QMD Learning Loop

Capture corrections, errors, feature requests, and recurring best practices in a QMD-native way for markdown-first agent workspaces. Use when a user corrects...

Security Scan

The skill is an instruction-only set of guidelines for capturing and promoting learnings inside a markdown-first workspace; its files and instructions align with that purpose and request no credentials, installs, or external network activity.

Purpose & Capability

Name/description, SKILL.md, and the five reference markdown files all describe and implement the same goal: classify events and route/promote them into workspace docs. There are no unrelated requirements (no env vars, binaries, or external endpoints) that would contradict the stated purpose.

Instruction Scope

Runtime instructions tell the agent to read, classify, and update workspace markdown files (logs, runbooks, decision memos, etc.). That scope is appropriate for a learning-loop skill. The SKILL.md explicitly prefers updating existing docs and keeping chronology separate from durable rules. There is nothing instructing the agent to read unrelated system files, fetch secrets, or transmit data externally.

Install Mechanism

This is an instruction-only skill with no install spec or code files to execute; nothing will be downloaded or written by an installer step. Risk from installation is minimal.

Credentials

No environment variables, credentials, or config paths are requested. The skill operates solely on workspace markdown documents, which is proportionate to its purpose.

Persistence & Privilege

The skill does not request always:true and uses platform defaults (agent-invocable and allowed to invoke autonomously). That is normal for a skill that will edit workspace files. It does not attempt to modify other skills' configs or system-wide settings.

Assessment

This skill is a set of editorial guidelines and templates for capturing and promoting learnings in your markdown workspace. It does not request credentials or install software. The only operational implication is that an agent using this skill will be reading and writing workspace markdown files — so: (1) ensure the agent's file-write permissions are limited to the intended workspace, (2) keep your workspace backed by version control so edits can be reviewed/reverted, (3) review any promoted/created documents for accidental inclusion of secrets or sensitive data, and (4) if you need stricter control, disable autonomous invocation or require manual approval before the agent edits durable docs.

Capture useful learnings without creating a parallel memory system. Route raw chronology to lightweight logs, durable operational lessons to indexed docs, and stable cross-cutting rules to existing workspace guidance.

Core workflow

1. Classify the event:
   • correction
   • error/failure
   • feature request
   • better practice / recurring pattern
2. Choose the lowest-appropriate destination.
3. Prefer updating an existing durable doc over creating a new one.
4. Use QMD-aware discipline: avoid duplicate policy documents.
5. Promote only when the lesson is recurring, cross-cutting, or policy-changing.

First-pass routing rules

Correction or misunderstanding

• Add chronological context to a daily log or other lightweight memory file when the correction matters.
• Promote only if the correction changes a stable rule, behavior, or operating convention.
• If promotion is warranted, update the most relevant durable target:
  • workflow guidance file
  • tooling notes file
  • behavior/principles file
  • decision or policy docs

Error, tool failure, or command failure

• Log immediate chronology in a daily log if the failure matters to continuity.
• If the failure is operationally meaningful or recurring, add it to an indexed incident or error log.
• If the fix becomes reusable, update or create a runbook.

Feature request

• Add a durable request to an indexed backlog or strategy file.
• If the request becomes accepted scope or changes policy, create or update a decision or project doc.

Better practice or recurring pattern

• For workflow/tooling guidance, update a runbook or workspace guidance file.
• For durable system rules, create or update a decision memo.
• For stable behavior or communication principles, update the relevant behavior/principles file cautiously.

Second-pass promotion rules

Use a second pass when the learning seems durable enough to influence future behavior, policy, or structure.

Second-pass sequence

1. Check whether the event still matters after the immediate task.
2. Inspect existing durable docs and prefer updating them over creating a new file.
3. Choose the promotion target using references/promotion-targets.md.
4. Write the promoted rule concisely and prevention-first.
5. Leave detailed chronology in lightweight logs instead of copying it wholesale into durable docs.

QMD-aware duplication check

Before creating a new durable rule, search the relevant QMD collections or inspect the existing indexed docs for the same topic. Do not create near-duplicate policy or workflow documents.

Trigger conditions for second-pass promotion

Promote when one or more are true:

• the issue recurs
• it affects multiple tasks or agents
• it changes operating policy
• it prevents meaningful repeated waste
• it defines a stable convention worth retrieving later

Do not turn every one-off event into a permanent rule.

Review loop

At natural breakpoints, review captured items and decide whether they should stay chronological, become operational logs, or be promoted into durable guidance. Read references/review-loop.md when doing this review.

Existing-doc first rule

Read references/routing-and-promotion.md when routing is ambiguous. Read references/templates.md when you need starter formats for incident entries, feature requests, lightweight memory capture, or a new decision memo.