Skip to content
Back to skills

Wikimem Observer

ASecurity

Run WikiMem's self-improvement Observer — score every wiki page on quality, surface the weakest/orphan/most-referenced pages and knowledge gaps, then optionally auto-improve a budget-capped set and commit one reviewable change. Use WHEN the user says "improve / clean up / lint / tidy my wiki", "what's weak in my notes", "score my vault", "fill the gaps", or wants the nightly self-evolution loop (Automation 3). This is the open-endedness moat — the wiki gets better while you sleep. Do NOT use ...

  • 7 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 20, 2026
ai-agentsgobashgit

Works with

  • cli

Security analysis

A100/100

Scanned October 20, 2026

npx -y skills add naman10parikh/wikimem --skill wikimem-observer --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Wikimem Observer?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Wikimem Observer
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/naman10parikh-wikimem-observer/badge)](https://www.skillsdirectory.com/skills/naman10parikh-wikimem-observer)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: wikimem-observer
description: Run WikiMem's self-improvement Observer — score every wiki page on quality, surface the weakest/orphan/most-referenced pages and knowledge gaps, then optionally auto-improve a budget-capped set and commit one reviewable change. Use WHEN the user says "improve / clean up / lint / tidy my wiki", "what's weak in my notes", "score my vault", "fill the gaps", or wants the nightly self-evolution loop (Automation 3). This is the open-endedness moat — the wiki gets better while you sleep. Do NOT use for ingesting new sources (use wikimem-ingest-pipeline) or querying knowledge.
---

# WikiMem Observer — The Wiki That Improves Itself

## What this skill does

The Observer is WikiMem's self-improvement loop (Automation 3 — the moat). It reads the whole vault, **scores each page**, finds the pages worth touching (weakest + newest + most-referenced), surfaces **orphans, contradictions, and knowledge gaps**, and — when allowed — **auto-improves a budget-capped set of pages and writes exactly ONE git commit** so the user can accept or reject the run via time-lapse restore. Budget-capped by design (default $2.00; a typical run is ~11 LLM calls ≈ $0.005–$0.01 on a tight cap).

Scoring dimensions: **recency/freshness, citation quality, wikilink density, category coverage, TLDR/summary quality.**

## Trigger — invoke this skill when

- "Improve / clean up / tidy / polish my wiki" · "make my notes better"
- "What's weak / stale / orphaned in my vault?" · "score my wiki" · "find gaps"
- "Find contradictions" · "which pages need work?"
- "Set up the nightly self-improvement" / "run the Observer"

Do **not** invoke to add new content (that is `wikimem-ingest-pipeline`) or to answer a question from saved knowledge.

## Steps

1. **Resolve the vault:**
   ```bash
   VAULT=$(find . -name "AGENTS.md" -maxdepth 3 | head -1 | xargs dirname 2>/dev/null || echo ".")
   ```

2. **Always observe before improving (read-only, $0):** get the report first so the user sees the diagnosis before any spend or change.
   ```bash
   wikimem observe --vault "$VAULT" --max-pages 100        # scores + orphans + gaps, no writes
   wikimem observe --vault "$VAULT" --json                  # same, machine-readable
   ```

3. **Present the diagnosis** — avg score, weak pages (< 60%), orphans, knowledge gaps, top recurring issues. THEN ask before spending: "Auto-improve the N weakest pages? (uses LLM, budget-capped)".

4. **Improve, budget-first, on approval.** Two equivalent entry points:
   ```bash
   # Observer with auto-improve, hard budget + page caps (preferred — explicit caps)
   wikimem observe --vault "$VAULT" --improve --max-improvements 5 --budget 0.50

   # Council-style improve against a quality threshold
   wikimem improve --vault "$VAULT" --threshold 80          # dry preview first:
   wikimem improve --vault "$VAULT" --threshold 80 --dry-run
   ```
   Each improve run writes **one** commit — keep it that way so it stays reviewable.

5. **Show the delta and offer to accept/reject:**
   ```bash
   wikimem status --vault "$VAULT"     # new avg score, orphans resolved
   wikimem history --vault "$VAULT"    # the Observer's commit; revert if unwanted
   ```

## Scheduling the nightly loop

```bash
# 3am daily — observe + improve up to 5 pages under a tight budget
0 3 * * * cd /path/to/wiki && wikimem observe --improve --max-improvements 5 --budget 0.50
```
Use `--budget` and `--max-improvements` on the cron path; an uncapped nightly run can spend.

## Expected output

```
Observer Report — 2026-06-06
Pages reviewed: 47/47   Avg score: 6.2/10
Weak (<60%): 8    Orphans: 3    Gaps: 5    Contradictions: 1
Top issues: missing TLDR (12), no tags (8), low wikilink density (6)

[after --improve] Improved 3 pages, $0.41 spent (cap $0.50), 1 commit written.
Accept this run, or `wikimem history` → revert?
```
Always report: avg score (before→after), what was changed, dollars spent vs cap, and that one commit was written (so it's reversible).

## Anti-patterns

- Auto-improving without showing the read-only `observe` diagnosis first.
- Running the nightly cron without `--budget` / `--max-improvements` caps.
- Writing more than one commit per run (breaks clean accept/reject).
- Treating Observer output as final truth — it is a reviewable suggestion, revertible via `wikimem history`.

## Related

| Resource | Path |
|----------|------|
| Observer engine | `src/core/observer.ts` · `src/core/observer-patterns/` |
| Improve (Council) | `src/cli/commands/improve.ts` → `src/core/improve.ts` |
| Observe command | `src/cli/commands/observe.ts` |
| History / revert | `src/cli/commands/history.ts` → `src/core/history.ts` |
| Slash command | `templates/claude-commands/wikimem-improve.md` |

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.