code-yeongyu/oh-my-openagent / packages/omo-senpi/skills/ulw-plan
ulw-plan
VerifiedExplore-first planning consultant that writes one decision-complete work plan before coding. Use only on an explicit user request for the ulw-plan workflow or a plan before implementation; it never self-activates on a bare ulw run.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 19.6 KB |
| references/full-workflow.md | markdown | 34.1 KB |
| references/intent-clear.md | markdown | 5.2 KB |
| references/intent-unclear.md | markdown | 7.2 KB |
| references/stance-calibration.md | markdown | 9.8 KB |
| scripts/plan-templates.mjs | script | 7.3 KB |
| scripts/scaffold-plan.mjs | script | 8.3 KB |
references/stance-calibration.md
View on GitHub---
name: ulw-plan
description: ulw-plan stance calibration - renderer selection, fork-reply classification, override gates, and the planning-style episode memory.
metadata:
short-description: How to deliver questions to THIS user, learned from behavior
---
# ulw-plan - stance calibration
Read this alongside your intent reference, before the first user-facing question. Intent routing decides WHAT gets asked (which forks survive); stance decides HOW those forks reach the user. **FIREWALL: stance chooses question form, pacing, and batching. It NEVER reinterprets what a reply meant, and no stored profile may bias fork-reply classification - every reply is classified fresh.**
## Renderers
Three delivery modes over the SAME surviving forks (the two filters and owner-decision rules are unchanged upstream). Every question this skill puts to the user goes through the question tool (`ask_user_question` or `request_user_input`, whichever this session lists) with its wait flag set to true, so it raises a prompt and a notification a user away from the screen still sees. When no question tool is listed or a call returns unavailable, put the questions in chat text and say in one line that the prompt is unavailable. Set `required: true` on the call that asks the approval gate (header `Approval`, `승인` in Korean) and on the call that asks the final authorization block (header `Authorize`, `권한` in Korean). A timed-out, dismissed, or unavailable answer to either is not an answer, whatever the tool result says: keep `status: awaiting-approval`, say in one line that the decision is still pending, and end the turn. Leave ordinary forks unflagged: a timed-out or dismissed fork falls to its recommended default.
- **batch** - all surviving forks in as few calls as the tool's limits allow, recommended default first; a skipped fork resolves to its default. Fits users who enumerate upfront and delegate the rest.
- **one-by-one** - one fork per call. Fits users who own decisions individually.
- **examples-first** - no open questions: present 2-3 contrasting concrete approaches and ask which is closest and what is wrong. Fits users who cannot yet externalize what they want - critique is cheaper than generation. Never ask such a user to produce criteria from a blank page.
The user switches renderer at any time by saying so in their own words; honor it immediately.
## Opening stance
Derive the opening renderer BEFORE the first question:
1. **Episodes exist** (planning-style block in projected memory): generalize from the episodes MOST SIMILAR to this session's context (repo, domain, familiarity) - never from a global average:
- consistent episodes -> open in the indicated renderer;
- episodes split by context (own project vs unfamiliar domain) -> apply the side matching this session;
- variable with no context pattern -> neutral opening (topology echo + escape hatch only) and let in-session signals decide. A single session NEVER flips the opening - do not chase the latest session's style.
2. **No episodes (cold start)**: CLEAR intent -> one-by-one; UNCLEAR intent -> examples-first. Never open an unknown user with silently adopted full defaults: that failure is expensive and invisible until the wrong thing is built, while an unnecessary question fails cheap and loud.
3. **Seeds** (`## Seed` entries): consult only while ZERO native episodes exist, and only for the opening renderer.
Announce the stance in one line beside the intent verdict, with the veto: "Recent sessions suggest batch mode - say the word to go one-by-one." The first time you ask anything, state that opting out is a legitimate answer: "you decide" resolves the fork to its recommended default and lands in the defaults ledger.
## Fork-reply classification
Classify every reply to an asked fork into exactly ONE state. Never infer difficulty from hedge words, filler, tone, or a missing option label - evaluate whether the fork became USABLE. Two tests:
- **Resolution**: given this reply plus existing constraints, could two materially different implementations both comply? No -> resolved.
- **Information gain**: did the reply eliminate an option, add a relevant constraint, challenge the framing, or ask a consequence needed to decide? Yes -> progress.
| State | Meaning | Action |
|---|---|---|
| RESOLVED | one implementable outcome: exact label, semantic equivalent, a new outcome rejecting the offered framing, a constraint leaving one option, or explicit delegation | record the semantic decision; continue |
| RESOLVED_BUT_UNINFORMED | names an outcome but marks it arbitrary, admits not understanding the options, or reveals a misconception about its consequence | correct the misconception in one line; defaultable fork -> adopt the recommendation with a visible veto; owner-decision -> one informed yes/no |
| UNRESOLVED_PROGRESS | still open, but the reply added decision-relevant information | use it, answer any question it raised, re-render THIS fork only |
| UNRESOLVED_BLOCKED | still open, nothing decision-relevant added | re-render THIS fork as 2-3 concrete outcomes, recommendation first, one material consequence each |
One reply may resolve several forks, correct facts, and add scope all at once - parse all of it: distribute answers to their forks, fold new scope into the draft, treat a factual challenge as UNRESOLVED_PROGRESS on the challenged claim.
After two UNRESOLVED_BLOCKED on distinct forks, offer the renderer menu once (continue as-is / defaults for the rest / examples-first). A session-wide renderer switch happens ONLY on an explicit user choice or request - never from counters, streaks, or hedge frequency.
## Override gates
"질문 그만", "니가 정해", "알아서", "stop asking", "you decide" and kin become instructions only after three gates, in order:
1. **Speech act** - a direct affirmative instruction from the user, now. Not negation ("알아서 하지 마"), quoted text, hypotheticals, or reported speech.
2. **Intent** - classify: delegate THIS fork / delegate the REMAINDER / stop the interview / request a recommendation / fix the process (fewer, better questions). Only the two delegations transfer decision ownership.
3. **Scope** - take the narrowest supported reading. A bare "알아서" answering one fork delegates that fork only; session-wide requires explicit breadth ("나머지는", "전부", "from now on"). Preserve stated exceptions exactly.
Actions: delegate-this-fork -> recommended default + ledger, renderer unchanged. Delegate-remainder -> defaults-ledger mode for the remaining defaultable forks; acknowledge the exact scope. Stop-interview -> stop incremental questions, adopt defaults, and consolidate ALL surviving owner-decisions into ONE final authorization block (recommended choice + material consequence each); "go" authorizes only when it follows that block. Recommendation-request -> give it; adopt only if the wording also delegated. Process-fix -> improve the questions; ownership unchanged.
**INVARIANT: no override phrase silently authorizes an irreversible, destructive, or spend decision.** When questions were stopped, those decisions arrive as the final authorization block - never as adopted defaults.
## Planning-style episode memory
The profile is an append-only log of observation episodes at `system/human/planning-style.md` (memory tool; create lazily on the first qualified episode, description: "Planning-interaction episode log; generalize at read time"). **NEVER store scores, counters, thresholds, settled flags, or a summarized persona** - generalization happens at read time, every session, from the raw episodes. A cached summary is what goes stale; raw episodes cannot.
**Qualified episodes only.** At close, record 0-2 lines - only when one of:
- the user explicitly chose or requested a renderer, or stated an interaction preference (record as `declared`);
- the same unsolicited stance showed across 2+ distinct forks: batched or superseded them, demanded per-fork ownership, or needed concrete examples to answer (record as `observed`).
Never record: sessions with no genuine fork, topology corrections, bare approvals ("go"), a single local delegation, frustration alone, or an answer shape the renderer forced. No qualified signal -> record NOTHING; forced observations bias the log toward tidy sessions.
Episode format - every field required:
```
- [YYYY-MM-DD] <context: repo/domain + familiarity evidence> - <observed behavior, one line> <!-- src: <session-id>; model: <session model>; pattern: observed|declared; confidence: low|medium|high -->
```
- **Minority protection**: contradicting episodes are context markers, not noise. Before applying the majority pattern, check whether THIS session matches the minority episode's context. Never prune minority or `declared` entries; cap the log near 15 by pruning the oldest majority-consistent entries first.
- **Memory tools absent**: note the session stance in the draft only; skip the profile.
### Seeds (onboarding-inferred)
`## Seed` entries in the same block are onboarding inferences from OTHER coding agents' session history. Hard rules:
- **`confidence: low`, ALWAYS.** Never raise it, however consistent the foreign evidence looks: behavior in another tool reflects that tool's affordances (no memory there, a different delegation surface), not necessarily this user.
- Every seed carries full provenance: source harness name(s), the model(s) under which the pattern was mainly identified, the session ids mined, and the date the seed was written (`seeded:`). Phrase the body as a hypothesis ("appears to ..."), never a conclusion.
```
- [YYYY-MM-DD] (seed) appears to <hypothesis + context> <!-- src: <harness>/<session-id>,...; harness: <names>; model: <models>; seeded: YYYY-MM-DD; pattern: inferred-from-external-sessions; confidence: low -->
```
- **Precedence: native episodes > the user's explicit statements > seeds.** Once 2-3 native episodes exist, stop consulting seeds entirely.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| planning-and-task-breakdownaddyosmani/agent-skillsBreaks work into ordered tasks. Use when you have a spec or clear requirements and need to break work into implementable tasks. Use when a task feels too large to start, when you need to estimate scope, or when parallel work is possible. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| screenpipethedotmack/claude-memSearch screen and audio history recorded by Screenpipe and optionally save selected context into claude-mem. Use when the user asks to recall desktop activity, a meeting, or something seen in another app through Screenpipe. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |
| contact-widgetnexu-io/open-designSelf-contained floating chat widget with welcome screen, social links, meeting button, and message input. Single HTML file, zero dependencies. | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |
| meeting-notesnexu-io/open-designMeeting notes page — title bar with attendees, agenda checklist, decisions block, action items table with owners + dates, and a "next meeting" footer. Use when the brief mentions "meeting notes", "minutes", "1:1 notes", "all-hands recap", or "会议纪要". | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |