arjunrajlaboratory/memex/observe-manual-patterns

Observe manual patterns

You are running as `agent:auditor` for this skill. Your job: detect repeated user-driven (actor me) mutations in log.md that don't yet have a skill wrapping them, surface them as skill-creation candidates.

Why this skill exists

The vault's hand-written skills were created when the user noticed a pattern: "I keep doing X manually, this should be a skill." This observer makes that detection systematic — instead of relying on the user noticing, it scans the canonical mutation log and flags repeated shapes.

The cost of NOT having this observer: useful skills get created late, after the user has done the same manual work 8–10 times. The value is short-circuiting the "I noticed too late" pattern.

Inputs

• Window (optional, default 30 days). The scan window from log.md.
• Threshold (optional, default 3). Minimum count for a pattern to be surfaced.

Step 0 — Orient

Read these to know the universe of existing skills (so you don't recommend creating a skill that exists):

• ls .claude/skills/ — list of installed skills.
• ls Agents/Prompts/ — paste-able prompts (each prompt is a skill candidate already documented).

Note: the canonical log line shape from AGENTS.md is:

<datetime> — <actor> — <verb> — <[[target]]> — <one-line summary>

Where actor is one of me, agent:planner, agent:auditor, agent:librarian, agent:capture, etc. (the vocabulary in AGENTS.md / log-mutation; user-driven lines use the bare token me).

Step 1 — Filter the log

Read the last <window> days of log.md. Filter to lines whose actor field is me (grep for — me — ; the em-dash delimiters keep it from matching prose). Drop lines whose verb matches a known skill-equivalent operation (since actor me here means the user themselves did it, NOT through a skill). The relevant verbs to KEEP (i.e., what the user did manually) include but are not limited to:

• update, schedule, commit, publish, send, move, rename, link, tag, delete, restore, convert, merge, archive

The verbs to DROP (these correspond to existing skills, even when the user invoked them — the action shape itself is already a skill):

• create of a typed note (covered by create-task, capture-decision, ingest-source, ingest-person, ingest-project)
• triage, brief, review, lint, promote, flesh-out, close, revisit, observe, digest (each is a named skill)

Step 2 — Bucket by shape

For each kept line, extract a normalized shape:

<verb> <target-type>

Where target-type is inferred from the wikilink target (e.g., [[X]] → person, [[Y]] → task, [[Z]] → decision, [[W]] → project). If the target isn't a vault entity (e.g., "calendar event for ...", "GitHub repo X"), use the target's category as the type (calendar-event, external-repo, email-draft).

Group by shape. Count occurrences. Drop shapes with fewer than <threshold> occurrences.

Step 3 — Format the report

## Manual patterns — <window>
 
| Shape | Count | Example log lines |
| --- | --- | --- |
| `update person.last_contact` | 5 | 2026-05-12 — me — update — [[X]] — last_contact bump; 2026-05-17 — ... |
| `schedule calendar-event` | 4 | 2026-05-09 — me — schedule — calendar event for ...; ... |
| ... | ... | ... |
 
### Proposed skill candidates
-**<verb-target>**: candidate skill `<suggested-name>` — <one-line shape of what the skill would do>.
  -Frequency: <count> in <window> days. Examples: <up to 3 target wikilinks>.

If there are zero patterns above threshold, write: "No repeated manual patterns above threshold (<threshold>) in the last <window> days."

Step 4 — Persist the report

Two paths (same as observe-skill-corrections):

• Open weekly review for this ISO week — append the ## Manual patterns — <window> section.
• No open weekly review — write a standalone Ops/Reviews/Observations/Manual patterns - <ISO-week>.md. Frontmatter: type: review, scope: observation-manual-patterns, date: <today>, generated_by: agent:auditor, sensitivity: private.

Step 5 — Log

<datetime-with-tz> — agent:auditor — observe — [[<output file>]] — manual-patterns; <N patterns above threshold>; top: <shape> @ <count>

What this skill does NOT do

• Does not create any skill. Proposals only. The user runs superpowers:writing-skills (or hand-writes) the new skill if they accept the proposal.
• Does not de-duplicate against `Agents/Prompts/`. If a proposed skill already exists as a paste prompt but not as a .claude/skills/ entry, surface it anyway — that's a different kind of suggestion (promote the prompt to an auto-skill).
• Does not infer intent beyond shape. "Update person.last_contact 5 times" might mean the user wants a bump-contact skill, or might mean they had a busy week with many follow-ups. The report surfaces; the user decides.

Model recommendation

sonnet. Structured grep over a small file + grouping. No judgment required beyond verb classification.

Related

• weekly-review — invokes this skill in its ## Learnings section.
• observe-skill-corrections — sibling observer; scans transcripts instead of log.md.
• superpowers:writing-skills — the skill the user would invoke if they accept a proposed candidate.
• log-mutation — the canonical log-append helper.