Doubt

Doubt

Produce a decision artifact that shows why a conclusion holds, what pushes against it, and what remains unknown. Do not use a graph to decorate an answer that has not been sourced.

Workflow

1. Frame one decision. Write a falsifiable question and one provisional

position. Narrow broad topics until a reader can tell what action or belief the map is testing.

1. Collect current source regions. Prefer direct observations and primary

sources. For each item, preserve the URL or local path, publisher, date, retrieval date, section/page/line locator, and a short verbatim excerpt copied from that source region. Never put an agent-written paraphrase in excerpt; keep interpretation in the evidence node's text. Read [references/evidence-ladder.md](references/evidence-ladder.md) when source quality is disputed.

1. Atomize the reasoning. Create only four node types:

• position: the single current verdict;
• claim: an intermediate proposition;
• evidence: a sourced observation;
• unknown: a specific missing fact that could change the result.

1. Type every edge. Use supports, contradicts, qualifies, or

missing. Add a plain-language note explaining why the source node bears on the target. Similarity is not support; opposite wording is not necessarily contradiction.

1. Preserve disagreement. Do not delete contrary evidence merely because

the verdict survives it. Represent scope differences as qualifies, not false contradictions.

1. Refuse false precision. Do not invent confidence percentages. Express

uncertainty as an explicit unknown, a qualifying edge, or a narrower verdict.

1. Write the map JSON. Follow

[references/map-schema.md](references/map-schema.md). Keep node IDs short, stable, and semantic.

1. Fail closed offline, explicitly verify, then render. The host exposes

the path of this SKILL.md when it loads the skill. Resolve scripts/validate.mjs relative to that directory and replace `` below with the concrete path; do not pass the placeholder literally.

``bash node /scripts/validate.mjs decision.doubt.json # Required before calling an externally sourced map finished or verified; # run only after the user explicitly permits source retrieval: npx doubt-ai verify decision.doubt.json --out decision.verified.doubt.json npx doubt-ai map decision.verified.doubt.json --out decision.html ``

The bundled validator uses only Node.js 18+ built-ins and must work without package installation or network access. Fix every validation finding. Never bypass missing-source, unused-evidence, unsupported-position, or false-precision findings. Only report validation success and a receipt after the validator exits zero and prints VALID followed by a 64-character receipt (or JSON with valid: true and that receipt). A file hash, node count, or syntactic JSON check is not a Doubt receipt. If the validator cannot run, report the run as blocked; do not claim success. Source verification is a required final gate for a finished externally sourced map, but do not run verify implicitly: it can make outbound requests for recorded HTTP sources. Ask for permission first. If permission is denied or unavailable, you may render the original JSON only as a recorded-only draft; do not call it finished or verified. Local file verification does not use the network. If verification fails, preserve the mismatch or retrieval failure instead of writing an attestation.

1. Inspect the HTML. Verify that the question, verdict, evidence cards,

contradictions, unknowns, edge focus, filters, and source links are readable. The JSON remains the canonical editable artifact.

Quality gates

A finished map must satisfy all of these:

• exactly one position has incoming reasoning;
• every evidence node names a source and participates in an edge;
• every source has a publication date, retrieval date, bounded region, and is used;
• every source excerpt is verbatim text from its bounded region, never a summary;
• every non-position node has a directed path to the position, without cycles;
• the map contains contrary or qualifying evidence when the source set contains

it;

• unresolved decision-changing facts appear as unknown nodes;
• every edge note states an entailment, contradiction, qualification, or gap;
• the verdict is no broader than the evidence.

Do not

• Treat an agent-written summary as evidence for itself.
• Attach a citation that merely mentions the topic.
• Generate decorative orphan nodes.
• Infer that two sources conflict when populations, versions, dates, or

evaluation conditions differ.

• Hide a failed retrieval behind authoritative prose.
• Replace a simple, settled fact with an unnecessary map.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: alsoleg89
• Source: alsoleg89/doubt
• License: MIT
• Homepage: https://alsoleg89.github.io/doubt/

Install and usage instructions live in the source repository linked above.