Code to Content Skill

Transform codebases into compelling developer content through a mandatory 6-phase process with verification gates.

Core premise: when AI makes competent prose free, grounded-but-generic content gets ignored. The thing that wins is information only the builder knows, written the way only they would write it. This skill's evidence and readability machinery stops content from being bad; Phase 2 (Differentiation Discovery) and references/differentiation.md stop it from being generic. The governing test, applied at delivery: cover the logo, swap in a competitor's name — if it still reads fine, it failed.

Specialized Agents

This skill includes 3 specialized agents for parallel execution:

Parallel Execution Pattern

Phase 1 - Code Analysis:

Launch 3 content-explorer agents in parallel:
├── Agent 1: "Find performance and optimization story angles"
├── Agent 2: "Find architecture and design decision angles"
└── Agent 3: "Find developer experience and tooling angles"

Consolidate: Merge findings, deduplicate, select top 3-5 angles

Phase 6 - Quality Review:

Launch 4 quality-reviewer agents in parallel:
├── Agent 1: "Review structure and format compliance"
├── Agent 2: "Review readability and audience fit"
├── Agent 3: "Review evidence grounding and technical accuracy"
└── Agent 4: "Review distinctiveness — swap-the-name test + AI-tells blocklist"

Consolidate: Merge issues, prioritize by severity, address blocking items

Reference Loading Rules (Lazy Loading)

NEVER load all references at once. Load only what's needed for the current format:

Always Load:

• phase-gates.md — Gate verification criteria (required for all phases)
• differentiation.md — Swap-the-name test, the three moves, AI-tells blocklist (required for Phase 2 and Phase 6)
• readability-guide.md — Readability validation (required for Phase 6)

Load On-Demand:

• analysis-prompts.md — Only if running Claude-native analysis (Phase 1)
• project-analysis.md — Differentiation Discovery protocol + developer interview (Phase 2)
• audiences.md — Only if audience calibration needed (Phase 3)
• brand-voice.md — Founder-voice application; tech-stack fallback (Phase 3)
• cognitive-load.md — Only if complex content requiring progressive disclosure (Phase 5)
• diagram-templates.md — Only if visual assets required (Phase 5)
• code-snippets.md — Only if content includes code examples for social sharing
• conference-talks.md — Only if generating CFP or talk outline
• documentation.md — Only if generating README, API docs, or architecture docs
• mcp-integration.md — Only if posting via MCP servers (Delivery)

Mandatory 6-Phase Process

You MUST complete all 6 phases in order. Each phase has a gate that MUST pass before proceeding.

PHASE 1 ──[Gate]──> PHASE 2 ──[Gate + Confirm]──> PHASE 3 ──[Gate + Confirm]──> PHASE 4 ──[Gate]──> PHASE 5 ──[Gate]──> PHASE 6 ──[Gate]──> DELIVERY
  Code              Differentiation               Audience &                   Content          Optimization      Verification
  Analysis          Discovery                     Format                       Generation                         & Delivery

Phase 2's founder interview is always offered, never blocking: if the user declines, proceed on code alone and carry a Distinctiveness: AT RISK flag to the Phase 6 report. See references/phase-gates.md for detailed gate verification criteria.

User Confirmation Checkpoints

Phases 2 and 3 require explicit user confirmation before proceeding.

The Phase 2 confirmation is also where the differentiation interview is offered. Skip confirmations when: User says "proceed without confirmations", using Quick Mode, or iterating on existing content. (Skipping confirmation does not skip the Phase 2 offer — it just doesn't wait; the offer can still be declined.)

Phase 1: Code Analysis

You MUST complete this phase before any content generation. This phase mines the codebase for the WHAT — tech stack, architecture, story hooks. The WHY, opinion, and roads-not-taken come from the human in Phase 2.

Execution Options

Choose based on task complexity:

Option A: Agent-Parallel Execution (Recommended)

Launch multiple content-explorer agents with different focus areas:

Agent 1: "Analyze [project] for performance and optimization story angles"
Agent 2: "Analyze [project] for architecture and design decision angles"
Agent 3: "Analyze [project] for developer experience and tooling angles"

Consolidate agent outputs into unified Project Brief.

Option B: Inline Analysis Protocol

Follow references/analysis-prompts.md for Claude-native analysis:

1. Tech stack detection (read package.json, requirements.txt, etc.)
2. Architecture pattern detection from directory structure
3. Story hook discovery (grep for TODO, FIXME, HACK, etc.)
4. Git history mining for narrative elements
5. Content angle generation (3+ angles with evidence)

Phase 1 Output: Project Brief (code-derived half)

## Project Brief: [Name]

### Tech Stack
- **Language:** [primary]
- **Framework:** [framework]
- **Architecture:** [pattern]
- **Voice Profile (fallback):** [precise/pragmatic/accessible/direct — replaced by founder voice if Phase 2 captures one]

### Content Angles (3+ Required)
1. [Angle with evidence: file:line or commit]
2. [Angle with evidence]
3. [Angle with evidence]

### Story Hooks
- [Hook 1 with source reference]
- [Hook 2 with source reference]

Phase 1 Gate: Code Brief Generated

You MUST verify before proceeding:

• [ ] Tech stack identified
• [ ] At least 3 content angles discovered
• [ ] Story-worthy element found (commit, pattern, or insight)
• [ ] All angles have evidence citations

STOP if: No content-worthy insights found after analysis.

Proceed directly to Phase 2 (no user wait yet — confirmation happens after the Differentiation Brief is assembled).

Phase 2: Differentiation Discovery

The code told you WHAT. This phase gets the WHY, the opinion, and the roads not taken — the material a competitor could not republish. Load references/differentiation.md and references/project-analysis.md.

This phase always runs. Its founder interview is ALWAYS OFFERED and NEVER BLOCKING. If the user declines, proceed on code alone and set Distinctiveness: AT RISK.

Required Actions

1. Make the offer (one message, three paths + opt-out):

"To make this unmistakably yours — something a competitor couldn't republish — I can (a) ask you 4–5 quick questions, (b) take raw material (a Slack thread, support tickets, a 2-minute voice-memo transcript, rough notes), or (c) polish a rough draft you write (write it ugly; I'll keep your voice). Or I can proceed from the code alone. Which do you want?"

1. Mess over spec — if raw material is offered, rank it ABOVE the polished README. Rough Slack/tickets/transcripts carry the signal.

1. Extract the three moves (see project-analysis.md §4 for the full question set):

• THE WHY — the thesis/stakes; the only-you-know insight to lead with
• THE SPIKY CLAIM — one defensible opinion a reader could disagree with
• ROADS NOT TAKEN — what they chose not to build + the trade-off

1. Capture founder voice as rejections ("never use semicolons", "never corporate") — this becomes the primary voice in Phase 3, displacing the tech-stack fallback.

Phase 2 Output: Differentiation Brief

Append to the Project Brief:

### Differentiation Brief
- **THE WHY:** [thesis to lead with]
- **THE SPIKY CLAIM:** [defensible opinion — or "none captured"]
- **ROADS NOT TAKEN:** [what they didn't build + trade-off — or "none captured"]
- **FOUNDER VOICE:** sounds like [anchor] · never [rejections] · source: [interview/transcript/draft/NONE]
- **RAW MATERIAL:** [slack/tickets/transcript/notes — or "none"]
- **DISTINCTIVENESS FORECAST:** [ON TRACK | AT RISK]

Phase 2 Gate: Differentiation Forecast Set (soft)

You MUST verify before proceeding:

• [ ] The offer was made (regardless of response)
• [ ] Differentiation Brief produced (fields filled or explicitly "none captured")
• [ ] Forecast set: ON TRACK (≥1 of WHY/opinion/roads captured) or AT RISK (code only)

This gate never STOPS delivery. AT RISK is allowed — it just propagates a loud warning to Phase 6. Do not block on missing founder input.

Phase 2 Confirmation

Present the combined Project Brief + Differentiation Brief and WAIT for confirmation.

Display: tech stack, 3+ angles, the WHY, the spiky claim, roads not taken, recommended angle.

Ask: "Proceed with this analysis? [Yes / Use angle # / Add context]"

DO NOT proceed to Phase 3 until user confirms (unless confirmations are skipped).

Phase 3: Audience & Format Declaration

You MUST declare audience and format before generating content.

Required Actions

1. Determine target audience (select ONE):

1. Select content format (select ONE):

• Blog post
• Tutorial
• Twitter/X thread
• LinkedIn post
• README
• Newsletter
• Video script
• Conference talk

1. Declare voice profile. Use the FOUNDER VOICE from the Phase 2 Differentiation Brief as the primary voice. Only if it is NONE (offer declined), fall back to the tech-stack heuristic — which is safe but swappable:

See references/audiences.md for full calibration rules and references/brand-voice.md for applying a captured founder voice.

Phase 3 Gate: Audience Contract Established

You MUST verify before proceeding:

• [ ] Single audience selected (no mixing)
• [ ] Format matches audience complexity
• [ ] Voice profile declared (founder voice if captured, else tech-stack fallback)

STOP if: Audience mixing detected or format incompatible.

Phase 3 Confirmation

Present the Audience Contract to user and WAIT for confirmation.

Display:

• Selected audience and what it means (grade level, jargon %)
• Selected format
• Voice profile (and whether it's the founder's or the fallback)

Ask: "Proceed with this targeting? [Yes / Change audience / Different format]"

DO NOT proceed to Phase 4 until user confirms.

This is the last confirmation before content generation begins.

Phase 4: Content Generation

You MUST ground all claims in project evidence — and lead with the Differentiation Brief.

Required Actions

1. Load format template from assets/templates/
2. Apply format-specific workflow (see below)
3. Lead with the WHY, not the feature list. Place the spiky claim and at least one road-not-taken into the draft using the template's differentiation slots (WHY/thesis lead · YOUR TAKE · ROADS NOT TAKEN). If the Differentiation Brief says none captured for these, do not invent them — generate honestly and let Phase 6 flag AT RISK.
4. If a rough founder draft was provided (write-ugly-first path), polish it — preserve their phrasing and order of thought; do not replace the voice.
5. Ground every claim in evidence:

• Code examples from actual codebase
• Metrics with source citations
• Commit hashes for historical claims

Format-Specific Workflows

Each format has a dedicated template in assets/templates/:

Load the appropriate template and follow its structure.

Phase 4 Gate: Draft Complete with Evidence

You MUST verify before proceeding:

• [ ] All code examples from actual codebase
• [ ] All metrics/claims traceable to source
• [ ] Template structure followed
• [ ] WHY leads (not a feature list); captured spiky claim / roads-not-taken are present if the brief had them

STOP if: Claims cannot be traced to evidence.

Phase 5: Optimization

You MUST optimize for audience and platform.

Required Actions

1. Verify voice consistency — Same tone throughout
2. Apply cognitive load management:

1. Apply SEO (for web content):

• Primary keyword in title
• Keyword in first 100 words
• Descriptive subheadings

1. Prepare visual assets:

• Syntax-highlighted code blocks
• Architecture diagrams (see references/diagram-templates.md for Mermaid templates)
• Screenshots with annotations

See references/optimization.md for full optimization rules.

Phase 5 Gate: Enhancement Applied

You MUST verify before proceeding:

• [ ] Voice consistent throughout (and matches the captured founder voice, including its rejections)
• [ ] Cognitive load appropriate for audience
• [ ] Visual assets enhance (not distract)

Phase 6: Verification & Delivery

You MUST NOT deliver content that fails verification.

Execution Options

Option A: Agent-Parallel Review (Recommended for important content)

Launch multiple quality-reviewer agents with different focus areas:

Agent 1: "Review for structure and format compliance"
Agent 2: "Review for readability and audience fit"
Agent 3: "Review for evidence grounding and technical accuracy"
Agent 4: "Review for distinctiveness: run the swap-the-name test and AI-tells blocklist"

Consolidate findings, address all Critical and High severity issues before delivery.

Option B: Inline Validation Protocol

1. Format Checklist — Load from references/checklists.md. Universal checks: hook, single takeaway, audience depth, voice, code syntax.
2. Readability — Use references/readability-guide.md. Thresholds: Beginner ≤8.0 grade/2% jargon, Intermediate ≤12.0/4%, Expert ≤16.0/8%.
3. Evidence — All claims cite actual files (path:line), commits, or measurements.
4. Distinctiveness — Use references/differentiation.md. Run the swap-the-name test, the AI-tells blocklist, and score the Distinctiveness Score (0–5).

Phase 6 Gate: Delivery Approved

You MUST verify before delivery:

• [ ] Format checklist: 100% blocking items pass
• [ ] Readability validation: PASS for declared audience
• [ ] Evidence verification: ALL claims grounded
• [ ] Voice consistency: Same tone throughout
• [ ] Distinctiveness reported: swap-the-name verdict + score recorded (this is REPORTED, never blocking)

DO NOT DELIVER until the first four checks pass. Distinctiveness is reported, not gated — but if the Distinctiveness Score is 0–2, surface the AT RISK warning prominently and offer to run Phase 2 before publishing.

Delivery Output

Include with final content:

## Quality Report

- **Audience:** [declared audience]
- **Format:** [format type]
- **Readability Grade:** [estimated grade] / [max for audience]
- **Jargon Density:** [percentage] / [max for audience]
- **Evidence Status:** All claims grounded
- **Checklist:** [X/X] items passed
- **Distinctiveness Score:** [N]/5 [(missing: ...)]
- **Swap-the-name test:** [PASS | AT RISK]
- **Founder input:** [used (source) | declined → AT RISK]

**Verdict:** APPROVED FOR DELIVERY

Stop Conditions

STOP and inform user when: No insights found (Phase 1), audience unclear or mixed (Phase 3), claims lack evidence (Phase 4), or readability fails (Phase 6). Do NOT stop when founder input is declined in Phase 2 — proceed on code alone and flag AT RISK.

Constraint Propagation

Each phase output is required input for the next: Code Brief → Differentiation Brief → Audience Contract → Draft → Optimized Content → Verified Deliverable. You MUST NOT skip phases.

Reference Files

Load on-demand. Always load: phase-gates.md, differentiation.md, readability-guide.md

Format-specific: newsletters.md, social-content.md, conference-talks.md, documentation.md, product-launch.md, build-in-public.md, code-snippets.md, code-explainers.md, posting-plan.md, mcp-integration.md

Slash Commands

See references/commands.md for full usage details. Commands are in .claude/commands/c2c/.

Examples

See examples/ for input/output demonstrations:

• blog-post/, twitter-thread/, readme/ — Positive examples with input and output
• negative/ — Gate failures with corrections
• workflow/ — Full 6-phase process

Evaluation

Test skill with evaluation/evaluation.xml (18 QA pairs covering template structure, audience thresholds, voice calibration, and quality gates).

Common Issues

Design Principles

1. Evidence Over Invention — Every claim traces to actual code, commits, or metrics. NEVER invent. (This is the floor — it stops bad content.)
2. Distinctiveness Over Competence — Grounded-but-generic loses. Lead with the WHY, hold one defensible opinion, show a road not taken. The swap-the-name test is the ceiling — it stops generic content. NEVER block delivery on it; warn instead.
3. Voice Is a Person — Capture the founder's actual voice (especially what they reject). The tech-stack table is a fallback, not a target.
4. Audience-First — Determine audience in Phase 3. Depth and tone flow from this choice.
5. Phase Gate Enforcement — Each gate MUST pass before proceeding (except the soft Phase 2 forecast). NOT optional.
6. Format Follows Function — Use the appropriate template and checklist for each format.

MCP Integration (Optional)

When MCP servers (twitter-mcp, linkedin-mcp, buffer-mcp) are available, this skill can post directly to platforms. See references/mcp-integration.md for setup. Without MCP, templates include copy-ready format with character counts.

Automatic Quality Hooks (Optional)

This skill supports automatic content validation via Claude Code hooks.

See references/hooks-guide.md for setup and customization.

Requirements

All core analysis is Claude-native with no dependencies. Optional Python 3 for automatic hooks validation.