Delegation Audit
Free for noncommercial use under PolyForm Noncommercial 1.0.0; commercial use needs a licence from Rico. View the source
The owner talks through everything they did this week. The kit interviews them one task at a time, grades each task with a fixed lookup table, and builds a printable one-page HTML audit. The audit shows what stays with the owner, what moves to someone on their team (with or without AI help), and how many hours come back each month. Dollar figures appear only when the owner states their own hourly rate.
Experimental
The problem
Owners of small businesses get stuck doing everything themselves. They rarely know how much of their month goes to work that does not need their judgment, and generic "use AI for everything" advice ignores the half of local business that works with its hands. A useful audit has to use the owner's own numbers, refuse to guess, and never claim software can do physical work.
How it works
Five markdown skills run a four-stage pipeline: task-dump captures the week without judging it; task-grilling asks question zero (does this happen on a screen or out in the world?) and then four fixed questions per task (the action, hours per week, what breaks if it goes wrong, and whose judgment it needs); a two-block lookup table turns those answers into one of five verdicts; audit-page renders the self-contained HTML page. A plain-Node core (skills/delegation-audit/core.mjs) computes the verdicts, frequency conversions and arithmetic. Two verifiers check every finished page (verify-html.mjs) and working file (verify-own-data.mjs).
Features
- Deterministic verdicts
- Named refusal: a task too vague to price is not graded
- Rate refusal
- No AI claims on physical work
- Stated arithmetic
- Prompt-injection hardening
- Mechanical page self-check
- Stated capability tiers
- Three worked examples
How it was built
Version 0.1.0 was built on 2026-08-12, out of an audit of the kit line the day before, as a delegation audit business coaches run with clients: take the owner's real week and show what could move off their plate.
It was a re-skin of the Idea Forge pipeline with no new mechanism invented. Its skills map one to one onto Idea Forge's, with a print-first page meant to be handed across a table.
A lookup table decides the verdict, not the model. The arithmetic is fixed and printed with its working, and no dollar figure appears unless the owner states their own rate.
Every hand-off names a person who takes ownership of the outcome, so the audit never frames work as removing a job.
The same afternoon a Claude Opus agent ran the kit as three invented buyers and found nineteen defects, all fixed. Rico kept the owner and the coach in charge of every judgment the page records.
Credits
The one-question-at-a-time interview mechanism is adapted from Matt Pocock's grilling and grill-me skills (github.com/mattpocock/skills, MIT). That licence is reproduced in skills/task-grilling/LICENSE.upstream and covers the adapted portions only.