agentskills.codes
CO

consolidate-knowledge-base

by gsemetSource

Interactively review the OKF knowledge base to detect contradictions between Findings, mark contradicted/superseded Findings, propose Experiments to resolve open questions, and propose promotions of converged Findings into stable Concept/Structure/Principle documents. Every mutation is confirmed by

Install

mkdir -p .claude/skills/consolidate-knowledge-base && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/15042" && unzip -o skill.zip -d .claude/skills/consolidate-knowledge-base && rm skill.zip

Installs to .claude/skills/consolidate-knowledge-base

Activation

This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.

Interactively review the OKF knowledge base to detect contradictions between Findings, mark contradicted/superseded Findings, propose Experiments to resolve open questions, and propose promotions of converged Findings into stable Concept/Structure/Principle documents. Every mutation is confirmed by a human before it is written. Use periodically or after a batch of recorded Findings.
385 charsno explicit “when” triggerlonger than Claude Code's old 250-char listing cap (fine on current versions)

About this skill

consolidate-knowledge-base

Turn a pile of dated Findings into consolidated, trustworthy knowledge — without ever rewriting history. This skill is interactive: it proposes, the human confirms, then it writes.

When to use

  • After a batch of Findings has been dumped by record-finding.
  • Periodically, to keep the empirical layer coherent with the stable layer.
  • When you suspect two findings disagree.

Guardrails

  • NEVER edit or delete a Finding's body. Only append lifecycle frontmatter (status, contradicted_by, superseded_by).
  • NEVER promote to a Principle without explicit human confirmation — principles are agreed with humans.
  • Ask before every write, using vscode_askQuestions.

Procedure

1. Scan

Read knowledge/findings/, plus principles/, structures/, concepts/. Build a map of claims, tags, and existing contradicts/supersedes links.

2. Detect contradictions

Identify pairs where a newer Finding disagrees with an older Finding, or where a Finding conflicts with a Principle/Structure/Concept. Present each candidate contradiction to the human for confirmation.

3. Mark (on confirmation)

For each confirmed contradiction, append metadata-only fields:

  • On the newer finding: contradicts: [<old-id>] (if not already present).
  • On the older finding: status: contradicted and contradicted_by: [<new-id>]. Use the same pattern with supersedes/superseded_by for replacements. The body of both files stays untouched.

4. Propose Experiments

For unresolved or low-confidence contradictions, propose a new Experiment document (hypothesis, steps, expected_signals, max_runs: 1–2) that would settle the question. Write it to knowledge/experiments/ only on confirmation.

5. Propose promotions

When several Findings converge on a stable truth, propose promoting the insight into the stable layer:

  • "how it works" → Structure in structures/
  • "the idea" → Concept in concepts/
  • "the standard/convention" → Principle in principles/ (always human-agreed) Link the promoted doc back to its supporting findings (supported_by / related_concepts). Write only on confirmation. Findings remain untouched.

6. Validate

just knowledge-lint
just knowledge-validate

Both must pass before finishing.

Output

A short chat summary: contradictions found and marked, experiments proposed, promotions made, and any open questions left for a future run.

Search skills

Search the agent skills registry