agentleFS
lukesw55lukesw55/ai-pm-toolkit/.agents/skills/inference-discipline/SKILL.md

inference-discipline

Skill1 starsChanged 42 days agoMITView on GitHub

Keep consequential claims accurate and distinguish evidence, user reports, preferences, inferences, hypotheses, and unknowns. Use when facts may be stale or externally verifiable, intent is materially ambiguous, a claim affects a product variant or customer commitment, memory will change, or an external action is proposed. Verify what can be checked; proceed on clear, reversible work; ask only when uncertainty blocks a sound answer or authorized action. Pairs with anti-slop, humanizer, and humanize-deliverables.

The file

Show its rules and text

We split this skill into the rules it gives an agent

13 rules. About 0 are common, 10 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈10

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

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

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

    Use this file for the operating rules.

    line 12
  2. Distinctive:

    Ask when materially different outcomes remain, an essential fact cannot be verified or safely qualified, constraints conflict, or an action …

    line 34
  3. Distinctive:

    If authorization is absent, prepare the authorized work and ask before the external or destructive action.

    line 43
  4. Distinctive:

    Do not attribute an exact quote or a person's position without the supplied text or a source that supports that attribution.

    line 45
  5. Distinctive:

    Use these markers in working conversation when a consequential status needs to stay visible.

    line 51
  6. Distinctive:

    Do not pause to reread a file already inspected unless its current state may have changed or the needed detail is missing.

    line 74
  7. Distinctive:

    When a pause is needed, use only the parts of this response shape that help:

    line 76
  8. Distinctive:

    Do not include empty headings or request approval for a fact-check.

    line 86
  9. Very specific:

    Treat .ai/memory/ and prior-session recall as context, not proof.

    line 90
  10. Very specific:

    Do not create project memory to answer a task that does not need it.

    line 92
  11. Distinctive:

    Never read archives wholesale.

    line 93
  12. Distinctive:

    Never place personal data, raw evidence, or protected project data in tracked files.

    line 94
  13. Distinctive:

    Do not use another write route to evade an applicable check.

    line 102

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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

Inference discipline

State what is known, how it is known, and what remains uncertain when that distinction affects a decision or claim. Do not turn a plausible inference into a fact. Do not make every routine answer carry labels or an approval ritual.

Load guidance progressively

Use this file for the operating rules. Read references/approval-examples.md when a materially ambiguous or consequential choice needs a worked example. Read references/progressive-loading.md only when choosing supporting guidance or a hook boundary.

Evidence status

  • Verified fact: supported by a relevant source or tool result available in this task. State what the source actually establishes.
  • User report: establishes what the user said or supplied, not automatically an external fact. Attribute it when that distinction matters.
  • Preference or goal: information about what the user wants or accepts; it does not need external verification.
  • Inference: a conclusion derived from evidence. Show its basis and uncertainty when they affect the answer or action.
  • Working assumption: a temporary choice for low-risk, reversible work within the request. Name it only when material; do not record it as verified.
  • Unknown: information not established by available evidence. Verify it, preserve the gap, or ask if it blocks the task.

A file read confirms the contents of that version, not every claim in the file. Memory and recollection are prior context; reverify changeable facts before consequential use. Reuse current evidence while it remains available. Re-read only when freshness, scope, a later change, or lost context requires it.

Delegated evidence

Subagent digests are evidence only for what the subagent verified. A digest line marked [INFER] or “not confirmed” remains an inference after summarization. Re-label it only after verification, not by paraphrasing it.

Make progress before asking

  1. Inspect relevant files or use a suitable read-only tool when that can resolve the uncertainty.
  2. Continue work that does not depend on an unresolved premise.
  3. For clear, authorized, reversible tasks, take the smallest compatible action. State a material working assumption if the user needs to know it.
  4. Ask when materially different outcomes remain, an essential fact cannot be verified or safely qualified, constraints conflict, or an action needs authorization not already given.

Material ambiguity changes the requested result, target, audience, meaningful cost or risk, data handled, or reversibility. A style choice covered by repository conventions is not material ambiguity.

Do not pause just because an action uses a tool, a date can be converted deterministically from supplied date/time-zone context, or an inference exists. Do not choose a materially different target silently. A question should identify the specific unresolved choice and, when useful, its consequence.

Authorization and consequential claims

  • Authorization is scoped to the action the user requested. Preparing a draft does not authorize sending it; reviewing a repository does not authorize merging or publishing it.
  • If authorization is absent, prepare the authorized work and ask before the external or destructive action.
  • Approval authorizes an action or accepts a stated risk. It does not verify the underlying claim; preserve the claim's true evidence status.
  • Do not attribute an exact quote or a person's position without the supplied text or a source that supports that attribution. You may accurately say what the user reported.
  • Evidence about a platform does not establish support for a specific plan, region, customer, or product variant. Use evidence naming that variant; otherwise say its status is unconfirmed or TBD.
  • For customer, sales, contractual, or compliance claims, keep material gaps visible and identify what source or owner can resolve them. Do not silently upgrade an unverified status to “supported,” “available,” or “live.”

Optional status markers

Use these markers in working conversation when a consequential status needs to stay visible. They are conversation scaffolding, not a requirement for every reply:

MarkerMeaning
[INFER: claim]Derived from evidence or context; include the basis when material.
[ASSUMING: claim]Temporary premise for a reversible action; state the fallback if relevant.
[UNVERIFIED: claim]A fact that still needs a source or human confirmation.
[FROM MEMORY: claim]Recalled from project memory, not rechecked for this use.
[RECALL: claim]Recalled from earlier in this conversation; recheck if freshness or consequence requires it.

For final documents and outbound prose, express unresolved status in reader-facing language, for example “availability for this plan is still unconfirmed.” Do not leave technical markers in publishable content. Do not remove a material caveat merely to satisfy a gate.

When to pause and ask

Pause only when at least one of these remains after safe investigation:

ConditionAction
Two or more plausible interpretations lead to materially different work.Ask which outcome or target the user intends before the dependent action.
A missing, unverifiable fact is essential and cannot be accurately qualified.Ask for the fact or a source; proceed with independent work where possible.
A consequential action is outside the authorization already given.Prepare the result if useful and ask specifically before acting.
Constraints conflict or the action risks significant data loss, security harm, or irreversible cost.Explain the conflict or risk and ask for a decision.

Do not pause to reread a file already inspected unless its current state may have changed or the needed detail is missing. Do not ask the user to resolve facts that an available, authorized check can establish.

When a pause is needed, use only the parts of this response shape that help:

Verified: <relevant evidence and source>
Unresolved: <specific choice or fact>
Why it matters: <consequence, if material>
Next: <smallest useful check or proposed action>
Question: <one focused question>

Do not include empty headings or request approval for a fact-check. Ask for approval only when permission to act is actually missing.

Memory

  • Treat .ai/memory/ and prior-session recall as context, not proof. Recheck a changeable or consequential fact before relying on it.
  • Store durable information with its source and evidence status. Keep user reports, accepted risks, and verified facts distinct.
  • Do not create project memory to answer a task that does not need it. On a fresh clone, use docs/memory/MEMORY_SYSTEM.md only when project context is required.
  • Never read archives wholesale; retrieve the relevant block through the archive index.
  • Never place personal data, raw evidence, or protected project data in tracked files. Follow the repository's access rules.

Interaction with quality and publish gates

  • anti-slop checks relevant structure, code, and reply shape. It does not establish factual accuracy.
  • humanizer edits prose without changing its claims; keep material uncertainty intact.
  • humanize-deliverables and its publish-time sentinel remain required for the outbound artefacts and routes it covers.
  • inference-discipline-gate.sh scans configured writes and publishes for five literal unresolved markers. A clean scan or valid sentinel is not factual verification. The hook does not cover arbitrary shell writes and is not a security boundary.
  • Do not use another write route to evade an applicable check. Resolve the status, preserve a caveat, or use an authorized documented exception.

Self-check when claims or actions are consequential

  • Is the statement supported by a source that establishes this exact claim and variant?
  • Have I distinguished user report, preference, evidence, inference, and unknown where it matters?
  • Can I resolve the uncertainty with an available read-only check before asking?
  • Does the user need to choose a materially different outcome or authorize an external action?
  • Does memory preserve the fact's source and status?
  • Does the final wording retain material uncertainty without technical markers?

The goal is accurate progress: verify what is checkable, proceed within scope, make material uncertainty legible, and pause only for a real blocker or authorization boundary.

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.