Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add donald-ada/workinggenius --skill compactgit clone --depth 1 https://github.com/donald-ada/workinggeniusWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/donald-ada/workinggenius/compact)<a href="https://agentmods.dev/skills/donald-ada/workinggenius/compact"><img src="https://agentmods.dev/badge/skills/donald-ada/workinggenius/compact/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/donald-ada/workinggenius/compact"><img src="https://agentmods.dev/badge/skills/donald-ada/workinggenius/compact.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00047 | $0.01465 |
| Opus 5 | $0.00023 | $0.00732 |
| Sonnet 5 | $0.00009 | $0.00293 |
| Haiku 4.5 | $0.00005 | $0.00146 |
Grade A, and why
compact scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 44 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Compact
Every close is supposed to compact — the question in genius-file's format asked of each line, the answers routed. Snapshots drift anyway. /reconcile finds a snapshot over its ceiling or carrying history; this is what fixes one — errata's moves cannot, because nothing here is wrong, it is only in the wrong file.
The concept: the same question, the same destinations, applied where a close did not — and nothing moves before the user has seen what moves. The format owns the question, its destinations and its tie-breaks; reading it there is the first step. What this command adds is the retroactive pass, the proposal, and the checks below.
The counts as this command was invoked — every snapshot, whole and roster excluded, against the ceiling — and both directions of the invariant: links whose target or anchor is missing, log entries nothing links, keys with prose after them. Where a policy notice shows instead (shell injection disabled), run it before proposing:
python3 ${CLAUDE_SKILL_DIR}/../genius-file/measure.py snapshots
python3 ${CLAUDE_SKILL_DIR}/../genius-file/measure.py links
- Scope: a snapshot with a log beside it, in flight or done. Unlike
/distill, being finished is no bar: routing a misfiled line is not correcting a wrong one (errataskill carries the carve-out). The log is appended to and never otherwise touched. A heading whose key carries prose after it is the case that most tempts a pass to break that rule; the format holds the repair, and it is the link, not the heading. ⚠ A link whose target is genuinely not in the log is reported, never repaired — check first that it is not the prose-after-key case. Repairing it would mean writing the log, which this command does not do, or guessing an anchor, which puts a wrong door where a missing one at least reads as missing; the call is the user's. - A work in some other shape is brought to the format as it is routed. A contract still sitting in the snapshot goes to a
CONTRACT.mdcreated for it, in the right layer; a work sitting flat in.genius/moves into its folder with the repository's own move, links rebased in both directions. Lines that answer "arrived with a slice" are never routed to the log to make the number work: that trade demotes live constraints to history. A brand-new work whose log has no first entry has nothing to route into; skip it. - Read before proposing. The snapshot, the
CONTRACT.mdbeside it, and the log's headings. Destinations are decided against the Slices list, so a full log read is never the price of this pass. - Propose, then move. Read-only until the user has seen, per section, what leaves and where to (categories and counts, not a line-by-line ballot), the character count now and after, and every line you were unsure about: one answering two branches at once (say which way the tie-break went), one whose owning slice is not obvious, one whose log heading carries prose after the key. One confirmation covers the batch; a move the user has not seen is a silent one.
- Count the way the format counts, instrument and exclusions both: a full-text count reads a thirty-slice migration as bloated when its prose is half the ceiling, and every remedy you would then reach for is wrong. The count after is the same command run again once the moves are made, never arithmetic on the proposal.
- Under-ceiling is not the same as compacted, and this pass says which it found. A snapshot under its ceiling can still carry a closed slice's paragraph; one over it can be a large roster that is exactly right. Where the question moves nothing, say "nothing to route": that is a result.
- Where the count stays over after routing, say so and stop. The remedies from there are not this command's: a Problem section carrying design, an Open section nobody drained, a cut that is really two pieces of work. Name which it looks like; never shorten a kill-reason to make a number.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago Changed · +8 lines 53c6ceeea911
- 5d ago First seen · 36 lines · 47 tokens per session scan A 6375d0d7ac73
compact is a skill published in the GitHub repository donald-ada/workinggenius (6 stars, last pushed 4d ago), licensed MIT. It adds 47 tokens to every session and 1,465 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-05.
Other skills, from other repositories
daily-working
End-to-end pipeline: pull a task from the project's task tracker by ID (Redmine or GitHub Issues today, more addable via a new adapter), sanity-check and impact-assess it against the codebase before touching anything, implement it with the Claude CLI, verify the result in a real browser via the Claude Chrome extension…
work
Dispatch a tracker ticket into an isolated worktree and drive it autonomously to an open pull request, then stop for review. Pass interactive to brainstorm the ticket with a human before planning it, or a sentence of editorial direction as a rider. Use when asked to work a ticket end-to-end.
batch
Dispatch a night's batch of Todo tickets — one worktree and one pull request each — and leave a morning summary. Run explicitly; never inferred.
prep
Spec a ticket before it is dispatched. Brainstorms it, decides everything it can recommend an answer for, asks a present human about the little that survives, and leaves the spec, the decisions and anything still open as one ticket comment. No worktree, no branch, no pull request, and the ticket stays where it was.
debrief
Review what agents accomplished, surface loose ends across sessions.
dw-prune
Walk .ai/backlog/ and decide every entry out loud: drop what is stale or already done, do what is cheap and unblocked, bundle what ships with a cousin, leave the rest. Explicit-invoke only.