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 genius-filegit 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/genius-file)<a href="https://agentmods.dev/skills/donald-ada/workinggenius/genius-file"><img src="https://agentmods.dev/badge/skills/donald-ada/workinggenius/genius-file/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/genius-file"><img src="https://agentmods.dev/badge/skills/donald-ada/workinggenius/genius-file.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 | $0.00084 | $0.01678 |
| Opus 5.5 | $0.00034 | $0.00671 |
| Sonnet 5.5 | $0.00017 | $0.00336 |
| Haiku 4.5 | $0.00008 | $0.00168 |
Grade A, and why
genius-file 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 21d 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 — 35 lines — stays where its author put it; the contents beside it link to each section on GitHub.
The Work File
One piece of work = one folder at .genius/<slug>/ (a ## Working Genius section in CLAUDE.md/AGENTS.md may pin a different parent), holding:
- the snapshot
<slug>.md, the work's current truth, bounded - the log, everything that happened on the way, append-only, as a tree: the root
<slug>.log.mdnaming one branch per stage, slice or contract version underlog/ - the contract
CONTRACT.md, what binds the slices not yet built - the work's own artifacts
The shape of all of it — templates, the compaction question, the ceiling and its instrument, the invariant, the link rules — is FILE-FORMAT.md, the one home for those rules; the backlog's shape is BACKLOG-FORMAT.md, and the measurements behind the rules are FORMAT-EDGES.md.
The concept: the files, not conversation memory, carry the work — and what a work knows grows by three different laws, so it gets three files with three different rules. Conflate any two and the file grows with every discovery until nobody reads it closely, which is how a record dies while still being written to.
What that takes:
- Read the snapshot before acting; write the moment a decision lands. It outranks whatever you remember about the work. Re-read it after your own context is compacted: a session summarised mid-slice is a new cold reader, and the file is the truth.
- Reading scope stays with the work at hand. Starting or resuming costs three reads: list
.genius/for the slug, read that snapshot, read.genius/DECIDED.mdif settled ground might be at stake.CONTRACT.mdis a fourth, earned when you are about to build, attack or verify against it. Every other slug's files stay closed, and the log stays unopened unless a specific question forces it in, and then only the one branch that question names. An anchor to link is never such a question:python3 ${CLAUDE_SKILL_DIR}/measure.py anchors <slug>lists every branch, every##key and what already links it, because a file opened whole stays in the context to the end of the work. Appending to the log opens nothing:python3 ${CLAUDE_SKILL_DIR}/measure.py append <slug> <branch> <key> [<about>], the entry on stdin, writes the## <key>heading, creates a new branch with its root line, continues a-wipentry, and refuses a key that already exists — the mechanics the format asks of every append, done by the one thing that cannot misremember them. A project with two finished works and one with two hundred cost a session the same. - The contract changes by version, never by patch. Requirements moving mid-build is the normal case. A change of shape — criteria, scope, slices, seams — is the new version whole in
CONTRACT.md, the old whole in the log with what overturned it, one changelog line in the snapshot; a single value that measured differently is theerrataskill's line-level supersede, no bump. So a change costs the same at churn one and churn ten. - Compaction is part of closing, and it has an object. Every slice close, contract bump and stage close puts the format's one question to each snapshot line — does this line still constrain work that isn't finished? — and routes the answer; without the question, "compact to current truth" is satisfied by changing nothing. The invariant — nothing leaves except into the log or
CONTRACT.md, a link left behind — licenses the snapshot alone: the in-flight log is never compacted or tidied, because a kill-reason shortened to a verdict has been destroyed. Tenacity distills a done work's log once, by one announced rule;/distillcatches up older done work;/compactcatches up a snapshot that drifted. - The snapshot is the only path anyone has to know. Every section links the log entries that back it. A cold session that must trawl the log to act is the sign that state leaked into history.
- A slice's evidence rides the commit that closes it: the code, the log entry naming per criterion what ran and what it showed, the snapshot compacted, and
CONTRACT.mdwhere the slice left something binding — one commit where.genius/is committed, the same moment where it is gitignored. The evidence goes in the log, not the commit message: a message can't be corrected, linked, read in the diff, or survive a squash. A close is one write per file, composed before it is written, because every tool call re-reads the whole context. - Absence is the record. A stage the user didn't type has no section and no log entry; not typing a stage is the decision, and when work goes wrong later the missing section is the first suspect.
- Assumptions are visible. A decision made without the user is an
assumed: <question> → <answer>line in Open, surfaced at next contact — an honest assumption beats a hollow confirmation. Consumed, it moves to the log with what consumed it. - Derived work gets its line the moment it's spotted. An edge worth testing, a refactor worth doing, a question worth its own work: one line in
.genius/BACKLOG.md, in the backlog format's shape, in any stage — a seed, not a plan./geniuslights these up. An Open item that is work reaches the backlog only through its own work's log, Tenacity's sweep included (FILE-FORMAT.md, Open); starting a seed removes its line only once the work has taken the seed's content up (BACKLOG-FORMAT.md), because a seed is lossy and one was measured reaching no section of its work. - What binds is rewritten; what records is appended to. The snapshot, the index line, the glossary term get rewritten in place; the log gets appended to and otherwise left alone, because its worth is that it is what was written then. The
errataskill holds the discipline;/reconcilesweeps the drift no session tripped over. - How the sentences are written — for a cold reader and the owner both — is the
record-proseskill's discipline, one place for every document under.genius/. - Checkpoints are live, and easy for the user to read. Each is a real exchange with the user — a model answering its own interview confirms nothing. Where you are asking for a decision, lead with it and keep the sentences short — at Wonder the story you offer for correction is the ask, at Discernment the consequences played back are — because a dense report buys agreement rather than judgement.
What ships with it
5 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
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.
- 21d ago Changed · +7 lines ec2d0ed66ade
- 25d ago Changed 2c9cacb52430
- 1mo ago Changed 551b08b0acfa
- 1mo ago Changed 5a273493adcc
- 1mo ago Changed · -3 lines · +15 tokens per session 9600be678a01
- 1mo ago First seen · 31 lines · 69 tokens per session scan A dcb6d4c06de3
genius-file is a skill published in the GitHub repository donald-ada/workinggenius (6 stars, last pushed 7d ago), licensed MIT. It adds 84 tokens to every session and 1,678 once invoked, about $0.0003 per session on Opus 5.5. 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 2026-08-31.
Other skills, from other repositories
weekly-metrics
Generate a weekly productivity metrics report for solo AI-assisted development. Analyzes git commits, TaskNotes, and collaboration patterns from the last 7 days, then proposes workflow optimizations. Use this skill whenever the user asks for a weekly report, productivity metrics, sprint retrospective, workflow…
claude-autopilot
A time-limited autonomous agent for Claude Code that follows a user directive until a specified deadline. It breaks work into tasks, tracks progress, and winds down before time runs out.
excalidraw
Create Excalidraw diagrams programmatically. Use when the user wants architecture diagrams, flowcharts, or any visual diagram with a hand-drawn aesthetic, where editable output and drag-and-drop editing add value over Mermaid's text-only approach.
ia
Evaluate and improve how information is organized for findability. Use when the user asks to "organize this", "restructure docs", "improve navigation", "where should this go", "review file structure", or when documentation grows beyond a single page.
naming
Evaluate and improve names in code using naming as a design diagnostic. Use when the user asks to "name this", "rename", "review naming", "what should I call", struggles to name something, or when a code review surfaces vague or misleading names.
testing
Design test strategy using Beck's Test Desiderata — which properties matter, which tradeoffs to make. Use when the user asks "how should I test this", "what tests do I need", "review my test strategy", "is this well-tested", or when planning tests for a new feature or refactor.