/pitwall:secretary — things repeated often → skill

Principle: anything that has to be said more than 2 times belongs in a file, not in someone’s memory · the secretary reports and proposes, the user chooses

Arguments: none = since the previous report (--since auto) · all = everything · <YYYY-MM-DD> = since that date · called from /pitwall:stop = short mode (step 6)

1. Pull the messages

Run from the repo’s main checkout (the transcript path is computed from cwd — a worktree gets a different folder):

node <pitwall dir>/scripts/secretary-scan.mjs --since auto --assistant > <scratchpad>/pitwall:secretary.txt

• 1 line = time session U|A message · U = what the user typed themselves (including /command args, [N image]) · A = what the agent told the user (cut at 240 characters — things deep inside a long message are lost)
• tokens / keys are already replaced with [secret] — still never quote anything that looks secret in the report
• transcripts exist only on this machine (Windows / macOS separate) → the report header states the machine name (hostname)
• no new messages → “nothing new since ”, done

2. Group by intent (not by matching words)

Read the whole file, group it into topics — mixed Thai / English and misspellings (“fash”, “bestpractics”) count as the same topic if they ask for the same thing

Type	Looks like	Example
Repeated procedure	The same kind of work starts with similar words	“here’s the QA report” → fix → “how do I test it” → /commit-push
Repeated correction / ban	The user calls out the same kind of mistake	“not pro”, “take the emoji out”, “sloppy”, “why didn’t you follow the loop”
Repeated question	The user has to ask because the agent did not say it on its own	“what next”, “do I have to wait”, “did it pass”, “show me the evidence”
Agent explains again (A)	The agent tells the same story several rounds	merge order, how to switch accounts

Count: times (messages) and sessions (different 2nd column) · a “[Request interrupted]” message followed immediately by the same text typed again = 1 time

Threshold	Result
≥ 3 times in ≥ 2 sessions · or the agent explains again ≥ 3 times	propose
2 times	watch (one line at the end of the report)
Once · generic commands (“let’s go”, “go”, “start”) · a skill already used normally (e.g. /commit-push called often = working well, not a problem)	do not report

3. Does it already have a home

Per topic that passes the threshold, grep keywords (Thai + English) in .claude/skills/*/SKILL.md, the rules files of the profile repo rules, CLAUDE.md, .claude/agent-loop.md, ~/.claude/skills/ (machine-level skills), this plugin’s skills and the project’s MEMORY.md

• Not there → new
• There, but the message repeats after the date that file added the topic (git log -1 --format=%cs -S'<word>' -- <file>) → exists but still needs reminding — the skill is not invoked (the description lacks the words the user says) or the agent does not follow it (the rule is buried / vague) · say which one it likely is
• There, and not repeated after that → do not report (the fix worked)

4. What it should become

Trait	Home
Several steps, ordered, invokable by a command	a new skill or a new section in an existing skill
A one-line ban / requirement, every task	the repo’s rules files (profile repo rules)
Repo-specific values (branch, check, port)	.claude/agent-loop.md
The user’s personal preference / used across repos	memory (feedback)
Agent explains again	move the explanation into a skill and have reports refer to the skill name (like /pitwall:merge-order)
Domain knowledge (WHAT / WHY)	the repo’s docs / knowledge base

5. Report (short, in the profile’s user language)

Secretary · <machine> · <N> messages · <M> sessions · since <date>

| Topic | Times / sessions | Example (≤ 12 words) | Now | Proposal |
|---|---|---|---|---|
| Get QA report → fix → how to test | 5 / 4 | "here's the QA report" | new | skill `/qa-fix` |
| Asks what to do next / whether to wait | 9 / 3 | "so I have to wait, right?" | new | every report ends with "You: …" (lap / rules) |
| Merge order | agent 11 | "merge in order, #101 first" | exists `/pitwall:merge-order` | — |

Watch: <topic (2)> · …

• Sorted by times, most first · at most 6 rows (the rest go to watch)
• The example quotes the user’s own words, short, 1 message — never quote tokens, passwords, unnecessary addresses, personal data

Then ask with AskUserQuestion (multiSelect) which ones to do — chosen items: /pitwall:ticket creates an [improve] task, then follow the normal round (branch → fix → /pitwall:gauge → PR) · items not chosen: note them as declined

Finally node <pitwall dir>/scripts/secretary-scan.mjs --since auto --mark > /dev/null (records that reporting has reached this point · state lives in secretary.json next to that machine’s transcripts)

6. Short mode (called from /pitwall:stop)

Do steps 1–4 but report only propose topics not reported before, one line per topic + “run /pitwall:secretary to choose” · none → say nothing · no --mark (leave marking to the full /pitwall:secretary)

Never

• Never create / edit a skill, rule, memory before the user chooses
• Never send transcript content off the machine (tracker / PR / artifact) — in tasks / PRs put only the topic name + the count
• Never draw conclusions about the user’s habits or abilities — report only what repeats