Agent-Skills.md

Agent Skills: Remember

Esta skill debe usarse cuando el usuario pide "recuerda que", "guarda

UncategorizedID: DieGopherLT/claude-kit/remember

Repository

DieGopherLT/claude-kit
DieGopherLT
1

Install this agent skill to your local

pnpm dlx add-skill https://github.com/DieGopherLT/dotclaudefiles/tree/HEAD/plugins/claude-management/skills/remember

Skill Files

Browse the full folder contents for remember.

Download Skill

Loading file tree…

plugins/claude-management/skills/remember/SKILL.md

Skill Metadata

Name
remember
Description
Esta skill debe usarse cuando el usuario pide "recuerda que", "guarda

Remember

Given a piece of information, classify its nature and write it to the correct Claude Code memory location for the current project.

Scope

Project-level only. The five valid destinations are:

  • CLAUDE.local.md — local environment, never versioned
  • CLAUDE.md — always-loaded project context, versioned
  • .claude/rules/ — on-demand conventions (path-scoped or always-on), versioned
  • .claude/skills/ — reproducible flows Claude can execute, versioned
  • Memory — project context not derivable from code, machine-local

User-global destinations (~/.claude/CLAUDE.md, ~/.claude/rules/) are out of scope.

Reference files (read on demand)

  • references/storage-decision-guide.md — full spec for each destination: when to use, when not to, write format, and a concrete example. Read when the classification table below doesn't resolve the case.

Decision sequence

Apply these questions in order. Stop at the first yes.

| Question | Destination | |---|---| | Is it a secret, credential, or API key? | Refuse — warn the user; none of these systems are appropriate | | Does it only apply on your local machine (localhost URLs, local paths, personal dev overrides)? | CLAUDE.local.md | | Is it project orientation Claude needs before opening any file (overview, structure, build/run commands)? | CLAUDE.md | | Does it apply only to certain file types or directories in the project? | .claude/rules/ with paths frontmatter | | Is it a cross-cutting coding standard for the whole project (naming, error handling, control flow)? | .claude/rules/ without frontmatter | | Is it a step-by-step procedure or workflow that Claude should execute reproducibly? | .claude/skills/ | | Is it project context not derivable from the code (infra, stakeholders, historical decisions, external constraints)? | Memory (topic file under ~/.claude/projects/<project>/memory/) |

If the information fits two buckets equally, use AskUserQuestion to clarify before writing.

Workflow

Step 1: Classify

Apply the decision sequence above. Use $ARGUMENTS as the information to store if provided; otherwise use the user's last message.

Check for secrets first: URLs with credentials, tokens, passwords, private keys. If found, stop and warn — do not write.

If the information applies to all of the user's projects (not just this one), it belongs in ~/.claude/CLAUDE.md or ~/.claude/rules/. This skill does not write to user-global destinations — tell the user and stop.

Step 2: Resolve ambiguity

Proceed without asking for clear cases. Ask only when:

  • The information could plausibly go to two different destinations and the choice changes how it behaves (e.g., "applies to all TypeScript" vs "applies to all code in the project").
  • The destination file is near its size limit and writing would make it worse (CLAUDE.md approaching 200 lines).

Step 3: Write

For each destination, see references/storage-decision-guide.md for the exact write format. Abbreviated rules:

  • CLAUDE.local.md: verify CLAUDE.local.md is listed in .gitignore before writing; if not, warn the user and stop. Append at end; create file if absent.
  • CLAUDE.md: insert under the most relevant existing section; never create a new top-level section unless no section fits. If the file is at or above 180 lines, suggest running rulify first and stop.
  • .claude/rules/ (path-scoped): check for an existing rule whose paths covers the same scope; merge into it if found, otherwise create a new file under .claude/rules/<category>/<topic>.md with paths frontmatter.
  • .claude/rules/ (always-on): check for an existing rule covering the same standard; merge if found. Create new file without any frontmatter if no match. Never start the file with --- — that converts it to a path-scoped rule with no triggers and it will never load.
  • .claude/skills/: check for an existing skill covering the same flow; merge into it if found. Otherwise create .claude/skills/<skill-name>/SKILL.md. See references/storage-decision-guide.md for format details.
  • Memory: write to a topic file (~/.claude/projects/<project>/memory/<topic>.md) using three-field frontmatter (name, description, metadata.type) plus body. Add or update the pointer line in MEMORY.md.

If the new information contradicts existing content in the target file, overwrite the stale entry rather than appending. Report the replacement in Step 4.

Verification clause for transferable practices. When the information to persist is a practice about an external tool, framework, or language (as opposed to a project-internal fact or convention), launch the practice-verifier agent with the claim before writing — what just worked in this session may be version-specific, incomplete, or accidentally right, and a persisted mistake propagates to every future session. Persist only what the agent deems actionable per its own contract (agents/practice-verifier.md owns the confidence bar), applying its corrections when the verdict is adjusted; report refuted claims to the user with the verifier's evidence. Skip verification for project-internal facts — the codebase, not the internet, is their authority, and the session just observed them directly.

Step 4: Report

One sentence: what was saved, where, and the deciding criterion.

Example: Saved to .claude/rules/code-standards/naming.md (always-on) — cross-cutting naming convention that applies to all project code.