsletters, social media posts, academic papers) from a single idea using the Ideon 7-stage deterministic pipeline. |
Ideon Skill: Agent-Driven Content Generation
This skill enables an agent to autonomously generate professional, multi-format content using Ideon's 7-stage deterministic pipeline. An agent using this skill will orchestrate planning, writing, image generation, link enrichment, and file output—with mandatory user approval gates at critical junctures.
What This Skill Does
Ideon is a deterministic content generation system that converts a single idea into multiple professionally-formatted outputs (articles, blog posts, newsletters, social media posts, academic papers, etc.) via a 7-stage pipeline:
- Shared Plan — Generate cross-channel content guidance
- Planning — Create title, slug, description, and detailed content plan
- Sections — Write long-form content (intro, body sections, conclusion)
- Image Prompts — Expand and refine image descriptions
- Images — Render images via Replicate AI
- Output — Generate channel-specific formatted content (markdown + metadata)
- Links — Enrich content with contextual external links via web search
The skill directs agents to execute this pipeline with mandatory approval gates where users can review and adjust outputs before proceeding.
When to Use This Skill
✅ Use for: Multi-format content generation with images, links, and metadata. Ideal for articles, blog posts, newsletters, social media posts with professional guidance.
❌ Skip for: Quick snippets, single posts without images/links, real-time editing, no API access (requires agent-driven writing and optional Replicate image access).
Required Inputs
An agent MUST collect these inputs from the user upfront before proceeding:
Required
- Idea (string, 10-500 chars) — The content topic or concept
- Example: "React suspense for data fetching", "AI risks in healthcare", "Building a SaaS side project"
- Primary Target Format (one of: article, blog-post, newsletter, x-post, x-thread, reddit-post, linkedin-post, press-release, science-paper) — The main output format
Optional (with defaults)
- Secondary Target Formats (list) — Additional output formats to generate alongside primary
- Default: None (only primary target)
- Custom Links (list of expression -> URL pairs) — Optional user-provided links to force into output during link enrichment
- Default: None
- Example:
"React Suspense API documentation" -> "https://react.dev/reference/react/Suspense"
- Writing Style (one of 13 options): professional, technical, friendly, analytical, empathetic, academic, persuasive, journalistic, authoritative, conversational, playful, minimalist, storytelling (default: professional)
- Content Intent (one of 13 options): tutorial, case-study, opinion-piece, deep-dive-analysis, cornerstone, interview-q-and-a, how-to-guide, listicle, announcement, critique-review, counterargument, personal-essay, roundup-curation (default: tutorial)
- Target Length (small | medium | large) — Target word count
- small: ~500 words
- medium: ~900 words
- large: ~1400 words
- Default: medium
- Image Count (integer, 1–10) — Maximum images to generate
- Keywords (list of strings) — SEO keywords to target in the content
- Default: None (auto-generated during planning if not provided)
- Supports compound keywords:
"organic marketing, content strategy, seo"
- When provided (or when the planner generates keywords), the planner assigns
primaryKeyword and per-section targetKeywords (0-2 each) with parity placement rules; the intro writer places primaryKeyword in the first 100 words; section writers use each section's targetKeywords in BLUF openers
- After section writing, the default-on
seo-check stage lints placement and may run a five-tool surgical editor agent (errors-only pass by default; --seo-check-mode strict for zero warnings; --seo-check-max-turns / seoCheckMaxTurns for turn limit; --no-seo-check to skip; ideon write resume --seo-check to re-run)
- Keywords can also be defined at the series level and are inherited by all articles in that series
- Author (slug) — Author profile for voice and expertise, injected into all writing prompts when resolved
- Resolution: run author → series
defaultAuthor → publication defaultAuthor
- Default: None (draft may use
[AUTHOR: …] placeholders)
- Experience notes (string) — Per-run anecdotes or first-hand observations to weave into the draft (supplements series
experienceNotes)
- Editorial checklist — After a successful run, review
meta.json editorialChecklist and CLI/MCP summary for pre-publish items (byline, AI disclosure, placeholder cleanup)
System Requirements (agent provides)
- Replicate API Token — For image generation (optional if image count = 0)
Workflow Overview
An agent executing this skill follows this sequence:
1. Collect all required + optional inputs from user
2. Validate execution requirements (agent writing capability, optional Replicate token)
3. Load writing guides for chosen (format, style, intent)
4. Execute Stages 1–2 (Shared Plan + Planning)
5. ⭐ CHECKPOINT 1: Present plan to user (title, slug, description, content structure)
└─ User must explicitly approve before proceeding
└─ Options: Approve, Edit (adjust style/intent/length), Restart, Cancel
6. Execute Stage 3 (Sections) — Write long-form content
7. Execute Stage 4–5 (Image Prompts + Image Rendering) — Generate images via Replicate
8. ⭐ CHECKPOINT 2: Present full content + images to user
└─ User reviews quality before proceeding to links
└─ Options: Approve, Skip links (faster), Retry links, Cancel
└─ Special: If link enrichment times out, offer user choice
9. Execute Stage 6 (Output) — Generate channel-specific formatted files
10. Execute Stage 7 (Links) — Enrich with contextual external links
11. ⭐ CHECKPOINT 3: Confirm output location + file list to user
└─ User approves file placement before writing to disk
└─ Options: Approve, Choose different directory, Cancel
12. Write all files to user-approved output location
13. Return meta.json summary + file paths to user
Output Structure
Each run writes a flat generation directory (default ~/.ideon/output/<timestamp>-<slug>/):
<output-dir>/<timestamp>-<slug>/
├─ article-1.md # Primary format + frontmatter
├─ article-1.links.json # Link sidecar (v2) when enrichment ran
├─ meta.json # Generation metadata (title, slug, outputs, checklist)
├─ job.json, plan.md, generation.analytics.json
├─ cover-1.png, inline-*.png # Images (relative paths from markdown)
└─ [... secondary .md + .links.json per target format ...]
Resume state lives in ~/.ideon/sessions/<project-hash>/state.json, not in the generation directory.
See output-spec.md for meta.json, links.json v2, and markdown frontmatter. CLI/MCP agent reference: skill/ideon-cli/references/output-structure.md.
API Requirements
Replicate (Image Generation - Optional)
- Only needed if image count > 0
- Default model:
flux (see replicate.md for details)
- Estimated cost: $0.10–0.25 USD per image
- Timeout: If image rendering >5 minutes per image, offer user choice (skip, retry, cancel)
For LLM-based stages (planning, sections, links), the agent using this skill handles all LLM calls directly.
Approval Checkpoints
Checkpoint 1: Plan Approval (MANDATORY)
When: After Stages 1–2 (Shared Plan + Planning) complete
Agent presents:
- Title (auto-generated)
- URL slug (auto-generated)
- Description (auto-generated)
- Content plan structure (sections, estimated word count, key topics)
User options:
- ✅ "Approve and continue to sections" — Proceed to Stage 3
- 🔄 "Edit plan" — Adjust style, intent, or length; regenerate plan
- 🔁 "Restart with different inputs" — Return to input collection step
- ❌ "Cancel" — Abort the entire run
Checkpoint 2: Content Review (with Link Timeout Handling)
When: After Stages 3–5 complete (Sections + Images)
Agent presents:
- Full rendered markdown content (with image placeholders showing where images will appear)
- All generated images with captions
- Estimated word count vs. target
- Preview of how content will look in primary format
User options:
- ✅ "Approve and continue to link enrichment" — Proceed to Stage 7
- 🔗 "Add or edit custom links" — Provide/modify expression -> URL links before Stage 7 merge
- ⚠️ "Skip links, save now" — Skip Stage 7 and go to output (faster, no external links)
- 🔁 "Start over with different inputs" — Return to input collection
- ❌ "Cancel" — Abort
Special: Link Enrichment Timeout Handling
- If Stage 7 (link enrichment) exceeds 30 seconds, agent pauses with message:
- "Link enrichment is taking longer than expected. Options: (1) Continue waiting, (2) Save content without links, (3) Cancel"
- User chooses, agent proceeds accordingly
Checkpoint 3: Output Confirmation (MANDATORY)
When: After all content is ready (post-Stage 7)
Agent presents:
- Output directory path (full path on user's machine)
- File list (all markdown files + images + metadata)
- Total estimated cost breakdown (LLM + images)
- Metadata filename (
meta.json)
User options:
- ✅ "Save to this location" — Write all files to disk
- 📁 "Choose different directory" — Specify alternate path, then save
- ❌ "Cancel" — Abort before writing files
Error Handling & Recovery
See troubleshooting.md for detailed guidance on:
- API failures (quota exceeded, rate limits, invalid tokens)
- Image generation failures (Replicate timeouts, unsupported models)
- Content quality issues (sections too short, links not found)
- Network timeouts and retry strategies
Reference Documentation
This skill references the following detailed guides. Agents should load these as needed:
| Document | Purpose |
|---|
| guides.md | How to load modular writing guides (format + intent + style + seo) |
| guide-map.json | Index mapping content types, styles, intents, seo to guide files |
| pipeline.md | Detailed 7-stage workflow orchestration |
| checkpoints.md | Approval checkpoint logic and user interaction patterns |
| output-spec.md | Output file formats (session.json, meta.json, markdown, links.json v2) |
| replicate.md | Image generation integration (models, costs, error handling) |
| examples.md | End-to-end workflow examples for common scenarios |
| troubleshooting.md | Edge cases, error recovery, and FAQ |
📌 Self-Contained Writing Guides
All writing guides (formats, styles, content intents, seo, references) are synced into this skill directory under guides/. This ensures agents have full access to all required guides without external dependencies.
Location: ideon/skill/ideon/guides/ (populated via npm run guides:sync)
To update guides: Run npm run guides:sync in the ideon directory to sync the latest guides from the canonical source (ideon/writing-guide/).
See guides.md for detailed sync instructions.
Quick Start Example
Request: "Technical deep-dive blog post on Rust async/await + LinkedIn version + 4 images"
Workflow: Collect inputs → Validate APIs → Load guides → Plan (✓ checkpoint) → Sections → Images (✓ checkpoint) → Output + Links (✓ checkpoint) → Save files → Done