Samples Curation

Samples Curation Skill

You are helping the owner capture and annotate gold-standard content examples. These examples are saved to the samples file registered in the project's ## Context files table (default: context/samples.md) and used by output-format skills as high-signal reference material for on-brand content generation.

Step 0: Recall learnings

If .claude/learnings.md exists, read it silently. Apply all entries relevant to this run. Do not announce this step. If the file is absent, continue normally.

Step 1: Identify the content

Ask the owner:

"Please share the content you'd like to save as a gold-standard example. You can either paste it here, or give me a file path and I'll read it."

Wait for the response.

• If the owner pastes content, acknowledge it and proceed to Step 2.

• If the owner provides a file path, read the file:

  cat "<path>"
  

  Confirm you've read it, then proceed to Step 2.

Step 2: Annotation interview

Walk through these questions one at a time:

1. Output type — "What type of content is this? (e.g., LinkedIn post, blog article intro, email subject line, whitepaper section)"

2. Key qualities — "What makes this example good? Describe the qualities you want future content to replicate. (You can skip this if you prefer.)"

3. Caveats — "Any caveats or context? For example: 'works for B2B, not B2C', or 'written for a product launch, tone may be more urgent than usual'. (Optional.)"

If the owner skips a question, omit that field from the entry — do not add placeholders.

Step 3: Preview and confirm

Present a formatted preview of the entry before saving:

─────────────────────────────────────────────
Sample entry preview
─────────────────────────────────────────────
Output type: <type>
Key qualities: <qualities>          ← omitted if skipped
Caveats: <caveats>                  ← omitted if skipped

Content:
<content>
─────────────────────────────────────────────

Ask: "Shall I save this entry to the samples file? (yes / no / edit)"

• Yes → proceed to Step 4.
• No / cancel → say "Entry discarded. Nothing was written." Then go to Step 5.
• Edit → ask what to change, apply the edit, show the preview again.

Step 4: Append to samples.md

Determine the target file: check the ## Context files table in CLAUDE.md for an existing row whose summary indicates it holds curated content samples. If found, use that file's path. Otherwise, default to context/samples.md.

Check whether the target file exists:

ls "<target-path>" 2>/dev/null && echo "exists" || echo "missing"

If missing, create it with a header first:

# Content Samples

Gold-standard content examples curated by the owner.
Skills reference this file to produce on-brand output.

---

If the file was newly created and isn't already registered, mention it to the owner and suggest running /register-context to register it in the ## Context files table.

Then append the entry in this format:

## Sample: <output-type> — <YYYY-MM-DD>

**Output type:** <type>
**Key qualities:** <qualities>
**Caveats:** <caveats>

Omit `**Key qualities:**` and `**Caveats:**` lines if the owner skipped those fields.

After writing, confirm:
> "✓ Saved to `<target-path>`."

Count the total `## Sample:` entries in the file and report:
> "The file now contains <N> sample(s)."

## Step 5: Loop

Ask: "Would you like to curate another example? (yes / no)"

- **Yes** → return to Step 1.
- **No** → proceed to Step 6.

## Step 6: Store learnings

Evaluate the full curation session — not individual loops — for patterns worth storing.

For each qualifying observation, append one tagged line to `.claude/learnings.md`
(create with standard header if missing):

```text
[cc-content:samples-curation] <concise observation> — <YYYY-MM-DD>

Qualifies: patterns discovered across the session — e.g., consistent length or format preferences in the samples selected, recurrently skipped annotation fields, or constraints the owner mentioned about what counts as a good sample — that are not already captured in any context/ file, CLAUDE.md, or any existing .claude/learnings.md entry under any plugin tag.

Does not qualify: individual sample content; facts about what was saved (those are in the samples file); standard curation behavior applied without deviation; facts semantically equivalent to an existing entry under any plugin tag — when in doubt, skip; redundancy is worse than a missed entry.

Check for the file before appending:

ls .claude/learnings.md 2>/dev/null && echo "exists" || echo "missing"

Standard header when creating the file:

# Learnings

Corrections and feedback collected during content sessions.
Entries are tagged by skill and dated.

---

If any entries were written, confirm: "✓ N learning(s) saved to .claude/learnings.md." If nothing qualified, skip silently.

Then say: "Samples curation complete. You can invoke this skill again any time to add more."