pack
/pack
Maintain the lll skill pack. Shows which skills fire in your transcripts and where none did, then adds or fixes a skill to CONTRACT.md, runs checks and evals, bumps the version and tags. Use when asked to add a skill, fix a skill that did not fire, which skills do I use, check the pack, release the pack. Not for building a product feature (lll:build).
- Author
- lilly-logic-labs ↗
- Repository
- lll-skills ↗
- Licence
- MIT
- Last repository change
- 12.11.2026
Assessment and security
- Files in this package
- 8
- Scripts
- 2
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 12.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
- 2 scripts in this package. Review them before installing, as Claude may execute them.
Scripts: scripts/lint.mjs, scripts/skill-usage.mjs
What it does and when to use it
Maintain the lll skill pack. Shows which skills fire in your transcripts and where none did, then adds or fixes a skill to CONTRACT.md, runs checks and evals, bumps the version and tags. Use when asked to add a skill, fix a skill that did not fire, which skills do I use, check the pack, release the pack. Not for building a product feature (lll:build).
Original author description
Maintain the lll skill pack. Shows which skills fire in your transcripts and where none did, then adds or fixes a skill to CONTRACT.md, runs checks and evals, bumps the version and tags. Use when asked to add a skill, fix a skill that did not fire, which skills do I use, check the pack, release the pack. Not for building a product feature (lll:build).
How it works
- When to use
- When not to use
- Modes
- Tools
- Steps
- Rules
- Output
SKILL.md preview
--- name: pack description: Maintain the lll skill pack. Shows which skills fire in your transcripts and where none did, then adds or fixes a skill to CONTRACT.md, runs checks and evals, bumps the version and tags. Use when asked to add a skill, fix a skill that did not fire, which skills do I use, check the pack, release the pack. Not for building a product feature (lll:build). metadata: origin: lll --- # Pack Keeps the lll skill pack small, sharp and proven: measures which skills fire, finds the tasks no skill covered, then adds or fixes one skill at a time with evals, checks and a release. A skill only earns its listing cost if it triggers when it should and stays quiet when it should not. ## When to use - "Which skills do I actually use", "why did the skill not fire", "add a skill for X", "improve the descriptions", "check the pack", "release the pack". - A task went badly because no skill fired, or the wrong one did. - After a batch of skill edits, before sharing or tagging the pack. - Checking the pack for another host that reads `SKILL.md` (Codex, Cursor). ## When not to use - Building a product feature or app code: `lll:build` (plan it with `lll:plan`). - Token cost of the whole setup (hooks, `CLAUDE.md`, MCP servers, plugins): `lll:tokens`. This skill covers only the pack's own listing. - A skill for another plugin, or for one project only: `anthropic-skills:skill-creator` runs the full draft, test and benchmark loop. Use it for deep description tuning with variance analysis. The contract here binds only this pack. - Install or manifest health: `/doctor` and `claude plugin validate .`. ## Modes - **Guidance** (default): run the read-only tools below, explain the numbers, draft a description or outline inline, name the fix. Write no files, start no subagents. - **Full**: only when asked to add, change or release a skill. Writes inside the pack, adds evals, runs every check, bumps the version when asked. Subagents are optional (`../../shared/agents.md`). Read `../../shared/ground-rules.md` first. Other sessions often edit the pack at the same time: change only the skills you were asked about, and never commit, stash, reset or tag without the owner. Transcripts hold private text: the usage tool prints counts and names; `--prompts` also prints 100 characters of each prompt, so use it only for the owner's own history and keep it out of reports. ## Tools Run from this skill's folder unless marked (root), which means the pack root (the folder with `.claude-plugin/plugin.json`). | Command | Does | |---|---| | `node scripts/skill-usage.mjs [--all] [--since 90] [--prompts]` | Per skill: auto-picked calls, typed `/lll:` commands, sessions, last use. Skills never used. Other suites' skills called that overlap an lll name. Sessions with many prompts and no lll skill, with their tools. | | `node scripts/lint.mjs [skill ...]` | Beyond `check.mjs`: missing "Not for", trigger list cut by a period, router line over 160 characters, references that point at each other, orphan references, scripts without `--selftest`, eval coverage, trigger phrases shared by two skills. | | `node scripts/check.mjs` (root) | Contract and relative-link check. | | `node scripts/gen-router.mjs` (root) | Rewrites `hooks/router.txt`, the skill map printed at session start. | | `node scripts/selftest.mjs` (root) | Self-test of the pack scripts and hooks. | | `claude plugin validate .` (root) | Manifest and skill validation. | | `claude plugin details lll` | Component inventory and "Always-on" token cost. Needs the pack installed. | | `claude plugin eval . --ablation none --tag trigger` (root) | Trigger evals. See `references/evals-howto.md`. | | `claude plugin tag --dry-run` (root) | Shows the `lll--v<version>` tag it would create. | ## Steps 1. **Orient.** Find the pack root. Read `CONTRACT.md`, `evals/README.md`, `NOTICE`, and the `SKILL.md` of every neighbour the change could touch. State the goal in one line: new skill, fix a trigger, retire or merge, or release. 2. **Measure use.** `node scripts/skill-usage.mjs --all --since 90`. Read the header first: how many sessions used any lll skill. A new install has no history; say so and rely on the user's examples instead. 3. **Find the gaps.** Method and signals in `references/gaps.md`: typed more than auto-picked (the description is failing), long sessions with no skill, other suites' skills doing an lll job, skills never used. Write each gap as one line: pattern, how often, existing skill or none, proposed fix. 4. **Choose the smallest fix.** In this order: reword the description of the skill that should have fired; add a reference or mode to it; add a new skill only when the work has its own trigger, recurs, and no built-in or `anthropic-skills:*` skill already does it. Nothing to fix is a valid answer. Never delete a skill without the owner. 5. **Write or change the skill.** Follow `references/authoring.md` and `references/descriptions.md`. Test the real task without the skill first, so the skill answers observed failures and not guesses. Write facts you verified; hedge the rest. Sources: adapt only MIT or Apache-2.0 text and list it for `NOTICE`; everything else is inspiration, in your own words. 6. **Add evals.** Three positive trigger cases phrased as a busy person types, and one near-miss that belongs to a neighbour. How: `references/evals-howto.md`. 7. **Verify, in order.** `lint.mjs`, `check.mjs`, the skill's own `--selftest` scripts, `selftest.mjs`, `gen-router.mjs` (read the token count it prints), `claude plugin validate .`, `claude plugin details lll`, then the trigger evals. Fix every error, including ones in other skills that block the run, or name the exact blocker. A failing positive means the description is wrong: change the description, not the grader. 8. **Release, only when asked.** Checklist, version rule and tag in `references/release.md`. Tag and push only with the owner's go-ahead. 9. **Report.** Before and after: skills touched, lint and check results (real output), eval scores, always-on token change, what was not tested. ## Rules - One concern per skill. Two skills that share a trigger phrase need a "Not for" line each, or one should absorb the other. - Describe when to use a skill, never its workflow. A description that lists the steps invites the model to follow the summary and skip the body. - No repository names, hostnames, client facts or product rules in any skill. Methods only. - Point to built-ins (`/code-review`, `/security-review`, `/simplify`, `/init`, `claude-api`, `run`, `dataviz`) and to `anthropic-skills:*` instead of copying them. - A script is allowed only for deterministic work: zero dependencies, Node 18 or newer, an assert-based `--selftest`. - Do not edit `NOTICE`, `CONTRACT.md`, `hooks/` or another skill unless the owner asked for that file. - The router and `house-style` text print at the start of every session; every character in a description is paid for in every session. ## Output - Guidance: the usage table summary, the gaps list, and the proposed fix or draft description, in chat. - Full: the changed files with paths, lint and check output tails, eval scores, always-on token cost before and after, version and tag status, and the lines for `NOTICE` if a source was adapted.
Files in this package (8)
- SKILL.md 7 KB
- references/authoring.md 6 KB
- references/descriptions.md 5 KB
- references/evals-howto.md 4 KB
- references/gaps.md 4 KB
- references/release.md 4 KB
- scripts/lint.mjs 10 KB
- scripts/skill-usage.mjs 16 KB
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · lilly-logic-labs/lll-skills | 100 | MIT |
| Trevojan/regency | 100 | MIT |
| kouko/monkey-skills | 100 | MIT |
| himanshu-mamgain/offerstar | 95 | MIT |
Database snapshot: 13.11.2026 · package version 7291508de0e7 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100tokens
Measure and cut what Claude Code spends. Finds session-start context, CLAUDE.md and memory size, skill and MCP listings, per-prompt hook text, cache misses, big tool outputs and unused plugins. Use when asked about token use, context bloat, cost, running out of context, Claude… - 100skilldiet
Shows what your agent skills cost you every session, which ones overlap, and which never fire. Measures the always-loaded skill listing in tokens per skill and per pack (installed skills, plugin skills, or any pack by path), flags skills with near-identical descriptions or the… - 100plan
Turn an idea into an approved design and a task-by-task implementation plan, or review a plan before building. Use when asked to brainstorm, spec, design a feature, "how should we build X", "write the plan", "poke holes in this plan", or CEO, eng or design review. Not for… - 100publishing-a-pack-locally
Lay out a pack so any skill-aware runtime recognises it without a bespoke loader. Use when a new skill pack is ready to be used by an agent runtime on the same machine - 100skill-pack-curator
Designs the membership and structure of a skill pack - persona, member selection, install sequence, cross-links, and what to add or cut. Use when someone asks "which skills should go in this skill pack", "design the membership of a new skill pack", "decide which skills to add or… - 100add-skill
Use in the agent-skills repo when the user asks to add a new shared skill, change or split one in skills/, move a section to a reference file, or fix a failing check. Writes the skill to the rules in AGENTS.md, runs the skill and prose checks and the temp home install, and…