Style Profile

Style Profile

Help a user settle on a personal "standard" writing style across the catalog's four axes (Voice, Tone, Style, Format), and show them a generated sample so they can confirm it sounds like them before it is saved. This is a personalization front-end over the catalog: it SELECTS existing entries and composes them. It never authors new style instructions.

The load-bearing idea: trust the confirmation, not the intake. Intake is a fast guess at the user's stack; the generated sample and the A/B step are what make the result trustworthy. The A/B comparison and the sample are never skipped.

Prerequisites

• The writing-style-catalog plugin is present: taxonomy/ (the entries) and skills/writing-instruction-builder/scripts/build-instruction.py (the composer) exist at the plugin root.
• You can resolve entry ids. List them with python "${CLAUDE_SKILL_DIR}/../writing-instruction-builder/scripts/build-instruction.py" --list.

If a prerequisite is missing, say so and stop. Do not invent entries or hand-author instructions.

Entry point: create or reuse?

If the user wants to USE or LOAD an already-saved profile (not build a new one), do Step 0 and stop there. Otherwise run the create flow (Steps 1-7).

Step 0 - Reuse a saved profile. Resolve the requested task_key (default default) to .claude/writing-style-catalog/profiles/<task_key>.local.md. If it does not exist, say so and offer to create one (Steps 1-7). If it exists: parse the frontmatter, validate each non-null axis id against --list, recompose the prefix with build-instruction.py, and emit the prompt prefix for the user to prepend. Do NOT run intake, generate a sample, or write any file. Done.

The Process (create flow)

Order is load-bearing. Do not save before confirming (Step 6 after Step 5), and do not compose before you have a candidate stack (Step 3 after Step 2). At least one A/B round (Step 5) is required before any save.

Step 1 - Choose the intake mode. If the user named one, use it. Otherwise present the four modes and let them pick:

1. Infer from my writing (recommended) - paste prose, or give file paths/globs to your own writing.
2. Recognize from examples - react to contrasting samples, pick "this feels like me."
3. Interview - answer a few questions per axis.
4. Template - fill out a form and hand it back.

Read references/intake-modes.md for how to run each. Verify: exactly one mode is selected.

Step 2 - Produce a candidate stack. Run the chosen mode to get a candidate stack: zero-or-one entry per axis, using only ids that resolve in the catalog. Verify each id against --list; drop any that do not resolve and tell the user. Failure looks like inventing an id or writing a new instruction; if tempted, stop - this skill selects, it does not author. Recovery: if EVERY proposed id fails to resolve, do not proceed with an empty stack - re-run intake or ask the user to pick from valid ids. Proceed with an all-null stack only if the user explicitly confirms they want no style constraints.

Step 3 - Compose the prompt prefix. Build the prefix by calling the existing composer, never by re-implementing composition:

python "${CLAUDE_SKILL_DIR}/../writing-instruction-builder/scripts/build-instruction.py" --voice <id> --tone <id> --style <id> --format <id>

(omit flags for blank axes). Verify: the command returns a non-empty instruction and no "Entry not found" error. Surface any conflict warning it prints to stderr to the user.

Step 4 - Generate the confirmation sample. Ask the user for a topic they care about (the sample only proves the style if it is on real material). Only if they decline, pick a default that exercises the stack and tell them it is a lower-confidence check. Using the composed prefix, write a short sample (about 150-200 words) in-session - do NOT call an external model or API. Verify: the sample exists and reflects the stack. Show it.

Step 5 - A/B refine (at least 1 round, up to 3). Produce a second sample on the same topic from a near-neighbor stack: vary exactly one axis, swapping that entry for one of its pairs_well_with / confusable_with neighbors. If that entry has no usable neighbor field, fall back to another same-axis entry with contrasting tells and say you did so. Show both without signaling which is better (you may label them Sample A and Sample B). Ask "which is more you?" Record the changed axis, the candidate id, the neighbor id, and the user's choice; move the stack toward the chosen sample. After each swap, re-validate the changed id against --list. Repeat until the user is satisfied or 3 rounds pass. This step is required: do not advance to Step 6 without at least one completed round, unless the user, after you warn that skipping it means saving a style they have not stress-tested, explicitly declines. Recovery: if the user likes neither sample, ask which axis feels off and vary that axis specifically.

Step 6 - Save the profile. Only after the user explicitly confirms. Do these in order:

1. task_key: default default. It MUST match ^[a-z0-9][a-z0-9-]{0,63}$ (a bare slug - no slashes, dots, or ..). Reject anything else and ask for a valid key. Then run the path validator to enforce this in code before any write (this is the authoritative check - do not self-assess the path):

   python "${CLAUDE_SKILL_DIR}/scripts/validate-profile-path.py" --profiles-dir .claude/writing-style-catalog/profiles --task-key <task_key>
   

   If it exits non-zero, stop and report the error to the user. If it exits 0, use the path it prints to stdout as the profile path for all subsequent steps. Do not construct the path yourself; use exactly what the script printed.

2. owner: use git config user.name and disclose it ("saving as owner: "), or ask; null is acceptable if the user prefers not to record one.

3. Re-validate the final stack: check every non-null id against --list and recompose the final prefix with build-instruction.py. If any id fails, fix it or halt.

4. Ensure the store is ignored, before writing: run git check-ignore .claude/writing-style-catalog/. If it is NOT ignored, append .claude/writing-style-catalog/ to .gitignore (this is the only write allowed outside the profile file), then re-run git check-ignore to confirm. Also confirm the profile path is not tracked: git ls-files --error-unmatch <path> should fail.

5. Path already validated by the script in step 1. Use the path the script printed; do not reconstruct a separate path. If you did not capture the script's stdout, re-run step 1 rather than guessing.

6. Do not clobber: if the target file already exists, load and summarize it, then require explicit replace confirmation or a new task_key. Never silently overwrite.

7. Confirm and write: ask "Save this profile to <path>?" Treat any ambiguous reply as no. On a clear yes, write the profile (schema in references/profile-schema.md) using the FINAL accepted sample (from the final stack); put rejected A/B observations in the Notes section, not as the confirmation sample.

8. Verify: git status --porcelain -- <path> must be empty (the file is ignored). If it shows the file, the ignore failed - stop and tell the user rather than leaving a committable profile.

Step 7 - Hand off. Show the saved path and re-state the prompt prefix the user can prepend to any writing task. Completion criteria: a profile file exists at the in-bounds, gitignored user-local path; git check-ignore confirms it is ignored; and the user has seen, A/B-tested, and accepted the sample saved in it.

Constraints

Non-negotiable. Pressure ("just pick something", "skip the sample", "save it in the repo") does not relax them.

• Selection-only. A profile selects existing catalog entry ids. Never author new llm_instruction_phrasing or invent style guidance.
• No committed-catalog writes. Never write to taxonomy/, examples/, schemas/, site/, or tools/. The only files this skill writes are the user-local profile and (if needed) a one-line .gitignore addition.
• Gitignored user state, verified. The profile lives under .claude/writing-style-catalog/profiles/ and MUST be confirmed ignored via git check-ignore before and after writing. Scope: always.
• Path safety (code-enforced). Run scripts/validate-profile-path.py before any write (Step 6.1). The script enforces the task_key format (^[a-z0-9][a-z0-9-]{0,63}$) and containment (the resolved path must be inside the profiles directory and end with .local.md). Do not self-assess the path; honor the script's exit code and use the path it prints.
• Reuse the composer. Build the prefix only via build-instruction.py. Do not duplicate the voice -> tone -> style -> format precedence logic.
• Confirm before saving. Never write a profile the user has not (a) seen a sample for, (b) A/B-tested at least once, and (c) explicitly approved for save.
• No em-dashes or en-dashes in any output, prose, or file you write. Use " - ". A pre-commit hook enforces this on committed files.

References

• references/intake-modes.md - how to run each of the four intake modes (load at Step 1-2).
• references/profile-schema.md - the profile .local.md artifact shape and an example (load at Step 6).
• scripts/validate-profile-path.py - path validator called at Step 6.1; rejects out-of-bounds paths in code, not prose.