agentleFS
vcxcviivcxcvii/rainmaker/skills/what-actually-worked/SKILL.md

what-actually-worked

Skill0 starsChanged 4 months agoMITView on GitHub

Answer three questions in order: what did we believe, what did we do, what actually happened: and always include a mandatory "What did nothing" section, since a retrospective containing only wins is a defect.

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 2 are distinctive, and 5 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.

  • 5

    Very specific

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

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

    Run rainmaker context --check first.

    line 22
  2. Very specific:

    Only the data/ files listed in this skill's Consumes table.

    line 30
  3. Very specific:

    If strategy.json.context_hash does not match the current hash of context/business.md, say exactly:

    line 32
  4. Very specific:

    If context/business.md carries confidence: stub, continue, and stamp every output with: "Built on a stub context.

    line 38
  5. Very specific:

    What did we do?: Every ledger event with a cause in the period, grouped by cause, effort hours summed, on-site and off-site work together.

    line 54
  6. Distinctive:

    Never assert causation. "Position improved after we shipped X" states timing.

    line 62
  7. Distinctive:

    All five sections are present, section 4 is genuinely non-empty whenever any shipped item failed to move, and no sentence in the report …

    line 88

≈ “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

what-actually-worked

Context load

Run rainmaker context --check first. It prints what exists, what is stale, and exits 1 if anything this skill requires is missing.

Then read, in this order:

  1. context/business.md in full. If absent, stop: "No business context. Run rainmaker audit, then the know-my-buyer skill."
  2. context/voice.md if this skill writes prose. If absent, stop and say so.
  3. context/glossary.md if this skill names products, features or competitors.
  4. data/strategy.json if this skill reads or writes strategy.
  5. Only the data/ files listed in this skill's Consumes table. Never crawl or call an API the core already covers.

If strategy.json.context_hash does not match the current hash of context/business.md, say exactly:

"Business context was edited after the strategy was written. Re-run know-my-buyer, or run rainmaker context --sync to accept the prose as authoritative."

Then stop.

If context/business.md carries confidence: stub, continue, and stamp every output with: "Built on a stub context. Nothing in it came from a buyer. Run know-my-buyer to replace it."

Consumes

FileWhyIf missing
data/ledger.jsonl, data/state.jsonevents with cause, baseline vs currentrun rainmaker audit
data/strategy.json and data/strategy-history/what we believed at the period's startrun know-my-buyer

Produces

Nothing persisted directly; may recommend a decisions entry demoting a pain point, applied via say-it-their-way or know-my-buyer.

Procedure

  1. What did we believe? Diff the current strategy.json against the version at the start of the period, using strategy-history/. Name every changed belief and the reason recorded for it.
  2. What did we do? Every ledger event with a cause in the period, grouped by cause, effort hours summed, on-site and off-site work together.
  3. What happened? baseline against current per finding, only for findings past their own verification window. Use src/ledger/query.ts pendingVerification to exclude anything not yet due, so nothing is judged early.
  4. What did nothing. Mandatory. Call src/ledger/query.ts didNothing, which separates true failures (comparable metrics, no movement) from unmeasured (no comparable metric at all, so no verdict either way). List every shipped intervention whose target metric did not move beyond its window, with the effort spent. A retrospective containing only wins is a defect and must be regenerated, not published as-is.
  5. What we cannot attribute. Findings that moved inside a window containing an algo_update event are reported as coincident and explicitly not causal, with the control: did non-matching pages move too?

Decision rules

  • Section 4 is never skipped, even when every shipped item worked. State plainly that nothing in the period did nothing, if that is genuinely true.
  • Never assert causation. "Position improved after we shipped X" states timing; it does not claim X caused it unless the control (non-matching pages did not move) is also shown.
  • Where a belief was contradicted twice across periods, recommend demoting it in strategy.json with a decisions entry, and name know-my-buyer as the next step if the drift conditions from spec/site-blueprint.md section 8 or the core spec's drift rule fire.

Output

## What did we believe?
<belief>: <held / changed to X, reason: Y>

## What did we do?
<cause>: <n> events, <n>h effort

## What happened?
<finding id>: baseline <x> -> current <y>

## What did nothing
<finding id>: <n>h spent, shipped <date>, <baseline> unchanged

## What we cannot attribute
<finding id> moved inside a window containing an algo_update on <date>.
Control: non-matching pages moved <n>% over the same window, so this is
coincident, not established as causal.

Done when

All five sections are present, section 4 is genuinely non-empty whenever any shipped item failed to move, and no sentence in the report asserts causation without showing its control.

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.