memoforge / style skill

You manage the user's custom style profiles. A profile is a directory under ~/.claude/plugin-data/memoforge/profiles/<name>/ containing prose-style.md (always), template.md (sometimes), meta.json, and supporting files. The memo skill reads these at Phase 1.5 when the user picks a profile for a new memo.

mf = ${CLAUDE_PLUGIN_ROOT}/scripts/mf. If that path is not executable in this host (Windows without a POSIX shell), call ${CLAUDE_PLUGIN_ROOT}/scripts/mf.cmd with the same arguments; every example below is written with <mf> standing for whichever of the two ran.

Authority hierarchy (highest wins, same shape as memo skill):

1. Cowork / Anthropic platform policy.
2. This skill and its arguments.
3. mf style <subcommand> — the canonical write path for profiles. Never bypass it.
4. User input from AskUserQuestion or text replies.

Key invariant. All user-facing strings (menu items, checkpoint questions, warnings, summaries) are in the user's language (the language they write in). The contents of generated prose-style.md and template.md are in the input language (English, Russian, etc.) — that is the extractor's concern, not yours.

Parse $ARGUMENTS

Split $ARGUMENTS by whitespace. The first token is the action:

• new <name> [flags…] — create a new profile. See §new below.
• list — list existing profiles. See §list.
• use <name> or use --clear — set or clear the default profile. See §use.
• show <name> — print profile contents. See §show.
• delete <name> — delete a profile. See §delete.
• (no action / empty $ARGUMENTS / menu / help) — go to §menu (interactive).

For any other first token, print: Unknown action. Use new | list | use | show | delete, or run /memoforge:style with no arguments for the interactive menu. End turn.

menu — interactive entry point

When $ARGUMENTS is empty, ask the user what they want to do. Use AskUserQuestion with these options (in the user's language; translate on the fly, profile names stay as typed):

• Question: "What would you like to do?"
• Header: "Style" (≤12 chars).
• multiSelect: false.
• Options:
  • label: "Create a new profile", description: "From example memos, written rules, or both"
  • label: "List existing profiles", description: "See your saved profiles and the current default"
  • label: "Set default profile", description: "Choose which profile /memo uses by default"
  • label: "Show profile contents", description: "Print prose-style.md and template.md of a profile"
  • label: "Delete a profile", description: "Remove a profile permanently"

If the host has no AskUserQuestion, fall back to a plain-text prompt listing the same options and end the turn. The user then re-runs /memoforge:style <action> ... with the chosen action.

Branch on the answer and continue inline to the matching section below (§new, §list, etc.).

new — create a new profile

Step 1 — Collect profile name

If <name> is missing from $ARGUMENTS, ask via AskUserQuestion:

• Question: "Profile name (lowercase, dashes, no spaces — for example, my-firm)"
• Header: "Name"
• multiSelect: false.
• Single option with description "Type the profile name as 'Other'". (AskUserQuestion always offers an Other / free-text input.)

If AskUserQuestion is unavailable, print the prompt as text and end the turn.

Validate the name:

<mf> style validate-name "<name>"

If exit code is non-zero, print the stderr message and re-ask the name (or end turn if no AskUserQuestion).

Also check that the name does NOT already exist. Read the list:

<mf> style list

If the JSON output contains a record with this name, ask:

• "Profile <name> already exists. Overwrite or cancel?"
  • "Overwrite" → delete the existing profile via delete <name> and proceed.
  • "Cancel" → end turn.

Step 2 — Determine input type

Parse the remaining $ARGUMENTS flags:

• --examples <path-or-paths> — one or more whitespace-separated paths. Can also be a directory; expand to its contents.
• --rules <text-or-path> — either inline text (typically quoted) or a path to a .md/.txt file. Distinguish: if the value is a path that exists on disk, treat as path; else treat as inline text.

input_type is computed from what was provided:

• Both flags set → both
• Only --examples → examples
• Only --rules → rules

• Neither flag set → ask via AskUserQuestion:

  • Question: "How would you like to define this profile?"
  • Header: "Input"
  • multiSelect: false.
  • Options:
    • label: "Examples only", description: "Provide paths to one or more example memos"
    • label: "Rules only", description: "Type or paste style rules as plain text"
    • label: "Both examples and rules", description: "Examples set the baseline; rules override on conflict"

  After the answer, ask for the missing input(s):

  • For examples: "Paths to example memos (space-separated, or a directory):" via AskUserQuestion (free-text via Other).
  • For rules: "Provide rules as text (paste below) or path to a .md/.txt file:" via AskUserQuestion (free-text via Other).

  Without AskUserQuestion: print the prompts as text, end the turn, and have the user re-invoke /memoforge:style new <name> --examples ... --rules ... with the values.

Step 3 — Validate paths (if examples were provided)

For each example path: check that the file exists — Read it, or Glob its directory; do not shell out to a POSIX test. If a path does not exist, print:

error: example path not found: <path>

and end the turn (no profile is created).

If --examples was a directory, expand to all .md, .txt, .pdf, .docx files under it (use Glob).

If --rules was a path, validate similarly.

Step 4 — Dispatch style-extractor

Print a one-line heads-up to chat: Extracting style profile '<name>' — this takes ~1 minute…

Dispatch the extractor via Agent. Substitute the real values before sending — the prompt is plain text, so it carries the expanded absolute path of the launcher chosen above (${CLAUDE_PLUGIN_ROOT}/scripts/mf, or ${CLAUDE_PLUGIN_ROOT}/scripts/mf.cmd on Windows), never the <mf> placeholder:

Agent(
  subagent_type="memoforge:style-extractor",
  prompt="""
  Extract a style profile.
  - profile_name: <name>
  - examples: <space-separated list of absolute paths, or empty>
  - rules: <inline text OR path to file, or empty>
  - input_type: <examples|rules|both>
  - work_dir: <a writable temp directory; create $TMPDIR/style-extract-<name> if needed>
  - mf: <absolute launcher path — run the `mf style` commands of your spec through it>

  Follow your agent spec in agents/style-extractor.md exactly.
  Init the profile via `mf style init-profile` first; write prose-style.md (always);
  write template.md only if structural input is present; write rules.md if rules were provided;
  copy examples into sources/; atomically write the final meta.json at the end.
  Return a ≤200-word summary in the user's language with any warnings.
  """
)

Wait for the extractor to return. Print its summary to chat verbatim (it already arrives in the user's language and concise).

Step 5 — Validate the written profile

After the extractor returns, validate the result:

<mf> style validate-profile "<name>"

If exit code is non-zero, print: Profile validation failed — the profile directory is malformed and was not registered. Please retry or report this issue. Then run delete <name> to clean up. End turn.

Step 6 — Offer to set as default

Ask:

• Question: "Make <name> the default profile for new memos?"
• Header: "Default"
• multiSelect: false.
• Options:
  • label: "Yes — make it default", description: "Next /memoforge:memo will preselect this profile"
  • label: "No — keep current default", description: "You can change later with /memoforge:style use <name>"

If "Yes": run <mf> style set-default "<name>". Print confirmation.

If "No": print a one-line note: Profile saved. Use /memoforge:style use <name> to select it later.

End turn.

list — print profiles table

Run:

<mf> style list

Parse the JSON and render a plain-text table in chat:

Profile          | Created    | Input    | Template | Default | Lang
-----------------|------------|----------|----------|---------|-----
my-firm          | 2026-05-25 | examples | yes      | ✓       | en
acme-rules       | 2026-05-20 | rules    | no       |         | en

If the list is empty, print: No profiles yet. Create one with /memoforge:style new <name>.

If any profile has valid=false, print a warning line below the table: ⚠️ <name>: <first error message>. Suggest re-running /memoforge:style delete <name> and creating fresh.

End turn.

use — set or clear default

If $ARGUMENTS is use --clear:

<mf> style clear-default

Print: Default profile cleared. /memoforge:memo will offer all profiles next time. End turn.

Otherwise, <name> is the second token. Validate and set:

<mf> style set-default "<name>"

If exit code is non-zero, print the stderr message (profile not found: <name> is the common case — suggest /memoforge:style list).

On success, print: Default profile set to '<name>'. /memoforge:memo will preselect it for new memos. End turn.

show — print profile contents

<name> is the second token. Validate it exists:

<mf> style validate-profile "<name>"

If exit code is non-zero, print the stderr message and end turn.

Print meta.json content:

<mf> style read-meta "<name>"

Then print the first 30 lines of prose-style.md and (if present) the first 30 lines of template.md, each with a header:

=== prose-style.md (first 30 lines) ===
<content>

=== template.md (first 30 lines) ===
<content>

(Open the full files at <profile_dir>.)

Use Read with offset=0, limit=30 for the lines; the profile dir path comes from:

<mf> style resolve-paths "<name>"

End turn.

delete — remove a profile

<name> is the second token. Check if it is currently the default:

<mf> style get-default    # -> {"default": "<name>"|null}

If the default field of that JSON equals <name>, ask first:

• Question: "<name> is the current default. Delete and clear default?"
• Header: "Confirm"
• Options:
  • label: "Yes — delete it", description: "Profile is removed; default is cleared; /memo will offer all remaining profiles next time"
  • label: "No — keep it", description: "Cancel deletion"

If "No": end turn.

Otherwise (or if not the default), run:

<mf> style delete "<name>"

The script atomically removes the directory AND clears the default file if <name> was the default.

Print confirmation: Profile '<name>' deleted. End turn.

Hard constraints

• Never write to the plugin data directory (profiles/) by hand — always go through mf style. It is the canonical write path: it validates names, writes atomically, and keeps the default-file consistent on delete.
• Never modify state.json of an in-flight memo task. This skill manages user-level style profiles only — it has no relationship to any specific memo task in progress.
• Never call Agent for anything other than style-extractor from this skill. Subagents for the memo pipeline are dispatched by skills/memo/SKILL.md, not from here.
• All user-facing strings (chat output, AskUserQuestion text) are in the user's language (the language they write in). The studio has no state and no pack: translate its AUQ options on the fly, profile names stay as typed. Profile body content language is decided by the extractor based on the inputs.