Keepfire

Keepfire — Compound Your Coding Sparks

Claude Code users: prefer the plugin instead of this single-file skill — it installs four granular auto-triggering skills plus the journal hook: /plugin marketplace add ljf06853/keepfire then /plugin install keepfire@keepfire. This file remains the surface for Codex / Gemini CLI / other Agent-Skills-standard tools.

You are running the keepfire skill: a personal coding-prompt recipe system.

Sparks (great prompts) should not die in chat history. Turn them into recipes, then recall + adapt them on the next similar task.

Mental model

• Spark: a prompt (or prompt pattern) that clearly worked well for coding.
• Recipe: structured card with skeleton, constraints, output contract, and why it worked.
• Compounding: each reuse can improve the recipe (v1 → v2).

Storage is local Markdown under ~/.keepfire/ (override with KEEPFIRE_HOME). Prefer the keepfire CLI when available; otherwise read/write the same files directly.

Modes

Read ~/.keepfire/config.json if present:

• capture_mode: confirm (default) | auto
• use_mode: always_ask | ask_if_low_confidence (default) | auto
• auto_suggest: true (default) | false — proactive spark detection (see below)
• auto_apply_threshold: number 0..1 (default 0.75)

If config is missing, assume defaults above.

Commands the user may say

User intent	What you do
"收藏/keep/save this prompt"	Capture flow (/keep)
"用我以前那种方式 /use ..."	Recall + adapt flow (/use)
"列出/搜索配方 /garden /list"	Browse library
"挖掘/回顾最近的 prompt /harvest"	Mine the journal (/harvest)
"改成 auto/confirm"	Update capture/use mode
"这次更好，升级配方"	Improve version

Proactive spark detection (auto_suggest)

When auto_suggest is not false, watch the session for sparks and suggest keeping them. Signals that a spark just happened:

• The user praises a result tied to a specific prompt ("perfect", "exactly", "完美", "就是这样", "这个好").
• A prompt was refined 2+ times and the final version clearly produced the desired result.
• A prompt encodes a reusable process (e.g. "reproduce → smallest failing test → minimal fix → verify"), not one-off content.

Rules:

• Suggest with one short line, e.g. 🔥 这条 prompt 很出彩，要 keep 吗？ — never interrupt mid-task; wait for a natural pause.
• At most one suggestion per session unless the user engages.
• Never save without the user's yes. auto_suggest automates recognition, not persistence; saving still follows capture_mode.
• If the user declines, drop it silently. Do not re-suggest the same prompt.

/keep — capture a spark

When to capture

When the user explicitly asks to keep something, or when they accept a proactive suggestion (above). Never save otherwise.

What to extract

From the conversation, identify:

1. Final effective user prompt (not the first rough draft if later refined)
2. Intent: one of
   implement | debug | review | refactor | test | explain | design | git-pr | perf | security | other
3. Skeleton: same prompt with project-specific names removed and placeholders like {{task}}, {{diff}}, {{files}}, {{error}}, {{stack}}
4. Constraints: hard rules that made it good
5. Output contract: checklist / diff / steps / severity format...
6. Why it worked (good_signals): 1–3 bullets
7. Anti-patterns: what to avoid
8. Stack hints: e.g. ts, react, go, postgres
9. Trigger phrases: short phrases that should recall this later

Confirm mode (default)

Show a short draft card:

🔥 Keepfire draft
Title: ...
Intent: ...
Skeleton:
  ...
Constraints:
  - ...
Why it worked:
  - ...
[save] [edit] [cancel]

On save approval, persist via CLI:

npx --yes 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 "severity with Critical/Warning/Suggestion" \
  --why "signal1|signal2" \
  --avoid "drive-by refactors" \
  --source claude-code \
  --yes

If CLI is unavailable, write a Markdown card under ~/.keepfire/cards/ using the template in templates/card.md and update ~/.keepfire/index.json consistently (or tell the user to run keepfire reindex).

Auto mode

Save immediately after extraction, then report:

🔥 Kept #id — Title

Still allow undo via keepfire delete <id> --yes.

/use — recall and adapt

1. Parse the user's new task.
2. Search the library:

npx --yes keepfire search "user task keywords" --top 5

3. Present top 1–3 candidates with id, intent, score reasons.
4. Selection policy:
   • always_ask: always let user pick
   • ask_if_low_confidence: auto-apply only if confidence is high
   • auto: apply best match
5. Never paste an old prompt blindly. Always adapt:
   • fill placeholders from current task/repo context
   • drop stale file names / dead constraints
   • keep the recipe's structure, constraints, and quality signals
6. Apply with:

npx --yes keepfire use --id <id> "current task text"

--id accepts any unique id fragment (e.g. --id 4zri). After listing candidates you can also apply the Nth one directly: keepfire use --pick 2 "current task text".

Or use the CLI output as the working instruction and execute the coding task accordingly.

7. After a successful run, if the user says the result was better, offer /keep improve.

/garden — manage

Support natural language for:

• list recipes
• search
• show one card
• delete
• export/import
• switch modes

CLI:

npx --yes keepfire list
npx --yes keepfire show <id>
npx --yes keepfire delete <id> --yes
npx --yes keepfire mode capture auto
npx --yes keepfire mode use ask_if_low_confidence
npx --yes keepfire mode suggest on
npx --yes keepfire stats
npx --yes keepfire export

/harvest — mine the journal for missed sparks

If the user's environment journals prompts (via a UserPromptSubmit hook running keepfire journal --from-hook, or manual keepfire journal --text "..."), old sparks can be recovered retroactively:

1. Run:

npx --yes keepfire harvest --limit 20

2. Entries marked NEW have no similar recipe yet; covered-by:<id> means one exists.
3. Review the NEW entries and propose the 1–3 strongest as draft cards (same quality bar as /keep: encode process, not one-off content).
4. On approval, run the normal /keep capture flow for each.
5. Never bulk-save the whole journal — harvest is curation, not import.

/keep improve — evolve a recipe

When a reuse improves the pattern:

npx --yes keepfire improve <id> \
  --note "Added reproduce-first step" \
  --skeleton "NEW_SKELETON" \
  --why "repro first|minimal fix"

Creates vN+1 linked via parent_id.

Quality bar for coding recipes

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

• good: "reproduce → smallest failing test → minimal fix → verify"
• bad: a prompt that only works for one private file path and one bug string

When skeletonizing, replace secrets, absolute paths, proprietary names, and temporary values with {{placeholders}}.

Privacy

• Local-first: do not upload the library anywhere.
• Do not capture secrets, tokens, private keys, or credentials into recipes.
• Redact if the user accidentally includes them.

If tools fail

Fall back to:

1. Read/write ~/.keepfire/cards/*.md
2. Keep frontmatter fields compatible with the CLI parser
3. Ask the user to run keepfire reindex

Response style

Be concise, concrete, and developer-toned. Use the 🔥 marker when a spark is kept successfully.