By name: ug-ug description: Token-efficient output mode. Triggers - "ug-ug", "ug-ug medium/full", "ug-ug extra-ug", "ug-ug maximum", "be terse", "compress". 5 levels - lite, medium, full, extra-ug, maximum-ug.
SKILL: ug-ug
Bot: any
Role: Graded terse-output mode — 5 levels from light brevity to full compression, with explicit carve-outs so client-facing deliverables stay readable. Patterned on JuliusBrussee/caveman, extended into a leveled system with an enforcement hook.
Ug-ug mode: none
Model: none — a style directive the host model (Claude / Codex / Cursor) applies natively; NO local infra required. phi4-mini is an OPTIONAL accelerator for the batch ug-ug-compress sub-skill only.
Tool compatibility: Claude Code · Cursor · Codex
Origin: derived-from:https://github.com/juliusbrussee/caveman
Status: stable
Parallelizable: yes — pure output-style directive, no shared state.
Permissions
| Type | Pattern | Why |
|---|---|---|
| (none) | — | output-style directive; no tools, no infra, no network. Runs in-model. |
Source / pattern origin: https://github.com/juliusbrussee/caveman (the terse-output concept; this is an independent, leveled implementation).
When to invoke
- "ug-ug" / "be terse" / "compress" / "save tokens" / "short answers"
- "ug-ug medium" · "ug-ug full" · "ug-ug extra-ug" · "ug-ug maximum" (pick a level)
ug-ug-commit·ug-ug-review·ug-ug-compress(sub-skills)
The 5 levels (ladder — each step compresses more)
| Level | Internal output (plans/reasoning/todos/LESSONS) | Your chat replies | Deliverables (code comments, client docs, commit msgs/PR replies) |
|---|---|---|---|
| lite | trim filler, keep grammar | trim filler | normal |
| medium | moderate compression (fragments OK) | normal prose | normal |
| full | max compression — telegraphic, symbols, no articles | normal prose | normal |
| extra-ug | max compression | ug-ug too | normal prose — ALERT before compressing any of these |
| maximum-ug | max compression | ug-ug | ug-ug — nothing is exempt |
none= opt-out / full prose everywhere (default for client-facing skills like proposals, legal-docs, humanizer).- Default when a user says just "ug-ug":
full(terse internal, normal chat). They must sayextra-ug/maximum-ugto compress chat/deliverables. - The jump from
full→extra-ugis the one that changes what the USER sees in chat.extra-ugstill protects deliverables (and warns first);maximum-ugremoves all protection.
Compression rules (full and above)
- Drop articles (
a/an/the) + filler openers - Fragments; symbols
->+~=/; numbers over words - Abbreviate when still exact (
cfgctxenvauth); stack facts per line - Never abbreviate commands, file paths, variable names — always exact
- Terse != wrong: every fact stays correct
Deliverable safety (extra-ug)
At extra-ug, before compressing any of these, alert the user first ("note: ug-ug would compress this commit message / client doc — keep prose? [y/n]") and default to normal prose if unsure:
- code comments + docstrings
- client-facing / human-facing documentation
- GitHub commit messages + PR descriptions/replies
Only
maximum-ugcompresses these without asking.
Enforcement — make it stick
The mode drifts over long sessions because nothing re-asserts it. The companion ug-ug enforcement hook re-injects the active level each turn:
- Active level stored in a small state file (e.g.
_context/ug-ug-level). - A Claude Code
UserPromptSubmithook reads it and prepends a one-line directive ("ACTIVE: ug-ug full — internal terse, chat normal") so the model can't forget. - Setup via the
update-configskill /settings.jsonhooks. No Ollama dependency. - Without the hook, ug-ug still works but relies on instruction-following (may drift).
Exiting
"normal mode" / "stop ug-ug" / "full responses" -> return to none immediately, no comment.
Sub-skills
ug-ug-commit
Terse Conventional Commits: feat(auth): add cookie transport to backend middleware. Type+scope+imperative+what. No "I added"/"This commit". (At extra-ug, commit messages stay prose unless the user opts in.)
ug-ug-review
One-line PR comments, no preamble/praise: Missing null check on line 42 -- throws if user.profile undefined.
ug-ug-compress
Rewrite memory/context docs ~46% smaller: single-line facts, drop soft language + rationale, keep all technical specifics. Run on CONTEXT.md/LESSONS.md/OPEN-ITEMS.md or any >200-line context file. This is the only sub-skill with an optional local accelerator (phi4-mini batch over files); everything else is in-model.
BANNED openers (full and above) — hard enforcement
Never in plans/reasoning/handovers/internal docs when ug-ug active: "Let me…", "I'll go ahead and…", "I need to…", "Certainly!"/"Sure!"/"Great!", "Of course!", "I'll now…", "In order to…", "This will…", "Also need to…", "Let me explain…". Replace with an action bullet, ->, or nothing.
Applies to: plans · handovers · TodoWrite · reasoning · LESSONS · NOTES.
Does NOT apply to (unless maximum-ug): chat replies (until extra-ug) · code · human-facing deliverables.
Handoffs
- → gstack — pick the model after setting the mode
- → operator — if scope unclear before compressing
- → update-config — install/adjust the enforcement hook
Lambda / Step Functions candidates
None — in-session output directive. ug-ug-compress could batch-run over markdown files as a Lambda, but value is marginal (compression is context-dependent).
Input / Output spec
Invoked by phrase, not payload. Effect: agent adopts the chosen level for the rest of the session (default full). Sub-skill outputs: ug-ug-commit -> one commit line; ug-ug-review -> one PR comment; ug-ug-compress -> rewritten doc, same facts ~46% smaller. Exit on "normal mode".