Capturing sparks

Capturing Sparks

A spark is a prompt (or prompt pattern) that clearly worked well for a coding task. Sparks die in chat history unless captured as recipes — structured cards the user can recall and adapt on the next similar task.

<HARD-GATE> NEVER save a recipe without the user's explicit yes. Proactive detection automates *recognition*, not *persistence*. If the user declines, drop it silently and do not re-suggest the same prompt. </HARD-GATE>

When you noticed the spark yourself

Check ~/.keepfire/config.json first: if auto_suggest is false, stay silent. Otherwise suggest with ONE short line at a natural pause — never mid-task:

🔥 这条 prompt 很出彩，要 keep 吗？ / That prompt was elite — keep it?

At most one suggestion per session unless the user engages.

Checklist

Create a task per item and complete in order:

1. Identify the final effective prompt — the version that worked, not the first rough draft.
2. Extract the recipe fields:
   • title — short, recognizable
   • intent — one of implement | debug | review | refactor | test | explain | design | git-pr | perf | security | other
   • skeleton — the prompt with project-specific names replaced by {{task}}, {{diff}}, {{files}}, {{error}}, {{stack}} placeholders
   • constraints — the hard rules that made it good
   • output_contract — expected output shape (checklist / diff / severity levels …)
   • good_signals — 1–3 bullets on why it worked
   • anti_patterns — what to avoid
   • stack_hints, trigger_phrases — for recall later
3. Redact secrets — tokens, keys, absolute private paths, credentials never go into a recipe.
4. Show a draft card and ask: [save] [edit] [cancel]. (Skip the confirmation only if capture_mode is auto in config — the keep intent itself is still required.)
5. Save on approval (see below), then report: 🔥 Kept #<id> — <title>.

How to save

Preferred — the keepfire CLI (if installed; check with keepfire version):

keepfire keep \
  --title "..." --intent review \
  --prompt "RAW_PROMPT" --skeleton "SKELETON" \
  --stack ts,node --tags pr,security \
  --triggers "审PR|security review" \
  --constraints "c1|c2" --output "..." \
  --why "signal1|signal2" --avoid "anti1" \
  --source claude-code --yes

For long prompts write them to a temp file and pass --prompt-file <path>.

Fallback — no CLI: write a Markdown card to ~/.keepfire/cards/<id>.md following the exact heading/frontmatter format in this plugin's templates/card.md (flat inline frontmatter only — the CLI parser does not handle nested YAML), with id YYYY-MM-DD-<slug>-<4 random chars>. Then tell the user to run keepfire reindex when they install the CLI, or update ~/.keepfire/index.json consistently yourself.

Quality bar

Prefer recipes that encode process, not one-off content:

• good: "reproduce → smallest failing test → minimal fix → verify"
• bad: a prompt bound to one private file path and one bug string

If the prompt is too task-specific to generalize, say so instead of saving noise.