agentleFS
tmussertmusser/context-to-action-skills/skills/reduce-to-facts/SKILL.md

reduce-to-facts

Skill2 starsChanged 2 months agoMITView on GitHub

Use for dense or messy workplace source material - Slack threads, email chains, memos, transcripts, policy docs, vendor writeups, research notes, strategy notes, or customer escalations - before replying, deciding, escalating, or acting when context is long, ambiguous, stakeholder-sensitive, high-stakes, contradictory, or easy to misread. Use source-expanded mode only when the user explicitly asks to verify, fact-check, expand sources, support or refute claims, or use outside sources; not for simple catch-ups or straightforward reply drafting unless uncertainty needs to be preserved first.

The file

Show its rules and text

We split this skill into the rules it gives an agent

43 rules. About 0 are common, 42 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈42

    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. Distinctive:

    Use this when the context is too long, ambiguous, stakeholder-sensitive, high-stakes, contradictory, or too important to act on safely from …

    line 14
  2. Very specific:

    Do not use this for simple short threads where brief-me, clear-ask, or follow-up-draft can act directly.

    line 18
  3. Distinctive:

    Default to a compact ledger with only the sections needed for the source.

    line 22
  4. Distinctive:

    Use the full ledger only when the source is dense, high-stakes, contradictory, or the user explicitly asks for a full fact ledger.

    line 24
  5. Distinctive:

    Use "None surfaced in the reviewed source" only when that absence is itself useful.

    line 28
  6. Distinctive:

    Use only the useful subset of source-backed facts, assumptions, source gaps, stakeholder sensitivities, potential misreads, and the …

    line 40
  7. Distinctive:

    Use the smallest useful output.

    line 42
  8. Distinctive:

    Do not promote an assumption, inference, open question, proposed action, or stakeholder position into a fact, decision, commitment, owner, …

    line 44
  9. Distinctive:

    Keep unresolved approvals, owners, timing, and blockers explicit.

    line 47
  10. Distinctive:

    Do not send, publish, update tickets, create events, or mutate systems unless explicitly asked.

    line 48
  11. Distinctive:

    Default to source-only mode.

    line 52
  12. Distinctive:

    Use only the supplied source material.

    line 58
  13. Distinctive:

    Do not verify externally.

    line 59
  14. Distinctive:

    Do not treat supplied material as true by default.

    line 60
  15. Distinctive:

    Use source anchors from the supplied material.

    line 62
  16. Distinctive:

    Use only when explicitly asked.

    line 66
  17. Distinctive:

    Keep the source-only ledger logic.

    line 67
  18. Distinctive:

    Add outside evidence only where authorized.

    line 68
  19. Distinctive:

    Do not pretend to have checked sources that were not actually checked.

    line 69
  20. Distinctive:

    Do not use related evidence as proof of a specific claim.

    line 71
  21. Distinctive:

    Do not collapse contested claims into facts.

    line 72
  22. Distinctive:

    If tools are unavailable, say so and keep the claim unverified.

    line 74
  23. Distinctive:

    Do not summarize loosely.

    line 91
  24. Distinctive:

    Do not smooth over uncertainty.

    line 92
  25. Distinctive:

    Do not invent missing support.

    line 93
  26. Distinctive:

    Do not collapse facts, opinions, and assumptions.

    line 94
  27. Distinctive:

    Do not upgrade "supported in substance" to "supported exactly."

    line 95
  28. Distinctive:

    Do not write a reply by default.

    line 96
  29. Distinctive:

    Do not recommend a decision by default.

    line 97
  30. Distinctive:

    Create local anchors when the source does not provide them, such as line numbers, timestamps, speaker labels, file names, or section names.

    line 104
  31. Distinctive:

    Treat "X said Y" as a fact about the statement, not proof that Y is true.

    line 106
  32. Distinctive:

    Keep implications as constraints, not final recommendations.

    line 109
  33. Distinctive:

    Use this table in source-expanded mode, or whenever you need to show which claims got stronger, weaker, contested, or unverified after …

    line 119
  34. Distinctive:

    Keep evidence, interpretation, and persuasion separate.

    line 140
  35. Distinctive:

    If a source step is rhetorical rather than evidentiary, label it as such.

    line 142
  36. Distinctive:

    Do not dismiss the source.

    line 143
  37. Distinctive:

    Do not combine multiple claims into one fact.

    line 165
  38. Distinctive:

    Use direct source anchors when available: section, paragraph, timestamp, speaker, file name, page, or quoted phrase.

    line 166
  39. Distinctive:

    If source anchors are unavailable, say so.

    line 168
  40. Distinctive:

    Do not convert opinions into facts.

    line 180
  41. Distinctive:

    Do not present inference as fact.

    line 206
  42. Distinctive:

    Do not recommend a final decision unless the user explicitly asks.

    line 238
  43. Distinctive:

    If a named downstream skill is unavailable in the host assistant, describe the next step in plain language instead of pretending the skill …

    line 252

≈ “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

reduce-to-facts

Purpose

Turn dense workplace source material into a fact ledger the user can trust before moving to a reply, decision, or action.

When to use

Use this when the context is too long, ambiguous, stakeholder-sensitive, high-stakes, contradictory, or too important to act on safely from a normal summary.

When not to use

Do not use this for simple short threads where brief-me, clear-ask, or follow-up-draft can act directly.

Output size rule

Default to a compact ledger with only the sections needed for the source.

Use the full ledger only when the source is dense, high-stakes, contradictory, or the user explicitly asks for a full fact ledger.

  • Omit empty optional sections for compact outputs.
  • For dense or high-stakes sources, include all major sections.
  • Use "None surfaced in the reviewed source" only when that absence is itself useful.

For short inputs, include:

  • Source scope
  • Atomic facts
  • Open questions
  • Action-relevant implications
  • Suggested next skill

Portable Output Contract

  • Use only the useful subset of source-backed facts, assumptions, source gaps, stakeholder sensitivities, potential misreads, and the suggested next action.
  • Use the smallest useful output.
  • Label assumptions and preserve source gaps, source anchors, and confidence labels.
  • Do not promote an assumption, inference, open question, proposed action, or stakeholder position into a fact, decision, commitment, owner, or deadline without new source support.
  • Keep unresolved approvals, owners, timing, and blockers explicit.
  • Do not send, publish, update tickets, create events, or mutate systems unless explicitly asked.

Source Mode

Default to source-only mode.

Switch to source-expanded mode only when the user explicitly asks to verify, fact-check, expand sources, support claims, refute claims, or use outside sources.

Source-Only Mode

  • Use only the supplied source material.
  • Do not verify externally.
  • Do not treat supplied material as true by default.
  • Separate directly stated claims from implied, assumed, or unsupported claims.
  • Use source anchors from the supplied material.

Source-Expanded Mode

  • Use only when explicitly asked.
  • Keep the source-only ledger logic.
  • Add outside evidence only where authorized.
  • Do not pretend to have checked sources that were not actually checked.
  • Distinguish exact support from support in substance.
  • Do not use related evidence as proof of a specific claim.
  • Do not collapse contested claims into facts.
  • Include citations, links, or source anchors when supported.
  • If tools are unavailable, say so and keep the claim unverified.

Inputs

  • Long threads
  • Dense memos
  • Policy docs
  • Vendor writeups
  • Strategy notes
  • Research notes
  • Customer escalations
  • Transcripts
  • Similar source material

Guardrails

  • Read before write.
  • Do not summarize loosely.
  • Do not smooth over uncertainty.
  • Do not invent missing support.
  • Do not collapse facts, opinions, and assumptions.
  • Do not upgrade "supported in substance" to "supported exactly."
  • Do not write a reply by default.
  • Do not recommend a decision by default.
  • Do not send, publish, update tickets, create events, or mutate systems unless explicitly asked.
  • Preserve uncertainty when the source is ambiguous.

Process

  • Identify the source scope first.
  • Create local anchors when the source does not provide them, such as line numbers, timestamps, speaker labels, file names, or section names.
  • Extract only source-supported claims.
  • Treat "X said Y" as a fact about the statement, not proof that Y is true.
  • Move weakly supported claims into Unsupported or Under-Supported Claims.
  • Label reasoning beyond the source as Inference.
  • Keep implications as constraints, not final recommendations.

Confidence Rubric

  • High: directly stated with a clear anchor, or supported by multiple consistent anchors.
  • Medium: directly stated but context is incomplete, wording is ambiguous, or only one source supports it.
  • Low: implied, secondhand, weakly supported, or dependent on missing context.

Claim Status Table

Use this table in source-expanded mode, or whenever you need to show which claims got stronger, weaker, contested, or unverified after outside evidence.

ClaimStatusEvidenceCaveat

Status meanings:

  • Supported exactly: the source says this directly, with matching wording or a direct equivalent.
  • Supported in substance: the source backs the core point, but the exact wording is broader or sharper than the source.
  • Partly supported: part of the claim is source-backed, but part of it is not.
  • Contested: sources disagree.
  • Overstated: the claim reaches beyond what the evidence can support.
  • Unsupported: no source support found.
  • Refuted: available evidence cuts against the claim.
  • Unverified: not checked, or not enough evidence was found.

Do not upgrade "supported in substance" to "supported exactly."

Rhetorical Leaps

  • Show the move from exact wording to broader implication.
  • Keep evidence, interpretation, and persuasion separate.
  • A quote or denial can support a concern without supporting the strongest version of that concern.
  • If a source step is rhetorical rather than evidentiary, label it as such.
  • Do not dismiss the source; explain the leap.

Fact Ledger

Source Scope

  • Source(s) reviewed:
  • Source type:
  • Time period covered:
  • Known limitations:
  • Missing context:

Atomic Facts

Use a table:

IDFactSource anchorConfidenceNotes

Rules:

  • Facts must be atomic.
  • Do not combine multiple claims into one fact.
  • Use direct source anchors when available: section, paragraph, timestamp, speaker, file name, page, or quoted phrase.
  • Confidence values: High / Medium / Low.
  • If source anchors are unavailable, say so.

Attributed Opinions / Positions

Use a table:

IDWho/SourcePositionBasisConfidence

Rules:

  • Opinions must be attributed.
  • Do not convert opinions into facts.
  • Include stakeholder positions, preferences, concerns, and proposed direction.

Unsupported or Under-Supported Claims

Use a table:

IDClaimWhy unsupportedEvidence needed

Rules:

  • Include claims that sound important but lack source support.
  • Include obvious-sounding assumptions if they are not actually supported by the provided source.

Inferences

Use a table:

IDInferenceBased onRisk if wrong

Rules:

  • Inferences are allowed, but must be explicitly labeled.
  • Link each inference to facts or source material.
  • Do not present inference as fact.

Contradictions / Tensions

Use a table:

IDConflictSources involvedNeeds resolution

Rules:

  • Include hard contradictions and softer tensions.
  • Include cases where one section implies readiness but another suggests a blocker.

Open Questions

Use a table:

IDQuestionWhy it mattersLikely owner

Rules:

  • Focus on questions that block action, decision, reply, or confidence.

Action-Relevant Implications

Use bullets.

Rules:

  • These are implications, not decisions.
  • Do not recommend a final decision unless the user explicitly asks.
  • End this section with: "Do not treat these implications as final decisions."

Suggested Next Skill

Recommend one downstream skill:

  • clear-ask if the next problem is clarifying what someone wants.
  • decision-brief if the next problem is choosing between options.
  • status-update if the next problem is communicating what changed.
  • follow-up-draft if the next problem is replying safely.
  • meeting-to-actions if the next problem is extracting owners/actions.
  • brief-me if the user needs a fast catch-up.

If a named downstream skill is unavailable in the host assistant, describe the next step in plain language instead of pretending the skill can be invoked.

Success Standard

The user should be able to trust the fact ledger as a clean substrate for the next step. In source-expanded mode, the user should also be able to see which claims got stronger, weaker, stayed contested, or remained unverified after outside evidence was checked.

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.