working-with-skills

Working with PostHog skills

This skill teaches agents how to use the llma-skill-* MCP tools well — minimum context, minimum round-trips, minimum mistakes. If you are not yet familiar with the tool surface itself, read the skills-store skill first for the catalog. This document is about how to choose between the tools and how to scale the workflow when skills get big.

Operating principles

1. Progressive disclosure is non-negotiable. Lists return descriptions, get returns body + manifest, file-get returns one file. Never preload bundled files "just in case" — every preloaded script is wasted context for the actual task.
2. Pick the smallest write primitive that does the job. A targeted edits or file_edits is cheaper, safer, and clearer in version history than a full body or full bundle replacement.
3. Reads are cheap; concurrent overwrites are not. Always have a recent version from llma-skill-get (or from the response of the previous write) before calling any write tool, and pass it as base_version.
4. Authoring follows the Agent Skills spec. Keep name kebab-case, descriptions trigger-rich, body short, bulky material in bundled files.

Decision tree: which tool do I call?

Need to know what's available?
  └─► llma-skill-list                    (names + descriptions only)

Need to use / inspect a specific skill?
  └─► llma-skill-get                     (body + file manifest, NO file contents)
        └─► llma-skill-file-get          (one file, on demand, only as referenced)

Authoring a brand new skill?
  └─► llma-skill-create                  (body + all initial files in one call)

Editing an existing skill?
  ├─ Body change?
  │    ├─ Substantial rewrite ............. update(body=...)
  │    └─ Surgical tweak .................. update(edits=[{old, new}, ...])
  ├─ Bundled file content change?
  │    └─ update(file_edits=[{path, edits:[...]}, ...])

If you find yourself reaching for update(body=...) plus a sprawling files=[...] to change one paragraph and one script, stop — that's two narrower calls (update(edits=[...]) plus update(file_edits=[...])) or even a single update carrying both edits and file_edits.

Discover before you fetch

llma-skill-list is the right tool to "find a skill" — it returns names and descriptions only. Reading the descriptions is the entire point: pick the right skill before pulling any body. If search doesn't narrow it enough, list without it and scan, but do not start fetching candidate bodies blindly.

llma-skill-get should be called once per skill per task, not per question. Cache the body in your working memory; fetch again only if you suspect the skill changed under you (e.g. a 409 on write — see "Concurrency" below).

Reading a large skill efficiently

Big skills (long body, many bundled files) are the case where lazy loading matters most.

1. llma-skill-get(skill_name=...) — read body + files[] manifest.
2. Scan the body's table of contents / headings. The body should already tell you which file goes with which task — that's why bodies stay short and reference files by path.
3. For each file the body explicitly points at for the current task, call llma-skill-file-get(file_path=...). Skip everything else.
4. If the body references "see scripts/X for the rare case Y" and you are not in case Y, do not fetch scripts/X.

When in doubt, fewer files. You can always fetch one more on the next turn.

Authoring a new skill

Use a single llma-skill-create call with body and initial files — the skill lands at version: 1 complete. Do not create the skill empty and then make N follow-up llma-skill-file-create calls; that's N extra versions and N extra round-trips for no benefit.

Authoring rules of thumb

• description is the discovery surface. It is the only thing llma-skill-list returns. Make it trigger-rich (what the user might say) and scope-honest (what the skill does and does not do).
• name — kebab-case, max 64 chars, no leading/trailing/consecutive hyphens. The spec validator rejects anything else.
• Body ≤ ~500 lines. Long preambles, exhaustive SQL, full example payloads, and runnable code belong in references/, assets/, or scripts/. The body should route to those files, not inline them.
• File layout convention — scripts/ for executable code, references/ for prose docs and examples, assets/ for templates / data. Agents can rely on this for orientation when they only have the manifest.
• allowed_tools lists the MCP / built-in tools the skill expects to be callable. Be honest — under-declaring causes silent failures, over-declaring is a security smell.

Updating an existing skill

The single most common mistake is using update(body=..., files=[...]) for a small change. That works, but it round-trips the entire skill, makes the diff unreadable in version history, and risks dropping files if files was incomplete. Use the smallest primitive instead.

Always read first, capture version

Note the returned version — pass it as base_version on every write. After a successful write, the response contains the new version; chain further writes with that.

Body: full replacement vs incremental edits

Full replacement when you are restructuring the body:

Incremental edits when you are tweaking a few lines (preferred for small changes — easier to review, lower error surface):

Each edits[].old must match exactly once in the current body, and body and edits are mutually exclusive in one call.

Bundled file content edits

file_edits patches one or more existing files in place — non-targeted files carry forward unchanged. This is the right primitive when you are tweaking script logic or fixing a typo in a reference doc:

file_edits cannot add, remove, or rename files — only patch existing ones. For structural changes, use the per-file tools.

Combining edits in a single call

You can combine edits (body) and file_edits (existing files) in one llma-skill-update call to publish a single coherent version when a change spans both:

File-path parameter naming (read this before guessing)

The same concept — a bundled file's path — is named differently depending on where it travels in the request, and this trips up agents working from memory. There is one rule:

• file_path — when the path is part of the URL (llma-skill-file-get, llma-skill-file-delete). These read/delete one file addressed by its path.
• path — when the path is a body field: llma-skill-file-create, the files=[{path, content, content_type}] array, and file_edits=[{path, edits}].
• old_path / new_path — body fields on llma-skill-file-rename.

Mnemonic: path is the field name on a file object (it sits next to content), so everything that carries a file object uses path; the two tools that address a file by URL use file_path. When unsure, check the tool's input schema rather than guessing — passing path to file-get yields a /files/undefined/ 404.

Adding, removing, renaming files

Each is its own call, each publishes a new version:

llma-skill-file-rename is a true move — it carries the existing content forward without resending it. Always prefer it over delete + create when the content is unchanged.

When to use update(files=[...]) (rare)

Passing files to llma-skill-update replaces the entire bundle — anything not in the array is dropped. This is the right tool only when you are intentionally wiping and reseeding the bundle (e.g. importing a fresh local SKILL.md tree). For almost every other case, prefer file_edits plus per-file CRUD.

Working with large multi-file skills

Skills with many files (10+) require extra discipline:

• Treat the manifest as the index. llma-skill-get's files[] is your map. Match each task step to one file and fetch only that one.
• Group structural changes into a sequence, not a fork. If you are renaming three files, do them sequentially: rename → rename → rename, each chained via the previous response's version. That gives you three small reviewable versions instead of one giant update(files=[...]) blob.
• Keep edits localised. A single llma-skill-update with file_edits targeting five files is fine. A single update(files=[...]) carrying ten full file bodies is almost always a sign you should have used file_edits.
• Refactor the body itself first. If the body has grown past ~500 lines, the right next step is usually to split content into new bundled files before adding more material.

Concurrency: base_version

Every write tool accepts base_version. Always pass it.

• The server compares base_version to the current latest version. If they match, the write succeeds and the new version is base_version + 1.
• If they differ, the write is rejected (someone else updated the skill). Re-run llma-skill-get, reconcile your changes against the new body, and retry with the fresh version.
• After a successful write, the response includes the new version. Chain further edits with that — do not re-get between back-to-back writes you control.

Skipping base_version does not speed things up — it just turns a clean "someone else won the race" error into a silent overwrite of their work.

Common pitfalls

• Calling llma-skill-list with no search and then fetching every body — defeats progressive disclosure. Read the descriptions first.
• Pre-fetching every bundled file after llma-skill-get — same mistake on the inner level. Fetch on demand from the body's directives.
• Using update(body=..., files=[...]) for a one-line fix — round-trips the entire skill, makes diffs unreadable, and risks dropping files. Use edits / file_edits.
• Using update(files=[...]) when you meant to add one file — drops every file you didn't include. Use llma-skill-file-create instead.
• Delete + create instead of rename — loses content history and costs an extra version bump.
• Stale base_version after chained writes — read the version from the previous write's response, not from your initial get.
• Leaving base_version off — accepts a silent overwrite. Always include it once you've done a get.
• Empty / vague description — the skill becomes effectively undiscoverable via llma-skill-list search. Treat the description as the trigger contract.
• Long body + no bundled files — when a body crosses ~500 lines, refactor into references/ and scripts/ rather than letting it grow.
• Mixing body and edits in one update call — they're mutually exclusive. Pick one.
• Guessing path vs file_path — file-get and file-delete take file_path (it's in the URL); create, rename (old_path/new_path), files, and file_edits take path (it's a body field). See "File-path parameter naming" above.

Archiving a skill

llma-skill-archive hides every active version of a skill by name. It is not version-scoped and cannot be undone — the skill drops out of llma-skill-list and llma-skill-get for the whole team.

Before archiving, llma-skill-get the skill if you need to inspect or copy it first. Archiving is the right tool for retiring a skill entirely; to remove a single bundled file use llma-skill-file-delete, and to roll back content publish a new version rather than archiving.

Porting a local SKILL.md tree into PostHog

When migrating a local skill folder (e.g. my-skill/SKILL.md plus scripts/, references/, assets/):

1. Read the local SKILL.md. Its frontmatter maps to name, description, license, compatibility, allowed_tools, metadata. The body after the frontmatter becomes body.
2. Walk the bundled subdirs and gather every file as { path, content, content_type }.
3. Call posthog:llma-skill-create once with everything — the skill lands at version: 1 complete. Do not split this into a create + N file-create calls.

After the create, the skill is live for everyone via llma-skill-get.

When a skill is the wrong answer

Not every persistent prompt belongs in the skills store:

• One-off task instructions belong in the conversation, not in a skill.
• Personal scratchpads belong in agent memory or local files.
• Code is not a skill — if it's something a service runs, it belongs in the repo.

A good skill is reusable, discoverable by description, and worth the cost of keeping it correct over time.