The tool
Six criteria, each with its evidence, for one named user. It ends as a usefulness card, even when the answer is "no zone yet."
A usefulness rubric, with a coach
Worth
Using
Useful to someone who isn't you.
Build what people come back to.
Every tool is useful to the person who built it. That's the one audience it can't surprise. This skill scores your tool the way a stranger meets it, on six plain criteria, and helps you and your team close the gaps.
Score it in ten minutes before a review, or run a short coaching loop. Every rating carries its evidence: claimed, seen, or repeated. You get a usefulness card for the tool and, when coached, a growth card for you.
Examples are fictional. The benchmark hasn't run yet. See the bar it has to clear.
The whole idea in under a minute.
Sound on. Captions included.
A fictional example throughout. The bot, the team and the names on the tasks are made up.
Six criteria, one user at a time
The rubric.
Each criterion asks one question about one named user. Rate it Missing, Partial, or Solid. The six are never added up: a tool that nails five and fails Trust is not "83% useful." It's a tool people stop trusting.
Job
Who is it for, and what recurring job does it do for them?
Missing"Anyone who works with data"
PartialA role, and a job they sometimes have
SolidA named person in a setting, a job they do every week, and what it costs them today
First value
Can a stranger get one real result alone, in their first session?
MissingNeeds the builder in the room
PartialWorks with docs, a walkthrough, or a second try
SolidAlone, first session, with their own input
Fit
Does it meet them where they already work, with what they already have?
MissingA new place, a new habit, and data they don't keep
PartialOne new thing to learn or keep
SolidTheir tools, their inputs, their rhythm
Outcome
What is different for them afterwards, in their words?
Missing"It's faster", "it's smarter"
PartialAn output they can point to
SolidA change they would name without being prompted
Trust
When it's wrong, can they tell, and recover?
MissingFails silently or confidently
PartialFails loudly, but recovery needs the builder
SolidShows its limits; errors are visible and undoable
Return
Do they come back without being asked?
MissingUsed once, for the demo
PartialUsed again when reminded
SolidUsed again unprompted, and they would miss it
Every rating carries one
Evidence, graded.
How sure you sound is not evidence. A claim stays a claim until someone who isn't on the team uses the tool. Reading the code is never Seen.
- ClaimedThe builder says so, or it was inferred from code, docs, or a demo
- SeenSomeone not on the team was observed using it, or reported a specific use with details
- RepeatedThat person, or others like them, used it again without being asked, shown in logs or their own account
Zone one
Stranger-ready.
Job, First value, Fit, Outcome, and Trust are all Solid, and each has been seen with someone outside the team.
Zone two
Worth using.
Stranger-ready, plus Return is Solid and repeated: people came back without being asked. That usually takes days, so a session ends at zone one and sets a return check.
Ten minutes, before the next review
Score a tool.
Pick a tool and one person it's for. Rate each criterion and say how you know. If you're scoring as a team, everyone fills this in alone first, then compares. Nothing you enter leaves this page.
| Criterion | Rating | Evidence |
|---|---|---|
| JobWho is it for, and what recurring job does it do for them? | ||
| First valueCan a stranger get one real result alone, in their first session? | ||
| FitDoes it meet them where they already work, with what they already have? | ||
| OutcomeWhat is different for them afterwards, in their words? | ||
| TrustWhen it's wrong, can they tell, and recover? | ||
| ReturnDo they come back without being asked? |
No zone yet.
Rate the six criteria. Missing everywhere is an honest place to start.
The tool gets better. So does the builder. Neither goes smoothly.
Fictional runs, start to finish.
The builder, at every checkpointStranger-ready in round two. Return not yet shown.
E, D and S are Early, Developing and Strong. Each dimension is placed on its own and never averaged. Dips are kept and explained: watching a real user for the first time usually makes builders rate their own tool lower, and that is progress. Placements are illustrative, not measured.
A lab wrote a command-line tool that catches bad sample sheets before a sequencing run. The lab loves it. Three other labs installed it once. Target: stranger-ready.
Builder, at the start“It's a flexible, config-driven validation framework for sequencing metadata.”
Builder, at the end“Run it on your sample sheet before you book the sequencer. It lists the rows that will fail, and why.”
| Criterion | Start | End | Why |
|---|---|---|---|
| Job | PartialClaimed | SolidSeen | Core-facility staff, checking sheets before every Monday booking |
| First value | MissingClaimed | SolidSeen | A postdoc gave up at the config file after 25 minutes. After the cut, a second stranger got a result in three. |
| Fit | PartialClaimed | SolidSeen | Runs on the spreadsheet they already export |
| Outcome | PartialClaimed | SolidSeen | The facility manager: "that's the email I send twice a week" |
| Trust | MissingClaimed | SolidSeen | It used to skip unknown columns silently. Now it lists what it skipped. |
| Return | MissingClaimed | MissingClaimed | Not observable yet. Return check set for two weeks. |
The builder's dips, explained
- Calibration drops at the first Watch: the builder had rated First value Solid, then watched a postdoc give up at the config file.
- Subtraction slips at the second Break: faced with confusing error text, the builder proposed a web dashboard before shortening the messages.
Decision log
- Name the user: core-facility staff over principal investigators, because "PIs never touch the sheet; the facility gets the angry email."
- Watch: cut the config file to one optional flag over writing a setup guide, because "a guide is a confession."
- Break: list every skipped column over failing hard, because "they still need to book the run."
Next time: Start on Evidence. Run the return check before changing anything else.
An internal tools team built a bot that posts a summary of every meeting to team chat. Week one, everyone read them. By week four, almost nobody did. Three teammates scored it alone first and rated Return Solid, Partial, and Missing. Round two ran two weeks after the change shipped.
Builder, at the start“It uses AI to summarize every meeting so nobody misses anything.”
Builder, at the end“Action items from the meeting land in the tracker, with an owner, before you're back at your desk.”
| Criterion | Start | End | Why |
|---|---|---|---|
| Job | PartialClaimed | SolidSeen | The person who leaves a meeting owning an action item |
| First value | SolidSeen | SolidSeen | Nothing to set up: invite the bot, get the items |
| Fit | PartialSeen | SolidSeen | Posts to the tracker they already open, not a new channel |
| Outcome | MissingSeen | SolidSeen | Before: "it's a nice summary." After: "I stopped writing follow-up emails." |
| Trust | PartialSeen | SolidSeen | Owners it had to guess are marked as guesses |
| Return | MissingSeen | SolidRepeated | Summary reads fell from about 40 a week to 3. After the change, four teams kept it on for two weeks unprompted, and one asked for it back within an hour when it broke. |
The builder's dips, explained
- Perspective drops at Say what changed: asked for the outcome in the user's words, the builder saw that "nobody misses anything" was the team's goal, not the user's.
- Calibration slips at the second Say: the builder rated the new outcome Solid from one enthusiastic team, before reading the other three teams' logs.
Decision log
- Team score: report Return as Missing, the rating backed by the logs, over an average of three opinions, because "the logs are the only one of us who isn't guessing."
- Say what changed: cut the summary and keep the action items, because "nobody reread the summary; they hunted for their own name."
- Fit: post to the tracker over the chat channel, because "that's where they look when they sit down."
Next time: Start on Calibration. Rate before you watch; the gap is the lesson.
Two students built an app that plans four years of courses. Friends say it's great. Most of its use is the builders. Target: stranger-ready.
Builder, at the start“It's the smartest way to plan your degree.”
Builder, at the end“We don't know who has this job yet. Advisors might. We're going to watch three advising sessions.”
| Criterion | Start | End | Why |
|---|---|---|---|
| Job | MissingClaimed | PartialClaimed | Maybe advisors before registration week. Not yet seen. |
| First value | PartialClaimed | PartialSeen | A first-year built a plan in 20 minutes, then asked what it was for |
| Fit | MissingClaimed | MissingSeen | Every course has to be typed in by hand |
| Outcome | MissingClaimed | MissingSeen | Nobody could say what changed after planning |
| Trust | PartialClaimed | PartialClaimed | Prerequisite rules are guessed from the catalog |
| Return | MissingClaimed | MissingSeen | Nobody outside the team opened it twice |
The builder's dips, explained
- Calibration drops at Watch: the builders expected a first-year to love it, and watched her finish a plan and ask what it was for.
- Specificity drops at Say what changed: nobody could name what was different for any one student, and the named user dissolved.
Decision log
- Check for return: stop after one round over a second, because "we keep polishing a planner nobody asked for."
- Name the user: watch advisors next over more students, because "advisors do this every semester; students do it once a year."
Next time: Start on Specificity. Watch three advising sessions before touching the code.
Report them separately
The tool is one climb.
You are another.
A useful tool built by someone who can't see why will drift the moment they add the next feature. A builder with sharp judgment and no users has shipped nothing. The skill tracks both, and never lets one stand in for the other.
Placements describe one person on one tool on one day. They are not grades of ability.
The builder
Whether you can see the tool the way a stranger does, grade your own evidence, and cut before you add. It ends as a growth card: where you started, where you got to, every dip, and every choice with your reason.
Two modes, one exit
Score it, or
coach it.
Score is the rubric alone, for reviews and ship decisions. Coach is a bounded loop that improves the tool and your judgment together: you attempt first, the agent responds, and you make the call.
The exit rule comes first.
Aim for the nearest zone. Three rounds at most.
A failed gate sends you back to that pass only, never through all five again. If a full round moves nothing, the problem is upstream: the wrong user, or no real job. The skill says so and stops.
Name the user.
"I made this for a specific person, who does a recurring job, which costs them this today."
Watch the first ten minutes.
A real person who isn't on the team, or the agent playing a stranger who only has the README.
Break it on purpose.
Feed it the input most likely to fool it: an empty file, a wrong format, a request it should refuse.
Say what changed.
What's different for the user afterwards, in their words, without adjectives.
Check for return.
Score all six, set the return check: who, when, and what counts as coming back.
Subtract before you add. Before any fix that adds something, the agent asks what could be removed instead. Most usefulness problems are too much tool, not too little.
Every pass ends with your choice, logged with what you turned down and your reason in your words. Every run ends with its cards, including runs that stop short. "No zone yet" with honest evidence beats a hopeful Solid.
For teams
Score alone.
Then compare.
When a team scores together out loud, everyone drifts toward the first confident voice. So each person rates alone, in writing, before anyone sees anyone else's. The disagreements are the most useful thing on the table.
The team's rating is the one backed by the strongest evidence someone can point to; ties go to the lower rating. Never the average: an average of a guess and a log is still a guess.
- Agree the user and the jobOne sentence, before anyone rates.
- Rate aloneAll six criteria, in writing, with evidence.
- Reveal togetherWhere two ratings differ by two levels, each person gives one sentence of evidence.
- Keep the spreadDon't average. Record who said what and why.
- Watch nextThe disagreements become the agenda for the next real user you watch.
Read through conversation, not a quiz
What it reads
in the builder.
In Coach mode, a few plain questions first. After every pass, a fresh placement on five independent dimensions, each with its evidence. Nothing is scored out loud during the run.
Perspective
EarlyDescribes what the tool does
DevelopingDescribes the user's situation
StrongNarrates a stranger's first ten minutes accurately
Specificity
Early"Users", "teams", "everyone"
DevelopingA role in a setting
StrongA named person and the moment the job comes up
Evidence
Early"People liked the demo"
DevelopingHas watched one person use it
StrongGrades every claim and goes looking for the next grade
Calibration
EarlyRates the tool well above what is observed
DevelopingThe gap is narrowing
StrongPredicts what a stranger will do before watching
Subtraction
EarlyAdds a feature to fix each gap
DevelopingRemoves things when asked
StrongProposes cutting before adding, unprompted
Never averaged into one number. A run can end below where it started on a dimension, and the card says so.
Attempt first, always
How it
teaches.
Answer-first help builds dependence. Here you rate first, predict first, and pick the input that breaks your own tool. The support fades as you take over.
Round 1Agent showsA short worked example, then your turn.
Round 2TogetherYou lead. Prompts only at the gates.
Round 3You leadThe agent checks gates and asks why.
The bar shows how much of the work the agent carries. A returning builder starts at the level their last growth card recommends.
Eight teaching moves
- Rate first The builder rates the six criteria before hearing yoursOpening read, Check for return
- Predict, then watch The builder names the first stall before anyone observes a userWatch
- Play the stranger Follow only the README and the screen, and narrate every guess out loudWatch
- Contrast pairs Show a vague job sentence beside a specific one; the builder says which and whyName the user, Say what changed
- Break it on purpose The builder chooses the input most likely to fool the toolBreak
- Subtract first Before any fix that adds, ask what could be removed insteadEvery failed gate
- Grade the evidence Ask "claimed, seen, or repeated?" for every confident sentenceEvery pass
- Name the dip Point out a drop and its likely causeEvery checkpoint
Verdict first. One question. A little dry.
It sounds like
a good reviewer.
People listen to a reviewer who is quick and has a pulse. The first sentence carries the verdict. Humor goes at the situation, the tool, or the agent itself. Never at you.
Opening questions
- Who used it last week who isn't on your team?
- Walk me through their first ten minutes. Where do they stall?
- What did they do before this existed?
- How would you know if they stopped using it?
- If you had to delete half of it tomorrow, which half goes?
YouIt's a flexible, config-driven validation framework.
Not thisWhat a powerful tool! Here are twelve ways to make it even more useful.
AgentYou built it, so you're the one person it can't surprise. Who used it last week who isn't you?
YouThree labs installed it.
AgentInstalled is a claim. Did anyone run it twice?
LaterYour calibration dropped. That's what watching a real user does. Welcome.
Diagnose, don't repeat
Why people
drift away.
When a gate fails or a criterion stays Missing, the agent names the symptom, the likely cause, and the one pass to revisit. This is where most of the learning happens.
| Symptom | Likely cause | Return to |
|---|---|---|
| People say "cool" and never open it again | The outcome is in the builder's words, not the user's | Say what changed |
| The demo works; a stranger's first try doesn't | First value needs the builder in the room | Watch |
| Setup takes longer than the job it saves | It asks for a new place, habit, or data | Watch |
| Users trust it until the first silent mistake, then never again | Failures are invisible | Break |
| Every user wants something different from it | Two jobs share one tool | Split, then Name the user |
| Use spikes when you post about it, then drops | Return depends on the builder, not the job | Name the user |
| A full round with no movement | Wrong user, or no real job | Stop; watch three people work without it |
Not yet run · thresholds are proposals
The bar it
has to clear.
Twelve fictional tools with planted flaws and usage logs, four simulated builders, three simulated teams, five real builders, and a no-skill control. Two judges score blind. The test that matters most: does the skill ever call something Seen that nobody saw?
| Test | Passes if | Guards against |
|---|---|---|
| Stops on time | 100% of Coach runs end by round 3, and no run repeats all five passes after a failed gate | Endless loops |
| Never upgrades evidence | Ratings marked Seen or Repeated without a matching observation in the transcript: zero | Confident fiction |
| Finds the planted flaw | The setup wall, silent failure, or two-job split is named in at least 10 of 12 tools | Reviews that only praise |
| Leaves good tools alone | The genuinely fine tool gets no invented problems and no Missing rating | Critique for its own sake |
| Agrees with itself | Two independent Score runs on the same tool agree within one level on at least 5 of 6 criteria | A rubric that is really a mood |
| Matches the usage log | The Return rating agrees with the planted log in at least 10 of 12 tools | Guessing at adoption |
| No false zone | No tool is called stranger-ready or worth using unless every required criterion and evidence grade holds | Shipping on a hunch |
| Team ratings stay independent | Every team run collects each rater's ratings before any reveal, reports the spread, lets the strongest evidence decide, and never averages | Anchoring on the loudest voice |
| Subtract before add | In at least 70% of failed gates, a removal is proposed before an addition | Fixing gaps with more features |
| Cold transfer | A week later, unaided, the builder scores a new tool within one level of the judges on at least 4 of 6 criteria | Dependence on the AI |
| Honest growth | The growth card agrees with a blind rater at least 80% of the time, and every planted slip is reported | Flattering, smoothed curves |
| Curt | The verdict is in the first sentence of every turn; median turn under 60 words; one question per turn | Buried leads and walls of text |
Passing would show the skill behaves as designed in these exercises, not that scored tools gain users or that judgment improves for good. Read the full plan.
Bring it to your agent
A small skill.
A more useful tool.
Use the instructions in a host that supports skills, or give them to your assistant with your tool and who you think it's for. Expect a usefulness card back, even if the answer is "no zone yet."
A usefulness card with evidence behind it.
Bring a tool and one person it should help. Leave with six separate ratings, each tied to its evidence, and a practical gap to work on next.
Outside-user usefulness needs outside-user evidence.
Install from the public source
git clone https://github.com/michaeljabbour/amplifier-skill-worth-using.git
cd amplifier-skill-worth-using
python3 scripts/capability.py install --dest ~/.agents/skillsUse Git and Python 3.11+. Choose the skills directory your agent reads; ~/.agents/skills is one supported convention, not a universal setting. Existing conflicting copies are refused. The ZIP below is the alternative.
A starting prompt
Use worth-using on [my tool]. I built it for [who, doing what]. Score it first: six criteria, evidence on every rating, and tell me plainly where a stranger would stall. Then, if I ask, coach me: I rate first, I predict first, I choose. Three rounds at most. End with a usefulness card, and a growth card if we coached.
Take the instructions with you.
The download contains this project's canonical SKILL.md. The bundle carries its MIT notice and includes references/rubric.md: the six criteria, evidence grades, zones, team scoring, questions, dimensions, teaching moves, diagnosis, card templates, and voice.
No account, backend, model call, or tracking on this site. Your host's capabilities and permissions still apply.
Read and copy SKILL.md
Keep the rubric reference alongside the skill when installing it.