checkpoint

Save developer context to Goldfish memory — checkpoint at meaningful milestones, not after every action

Safety Notice

This listing is imported from skills.sh public index metadata. Review upstream SKILL.md and repository scripts before running.

Copy this and send it to your AI assistant to learn

Install skill "checkpoint" with this command: npx skills add anortham/goldfish/anortham-goldfish-checkpoint

Checkpoint — Save Developer Memory

When to Checkpoint

When in doubt, checkpoint — a few extra checkpoints are better than lost context.

  • Completed a deliverable — feature slice, bug fix, refactor step
  • Made a key decision — architecture, tradeoffs, approach choices that future sessions must follow
  • Committed or pushed work — natural checkpoint moment after code lands
  • Reaching a stopping point — end of a work session or switching tasks
  • Before context compaction — preserve active state (PreCompact hook handles this automatically)
  • Found something non-obvious — blockers, root causes, discoveries worth preserving
  • User shared requirements/constraints — preserve what future work must honor

Space out checkpoints so each one captures a distinct piece of progress — one per logical milestone is the right cadence.

How to Write Good Descriptions

Your description becomes the markdown body of a .md file. Format it with structure — headers, bullet points, bold, code spans.

The WHAT/WHY/HOW/IMPACT Formula

mcp__goldfish__checkpoint({
  description: "## Implemented JWT refresh token rotation\n\nThe existing single-token approach was vulnerable to token theft.\n\n- **Approach:** Rotate tokens on each use, limiting the attack window\n- **Added:** `RefreshTokenStore` with atomic file writes and 7-day expiry\n- **Tests:** All 12 auth tests passing\n- **Impact:** Unblocks the session management PR",
  tags: ["feature", "auth", "security", "jwt", "refresh-token", "token-rotation", "session-management"]
})

Good vs Bad

GOOD (structured markdown):

## Fixed race condition in checkpoint writes

Concurrent saves could corrupt the daily markdown file.

- **Root cause:** Non-atomic write pattern
- **Fix:** Switched to write-tmp-then-rename with file locking
- **Verified:** Reproduced with parallel test, confirmed fix

BAD (no structure): "Fixed race condition in checkpoint file writes where concurrent saves could corrupt the daily markdown file. Root cause was non-atomic write pattern."

BAD (no context): "Fixed file writing bug"

Structured Fields

Use type to classify your checkpoint for better searchability:

  • type: "decision" → include decision + alternatives
  • type: "incident" → include context + evidence
  • type: "learning" → include impact

All types benefit from symbols, next, and impact.

Tags — Think About Future Search

Tags power fuzzy search recall. Write them for discoverability — how would future-you search for this?

Category tags (1-2):

  • Type: feature, bug-fix, refactor, docs, test
  • Area: auth, api, ui, database, build
  • Status: wip, blocked, discovery, decision

Concept tags (2-5) — the important part:

  • Include synonyms and related terms for the core topic
  • Include the problem domain, not just the solution
  • Think: "what words might I use when searching for this later?"

Example: a checkpoint about implementing retry logic for payment webhooks:

  • BAD tags: ["bug-fix", "payments"]
  • GOOD tags: ["bug-fix", "payments", "retry", "resilience", "webhooks", "fault-tolerance", "idempotency"]

Example: a decision to use WebSockets over SSE for real-time updates:

  • BAD tags: ["decision", "api"]
  • GOOD tags: ["decision", "real-time", "websockets", "sse", "push", "streaming", "notifications"]

What Gets Captured Automatically

You don't need to include these — Goldfish captures them:

  • Git branch — current branch name
  • Git commit — current HEAD short hash
  • Changed files — files modified since last commit (.memories/ excluded)
  • Timestamp — UTC, always

Focus your description on the MEANING, not the mechanics.

Source Transparency

This detail page is rendered from real SKILL.md content. Trust labels are metadata-based hints, not a safety guarantee.

Related Skills

Related by shared tags or category signals.

General

recall

No summary provided by upstream source.

Repository SourceNeeds Review
General

plan

No summary provided by upstream source.

Repository SourceNeeds Review
General

plan-status

No summary provided by upstream source.

Repository SourceNeeds Review
General

standup

No summary provided by upstream source.

Repository SourceNeeds Review