agentleFS
ai-analyst-labai-analyst-lab/ai-analyst/.claude/skills/always-compare/SKILL.md

always-compare

Skill304 starsChanged 2 months agoMITView on GitHub

Never present a metric or number in isolation; anchor every number to a comparison (prior period, benchmark, or another segment) or state that none is available. Use every time you show a number: rates, counts, revenue, averages, query results, chart labels, summary stats. Trigger on output containing "the rate is", "we saw", "total", "average", "conversion", "revenue", "users", "sessions", "churn", "AOV", "NPS", or any figure pulled from data.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 5 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Before presenting ANY number to the user — in chat, in a report, in a chart caption, in a Slack message, or on a slide.

    line 19
  2. Distinctive:

    Do the subtraction for the reader.

    line 54
  3. Distinctive:

    Do not silently drop the comparison.

    line 59
  4. Distinctive:

    Every number has an anchor.

    line 90
  5. Distinctive:

    Never present a bare number: If you catch yourself writing "X is N", stop and add the comparison before sending.

    line 94
  6. Distinctive:

    Never fabricate a comparison: If the prior-period data doesn't exist, say it doesn't exist — do not estimate, extrapolate, or reach for a …

    line 95

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Skill: Always Compare

Purpose

A number alone is not an insight. "Conversion rate is 3.2%" tells the reader nothing actionable — they cannot tell if that is a crisis or a record high. This skill enforces one rule: every metric ships with a comparison.

When to Use

Before presenting ANY number to the user — in chat, in a report, in a chart caption, in a Slack message, or on a slide. This runs on every analysis output, alongside question-framing (which runs at the start; this one runs at the end).

The Rule

Never show a number alone. Anchor it to at least one comparison.

Pick the comparison that best serves the decision. In priority order:

#Comparison TypeUse WhenExample
1vs. prior periodThe question is "is this changing?""down from 4.1% last month"
2vs. benchmark / averageThe question is "is this normal?""below the 3.8% site-wide average"
3vs. another segmentThe question is "who is affected?""vs. 5.4% on desktop"

Two comparisons beat one. A prior-period delta plus a benchmark tells the reader both the direction and the altitude. Use both when you have both.

Instructions

Step 1: Find every number in your draft output

Scan the response you are about to send. Every figure — headline stats, table cells, chart annotations, sentences in the narrative — is in scope.

Step 2: Attach a comparison to each one

For each metric, ask: compared to what? Then pull the comparison from the data:

  • Prior period: same metric, previous week / month / quarter (match the grain of the metric)
  • Benchmark: site-wide average, cohort average, target, or historical baseline
  • Segment: the same metric for a contrasting slice (mobile vs. desktop, new vs. returning, channel A vs. B)

Compute the comparison in the same query where practical — it is cheaper and less error-prone than a second round trip, and it guarantees the filters match.

Step 3: State the delta, not just both numbers

Do the subtraction for the reader. "3.2%, down from 4.1%" is better than "3.2% (last month: 4.1%)". Give direction (up/down) and magnitude (absolute points or relative %) — and be explicit about which you are using: "down 0.9pp (a 22% relative decline)".

Step 4: If you have no comparison data, SAY SO

Do not silently drop the comparison. An unanchored number must carry an explicit flag:

"Conversion rate is 3.2% (no prior period available for comparison)."

Other honest forms:

  • "(first month of data — no baseline yet)"
  • "(no site-wide benchmark defined; recommend establishing one)"
  • "(segment too small to compare — n=14)"

This is a standing obligation of the method (analyst-core): always flag when data is insufficient.

Examples

Bad -> Good

BadGood
"Conversion rate is 3.2%""Conversion rate is 3.2% — down from 4.1% last month, and below the 3.8% site-wide average."
"We had 12,400 orders in June.""We had 12,400 orders in June, up 8% from May (11,500) and the highest month of 2024 so far."
"AOV is $58.""AOV is $58, essentially flat vs. Q1 ($57), but members average $74 vs. $51 for non-members."
"Checkout drop-off is 34%.""Checkout drop-off is 34% on mobile vs. 19% on desktop — mobile accounts for 78% of all abandoned carts."
"NPS is 41.""NPS is 41 (no prior quarter available — this is the first survey wave, so treat as the baseline)."

Example: a full finding, done right

Mobile conversion is the problem. Mobile converts at 2.1% vs. 5.4% on desktop — a 3.3pp gap (61% lower relative). The gap widened from 1.9pp in Q1, driven entirely by the payment step, where mobile drop-off is 44% vs. the 26% funnel-wide average. Source: sessions + events, Jan 1-Jun 30 2024, excludes bot traffic.

Every number has an anchor. The reader knows instantly what to do.

Anti-Patterns

  1. Never present a bare number. If you catch yourself writing "X is N", stop and add the comparison before sending.
  2. Never fabricate a comparison. If the prior-period data doesn't exist, say it doesn't exist — do not estimate, extrapolate, or reach for a plausible-sounding benchmark you didn't compute.
  3. Never compare across mismatched filters. The comparison must use the same definition, filters, and exclusions as the metric — otherwise the delta is an artifact. Re-check the WHERE clause on both sides.
  4. Never mix up percentage points and percent. 4.1% -> 3.2% is down 0.9pp, which is a 22% relative decline. Saying "down 22%" without the "relative" qualifier misleads; saying "down 0.9%" is simply wrong.
  5. Never compare against a period distorted by a known event without flagging it — a holiday spike, an outage, a launch, or a backfill. Check .knowledge/datasets/{active}/quirks.md before choosing a baseline period.
  6. Never bury the comparison in a footnote. It belongs in the same sentence as the metric — that is where the reader forms their judgment.
  7. Never let charts escape the rule. A bar chart of one period is a bare number in visual form. Show the prior period, a benchmark line, or a segment split.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.