agentleFS
AhmedQureshi89AhmedQureshi89/lean-experiment-kit/.claude/skills/pull-evidence/SKILL.md

pull-evidence

Skill0 starsChanged 19 days agoMITView on GitHub

Pull the current numbers for a Running experiment from the sources its test card names and append a dated entry to its daily-evidence-log.md, compared with the bar. Reads only; never decides, never sends. Use when the founder says "pull the numbers" or during a measurement window.

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read only. Never run a query that changes data, and never send anything.

    line 19
  2. Very specific:

    Context: Read experiments/NNNN-*/test-card.md: the metric, the window, the bar, the exclusions, and the "How we measure" table, which names …

    line 22
  3. Distinctive:

    If a source cannot be reached, say which and why, and write the entry with what you have.

    line 31
  4. Very specific:

    When a source has a connector the Dashboard can query by itself (for example a SQL database), the page can show the count fresh on every …

    line 62

≈ “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

Pull the evidence for experiment NNNN (the argument). This is a read, not a decision: the learning card and the board decision belong to /log-experiment, and only after the window closes.

Rules

  • Exclude test and internal accounts from every count. The test card names them; if it does not, ask the founder once and write the answer into the card.
  • Say when a window overlaps the weekend (weekend in kit.config.json).
  • Personal data (email addresses, phone numbers, names) never goes into a .md file: every .md syncs to the dashboard. Refer to people by count or by an internal id.
  • An interim read is labelled interim. A read after the window closes is labelled final, and still makes no decision.
  • Without per-person tracking the count covers the whole site or product, not the people the test reached. Say so.
  • Read only. Never run a query that changes data, and never send anything.

Steps

  1. Context. Read experiments/NNNN-*/test-card.md: the metric, the window, the bar, the exclusions, and the "How we measure" table, which names each source and the exact query, event or tag. Read the last entry of daily-evidence-log.md so this entry continues it (reuse its query when it has one).
  2. Read each source the table names, with whatever this session can reach:
    • a connected tool (a database, an email service, web analytics) through its connector: load it with ToolSearch, run the exact query or filter from the card, SELECT only;
    • a script in the repo the card names, run as the card writes it;
    • numbers the founder pastes in chat (interviews, a form's export, a payment page). If a source cannot be reached, say which and why, and write the entry with what you have. On the first pull for an experiment, write the exact query into the log entry so later pulls reuse it.
  3. Write the entry. Append to experiments/NNNN-*/daily-evidence-log.md (create it if missing): a heading with the date and time, the window and how much of it has elapsed, a table of the counts, the count against the bar ("N of M needed"), interim or final, the caveats, and the queries used.
  4. Update the card header so the dashboard shows the read. result stays the record in words (the whole read in one sentence), and six fields beside it are what the dashboard draws as a scoreboard, so fill them in the same edit:

    FieldWhat goes in it
    read_atwhen this read was taken, ISO with the utc_offset from kit.config.json. Required while the card is running: build_data.py refuses to build without it.
    score_nowthe one count that answers the bar, as a plain number
    score_targetthe bar as a plain number, so the page can draw how far there is left to go
    score_labelwhat those two numbers count, in a few words ("visitors who clicked Subscribe")
    score_partsthe counts behind it, each { n: <number or short string>, label: "what it counts" }. At most 12 characters in n, never a | or a ·. Leave out any count the page already shows live (live_dataset).
    score_caveatone line: what not to trust about this read

    score_now and score_target must be comparable: where the bar is a share, convert it to a head count. Where a read has no number against the bar yet, leave the score_* fields empty and still set read_at. Check status matches reality (running while the window is open; done only after a learning card, which is /log-experiment's job). Set window_start and window_end if they are missing.

  5. Sync. Run /sync-dashboard.
  6. Tell the founder in three lines: the count against the bar, interim or final, and when the next read is due. If the window has closed, say that /log-experiment NNNN is the next step and wait for their word.

A live count on the dashboard (optional)

When a source has a connector the Dashboard can query by itself (for example a SQL database), the page can show the count fresh on every open: write the window-count query, run it once, then ArtifactData set collection datasets, doc live-NNNN on the dashboard URL with {title, description, source: {kind: "query", connector: "<connector name>", tool: "<its query tool>", args: {<the tool's arguments, with the query as run>}}}, and put live_dataset: live-NNNN in the card header. Every column of the query's first row becomes a tile on the running card.

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.