Authoring Skill Content

Overview

Guide the authoring of effective SKILL.md files, agent definitions (agents/*.md), and supporting resources within a bundle-plugin. Good content is the difference between skills that agents consistently find and follow — and ones that get ignored or misinterpreted.

Core principle: Write for the agent's experience. Every instruction should be discoverable (good description), loadable (right size), and followable (clear, motivated instructions). Skills are the first-class source of truth in a bundle-plugin — docs and README must not contradict skill content (see bundles-forge:auditing — references/source-of-truth-policy.md).

Skill type: Hybrid — follow the execution flow rigidly (Entry Detection → Path steps → Validation), but apply writing guidance flexibly based on context. The process is discipline-enforcing; the content decisions are pattern-based.

Announce at start: "I'm using the authoring skill to help write / complete / improve / adapt] skill / agent] content."

Entry Detection

Determine the authoring path from context:

| Context | Path | |---------|------| | skill-inventory from blueprinting, or user requests writing new SKILL.md / agent definition from scratch | Path 1: New Content | | User provides an existing/external skill to add into a project, or asks to adapt a skill to match project conventions | Path 2: Integrate Content | | scaffold-output directories exist but SKILL.md body has < 10 non-empty lines | Path 3: Complete Content | | User provides existing in-project skill-md to improve, or optimization-spec from optimizing with specific changes | Path 4: Improve Content |

When the target is an agent definition (agents/*.md) rather than a skill, follow the same path logic but use the agent authoring conventions from references/agent-authoring-guide.md.

Step 0: Project Context (all paths)

Before writing any content, verify scope and detect the project context:

1. Triage: should this be a skill? — Before writing, verify the content warrants a skill:

• One-off, project-specific conventions → belongs in CLAUDE.md / AGENTS.md, not a skill
• Mechanically enforceable constraints (regex, schema validation) → automate with scripts, not documentation
• Standard practices well-documented by the platform → don't duplicate, cross-reference instead
• Skip this check when arriving from bundles-forge:blueprinting (triage already done during design)

1. Detect project root — look for skills/ directory + package.json above the target
2. If project exists, read 2-3 existing SKILL.md files to extract the project's conventions:

• Description style (verb form after "Use when", scoping patterns)
• Section structure (which headings, in what order)
• Cross-reference format (project:skill-name prefix)
• Token efficiency patterns (use of references/, line counts)

1. If no project (standalone authoring), use the conventions from references/skill-writing-guide.md directly

Path 1: New Content

Write skill or agent content from scratch.

1. Gather requirements — from skill-inventory and design document context (blueprinting), user description, or conversation context. Identify: skill purpose, triggering scenarios, expected inputs/outputs, relationship to other skills. When a design document is available, leverage its project overview, target users, and use cases to write more targeted descriptions and overviews
2. Load writing guide — read references/skill-writing-guide.md (frontmatter conventions, description rules, instruction style)
3. Write frontmatter — name (kebab-case matching directory), description (start with "Use when...", under 250 chars, triggering conditions only)
4. Write Overview — 1-3 sentences: what the skill does, core principle, skill type declaration (rigid / flexible / hybrid)
5. Write the process — step-by-step execution flow. Use imperative form. Explain why, not just what. Include at least one concrete example per key instruction
6. Write Common Mistakes — table of pitfalls and fixes (at least 3 entries)
7. Write Inputs / Outputs / Integration — declare artifact IDs, calling relationships, and pairing skills
8. Check external dependencies:

• Declaration syntax — if the skill references MCP tools or CLI commands, read references/skill-writing-guide.md "External Tool References" section for allowed-tools declaration, fallback patterns, and CLI vs MCP guidance
• Prerequisites section — if allowed-tools declares external CLI tools (not git, python, node, npm, npx, bash, or paths under bin//scripts/), confirm the body includes a ## Prerequisites section with a Tool/Check/Install table. Read references/skill-writing-guide.md "Prerequisites Writing" for the standard format

1. Evaluate token budget — if body exceeds 300 lines, extract heavy sections to references/. Front-load critical instructions in the first ~5,000 tokens — after context compaction, only this portion survives
2. Run validation (see Post-Action Validation below)

Path 2: Integrate Content

Adapt an existing/external skill to fit a project's conventions and workflow.

1. Read the incoming skill — understand its purpose, triggering scenarios, and current structure
2. Read project conventions (from Step 0) — identify gaps between the incoming skill and project patterns
3. Load writing guide — read references/skill-writing-guide.md (frontmatter conventions, description rules, instruction style)
4. Adapt frontmatter — rewrite description to match project style (verb form, scoping), ensure name follows project kebab-case convention
5. Adapt body structure — restructure sections to match project patterns (Overview, Process, Common Mistakes, Inputs/Outputs/Integration)
6. Wire Integration section — add cross-references to existing project skills, declare Inputs/Outputs that connect to the project's workflow graph
7. Adapt instruction style — align with project conventions (imperative form, reasoning over directives, example density)
8. Evaluate token budget — ensure the adapted skill fits project token norms
9. Run validation (see Post-Action Validation below)

Path 3: Complete Content

Fill in scaffolded skill stubs with substantive content.

1. Read scaffold structure — identify which directories and stub files exist
2. Read project conventions (from Step 0) — match the style of already-completed skills
3. Load writing guide — read references/skill-writing-guide.md (frontmatter conventions, description rules, instruction style)
4. Complete frontmatter — fill in description (triggering conditions, "Use when...", under 250 chars). If name is already set, verify it matches directory name
5. Write Overview — core principle + skill type declaration
6. Write the process — step-by-step flow based on the skill's intended purpose from the design document or user context
7. Write remaining sections — Common Mistakes, Inputs/Outputs/Integration
8. Create supporting resources if needed — references/ for heavy content, assets/ for templates. Read references/skill-writing-guide.md "Supporting Resources" section for thresholds. Front-load critical instructions in the first ~5,000 tokens — after context compaction, only this portion survives
9. Run validation (see Post-Action Validation below)

Path 4: Improve Content

Enhance existing in-project content based on user feedback or optimization specs.

1. Read existing content — understand current structure, strengths, and gaps
2. Identify improvement targets — from user request, optimization-spec, or self-diagnosis:

• Description not triggering reliably → rewrite following description rules
• Token budget exceeded → extract to references/, cut redundancy, front-load critical instructions in the first ~5,000 tokens
• Missing sections → add Overview, Common Mistakes, Inputs/Outputs
• Instruction style issues → reframe directives as reasoning, add examples

1. Load writing guide if needed — read references/skill-writing-guide.md (frontmatter conventions, description rules, instruction style)
2. Apply changes — make targeted improvements. Preserve what works; don't rewrite content that already follows conventions
3. Verify Integration section — ensure cross-references still resolve after changes, artifact IDs match consuming skills
4. Run validation (see Post-Action Validation below)

Post-Action Validation

After completing any path, validate the authored content:

1. Run lint — bundles-forge audit-skill <skill-directory> on each authored/modified skill

• If audit script is unavailable: review against references/quality-checklist.md manually

1. Act on findings — fix Critical immediately, fix Warning if straightforward, report Info as suggestions
2. Report results — tell the user/calling skill what was validated and any remaining warnings

Common Mistakes

| Mistake | Fix | |---------|-----| | Description summarizes workflow | Describe triggering conditions only — agents shortcut to description | | Piling on MUST/ALWAYS/NEVER | Explain why the rule exists — understanding beats compliance | | Putting everything in SKILL.md | Extract heavy content (100+ lines) to references/; keep body under 500 lines | | No examples, only abstract rules | Add at least one concrete example per key instruction | | Writing for humans, not agents | Use imperative form, clear structure, explicit output formats | | Description too narrow | Be pushy — list related scenarios, edge cases, alternative phrasings | | Description too broad | Scope to the right context (e.g., "bundle-plugins" not just "any project") | | Skipping project conventions | Always read existing skills first when working in an established project | | Not wiring Integration section | Every skill needs Called by / Calls / Pairs with to connect to the workflow graph | | No defensive instructions for rigid skills | For rigid/hybrid skills, add explicit loophole closers, rationalization tables, and red-flag lists — see Defensive Writing in references/skill-writing-guide.md | | Forgetting validation | Always run bundles-forge audit-skill after authoring — catches issues before they propagate |

Inputs

• skill-inventory (optional) — list of skills and agent definitions to write, from bundles-forge:blueprinting design document
• scaffold-output (optional) — scaffolded skill directories needing content (from blueprinting pipeline or standalone scaffolding)
• skill-md (optional) — existing SKILL.md or agent definition to improve, complete, or adapt
• optimization-spec (optional) — specific content changes requested by bundles-forge:optimizing (e.g., rewrite description, reduce tokens, restructure sections)

Outputs

• skill-content — completed or improved SKILL.md files following authoring conventions (frontmatter, description, body structure, token efficiency)
• agent-content (optional) — completed agents/*.md definitions following agent authoring conventions

Integration

Called by:

• bundles-forge:blueprinting — dispatched as Phase 2 (content writing) in the blueprinting pipeline (write all skills + agent definitions)
• bundles-forge:optimizing — content rewriting for description, token, and structural improvements
• User directly — for standalone SKILL.md or agent definition authoring

Pairs with:

• bundles-forge:scaffolding — scaffolding generates directory structure, authoring fills it with content
• bundles-forge:auditing — auditing validates authored content for quality, cross-references, and security