SciAgent Skill Creator

Repo-local scaffolder for skills/ entries. Mechanizes the boilerplate from CLAUDE.md Steps 1, 2, 4, 5, 6 so authoring effort stays on content (When to Use, Workflow, Recipes, References) and not on field plumbing.

When to invoke this skill

• User asks for a new SciAgent skill entry on a specific tool, library, database, or guide topic
• User invokes /sciagent-skill-creator directly
• The agent is about to hand-edit registry.yaml and create a skills/<cat>/<name>/SKILL.md from scratch — use this instead

Do not invoke for:

• Editing an existing entry's content (just edit the file)
• Migrating an existing entry (read CLAUDE.md "Migrating from Existing Entries" first — the scaffolder generates a skeleton, but migration requires content judgment)
• Updating registry.yaml only (use a normal edit)

What you need to collect from the user

Before calling the scaffold script, gather these — in conversation, not via flags hidden from the user:

1. Topic — concrete tool/library/concept name. Reject vague topics ("ML stuff") with a clarifying question.
2. Sub-type — pipeline | toolkit | database | guide. Use the decision rule from CLAUDE.md Step 1b. If unsure, ask the user.
3. Category — primary category directory. List the table from CLAUDE.md Step 2 if the user is unsure.
4. Entry name — kebab-case slug. Convention: {tool-name}-{purpose} (e.g., pydeseq2-differential-expression). Confirm with the user.
5. License — underlying tool's license. Default to CC-BY-4.0 for original prose-only content.
6. Description — 1-2 sentences, max 1024 chars. Lead with tool/domain keyword in the first 120 chars. Anti-patterns are in CLAUDE.md Step 5 "Description writing rules".
7. Tags (optional) — only if the entry meaningfully spans multiple categories (e.g., literature DB stored under scientific-writing, tag with ["databases", "literature"]).

Duplicate check before scaffolding

Before calling the scaffold script, search the registry and legacy/ for similar names:

grep -i "<topic-keyword>" registry.yaml
ls legacy/ | grep -i "<topic-keyword>"

If a near-duplicate exists, surface it to the user before continuing. Authoring a parallel entry usually means the existing one needs updating, not duplication.

How to run the scaffolder

Call scripts/scaffold.py with explicit arguments. The script is non-interactive — the agent provides all values:

python .claude/skills/sciagent-skill-creator/scripts/scaffold.py \
  --sub-type pipeline \
  --category genomics-bioinformatics \
  --name my-tool-purpose \
  --description "MyTool short-form description starting with the tool name. Brief on inputs, outputs, when to pick this over alternatives." \
  --license MIT \
  --tags databases,literature   # optional, comma-separated

Behavior:

1. Validates name (kebab-case, not already in registry.yaml, not in legacy/)
2. Validates category exists as a directory under skills/
3. Validates description with validate_description.py (length + first-120-char keyword lead)
4. Validates tags (kebab-case if provided)
5. Creates skills/{category}/{name}/SKILL.md from the matching template, substituting frontmatter fields
6. Appends a new entry to registry.yaml with date_added = today (UTC)
7. Runs pixi run validate to confirm the registry is still well-formed
8. Prints next steps (fill in Overview, Workflow, Recipes, References)

On any validation failure, the script aborts without writing anything. Fix the offending value and re-run.

After scaffolding

The generated SKILL.md is a skeleton with placeholders. The agent's remaining job:

1. Fill Overview, When to Use, Prerequisites, Workflow / Core API / Key Concepts, Common Recipes, Troubleshooting, References
2. Match the section structure required by the sub-type (see CLAUDE.md Step 4 format rules)
3. Run pixi run test — full suite, not just validate — to catch sub-type-specific structural failures (code block counts, table row counts, section presence)

The scaffold script does not pretend to write content. Content stays with the agent and the source material.

Content authoring rules (what NOT to bake into a SKILL.md)

Skills document a tool's analysis surface, not the consumer's house style. A SKILL.md is read by many agents for many downstream tasks — visual choices that fit one analysis brief leak into every future invocation. Strip the following before committing:

• Color palettes, cmaps, themes — no hex codes (#08306b), no LinearSegmentedColormap.from_list(...), no ListedColormap([...]), no prescribed cmap= arguments unless the cmap is the tool's API (e.g., a tool that ships its own palette). Let matplotlib pick defaults; the consumer overrides downstream.
• Per-replicate / per-condition color dicts — e.g., colors = {"rep1": "#1f77b4", ...}. Matplotlib auto-cycles colors.
• Font choices, dpi presets, figure sizes tuned for one report — figsize=(8, 4) for a routine line plot is fine; figsize=(12, 4) chosen to fit a slide deck is not.
• One-shot user-brief specifics — if the user asked for "blue for low, red for high" in their analysis, that belongs in their code, not the skill. The skill teaches how to compute phi/psi density; how to color it is consumer choice.
• Hardcoded paths beyond the tool's defaults — "figures/", "results/", f"{pdb_id}_protein.pdb" are fine as illustrative outputs; "/Users/me/proj42/output" is not.

What to keep: the analysis logic, the data shape, the units, the parameter semantics, the expected output structure (columns, axes, units), and any visual choice the tool itself enforces.

Rule of thumb: if a downstream consumer would override the choice, don't ship the choice in the skill.

Writing style: be succinct

A SKILL.md is reference material for agents, not a tutorial. Token cost matters — every line is paid for on every retrieval. Write like documentation, not like a walkthrough:

• One sentence per section intro, not a paragraph. Drop hedging ("typically", "in general", "you may want to"), filler ("note that", "it's worth mentioning"), and softeners. State the rule.
• Comments inside code blocks earn their place. Only annotate non-obvious lines — units, gotchas, why this parameter. Don't restate what the code already says (# load the trajectory above traj = md.load(...) is noise).
• Code over prose when possible. A four-line code block beats a paragraph describing the same call. The agent runs the code; the prose is for what the code can't show.
• No preamble before code. "The following snippet demonstrates how one might..." → delete. The header and code speak for themselves.
• No closing recap. Each section ends when the information ends. No "In summary..." or "As shown above...".
• Cut redundancy across sections. If Workflow Step 2 already shows the parameter, the Key Parameters table row doesn't need to re-explain it — just list it.
• Tables for enumerations, not bullets. Parameters, troubleshooting, codes → table. Prose lists waste vertical space.

Target density: a reader scanning the file should reach the next code block within ~5 lines of prose. If a section's prose is longer than its code, tighten the prose.

Files in this skill

• SKILL.md — this file (when/how/what)
• scripts/scaffold.py — non-interactive scaffolder (create files, append registry, run validate)
• scripts/validate_description.py — description linter (length + first-120-char keyword rule). Reused by scaffold.py and standalone.

Failure modes to surface to the user

• Name already exists → suggest a different suffix
• Category not in list → present the category table and ask
• Description starts with stop-verb (Use, A, An, The, Query) → rewrite leading with the tool name
• Description too long → trim disambiguation tail; keep keyword carrier