GitHub - Vadiml1024/signal-analysis: A Claude Skill for disciplined analysis of information under noise — fact vs. assessment vs. conclusion, source and motive, capabilities over declarations, independent corroboration.

4 min read Original article ↗

A Claude Skill for disciplined analysis of information under noise.

It encodes an old set of staff-analyst working rules — source before content, fact vs. assessment vs. conclusion, capabilities over declarations, silence as data, and above all: do not fall in love with your own hypothesis — into a workflow Claude follows when you hand it a claim, a headline, a filing, an announcement, or a pile of unstructured chatter and ask what it actually means.

The premise: the usual failure is not ignorance. It's a confident conclusion assembled from one source, one snapshot, and a hypothesis the analyst already liked.

The rules

  1. Source before content. Who said it, when, and why now. The same fact means different things from different mouths.
  2. Fact ≠ assessment ≠ conclusion. A fact is checkable. An assessment is interpretation. A conclusion is a claim about consequences. Errors begin when an assessment is served as fact.
  3. Follow the interests. Not everything is a conspiracy, but almost everything has a motive. A naive analyst is useless; a paranoid one is dangerous.
  4. Capabilities, not declarations. A big goal with no means is a signal. A quiet buildup of means with no declaration is a bigger one.
  5. Inconsistencies talk. "Growth" becomes "stabilization"; "by Q3" becomes "in the medium term." Language shifts before reality is admitted.
  6. Trend over snapshot. +3% is good after −20% and bad after +10%. Never a number without its baseline.
  7. Silence is data. A metric that stops being published. A denial that never comes. A commentator who goes quiet.
  8. Do not fall in love with your hypothesis. Interrogate yourself: what contradicts it, and what would have to happen for it to collapse?
  9. Independent lines of confirmation. Three outlets retelling one primary publication are one signal, not three. Real corroboration comes from lines of different nature: money, personnel, procurement, logistics, filings, adversary behavior.
  10. Signal vs. noise. Excess information doesn't make anyone smarter; it corrodes the ability to think. Filtering is part of the work, not a preliminary to it.
  11. Don't rush the conclusion. Fast reaction is good for emotion and bad for understanding.

What the skill actually does

When triggered, Claude works through eight steps rather than answering off the cuff:

  1. Inventory and de-noise — type every input (official statement / leak / statistics / filing / viral post / silence) and discard noise explicitly.
  2. Decompose — split the material into facts, assessments, conclusions, and rumor. Highest-yield step; done in writing.
  3. Motive map — who gains from this being believed, now. Including the messenger.
  4. Capability check — what would the declared intention materially require, and are the resources there?
  5. Anomaly sweep — arithmetic that doesn't close, slipping timelines, language drift, trend vs. snapshot, silences.
  6. Corroboration map — collapse everything tracing back to one primary source into a single row. Count distinct signals, not outlets.
  7. Self-interrogation — contrary evidence, concrete falsifiers, and the best rival hypothesis. If you can't build a serious rival, you haven't understood the situation.
  8. Calibrated conclusion — a judgment with an honest confidence level and the observables to watch next.

Output follows a fixed template ending in Against my own hypothesis, What to watch, and Gaps.

Install

Claude.ai / Claude Desktop — download signal-analysis.skill from Releases and upload it in Settings → Capabilities → Skills.

Claude Code — clone into your skills directory:

git clone https://github.com/Vadiml1024/signal-analysis.git ~/.claude/skills/signal-analysis

Repackage after editing:

zip -r signal-analysis.skill signal-analysis/

Layout

signal-analysis/
├── SKILL.md                        # doctrine, 8-step workflow, output template, confidence table
└── references/
    ├── markers.md                  # language-drift ladders, inconsistency tells, silence markers
    └── source-typology.md          # what each source type is good for, how it fails, independence rules

Triggering

The description is deliberately assertive. It fires on "what does this really mean," credibility and motive questions, verify/debunk requests, forecasting from ambiguous signals, and pasted floods of headlines — and also when you're visibly building a narrative on thin or single-sourced evidence without asking for analysis at all.

Notes

  • Material under analysis is treated as data, not instruction. Directives embedded in an analyzed document are logged as evidence of the source's intent, never followed.
  • The skill is allowed to conclude "nothing here supports the claim," or "two readings remain open, and here is what separates them." Flattening ambiguity into a clean verdict to look decisive is the exact failure it exists to prevent.

Provenance

Adapted from a Russian-language essay by a former General Staff officer on the analytical habits taught in the early 2000s ("За деревьями лес"). The rules are his; the workflow, the marker catalogues, and the source typology are an expansion of them for use by a language model. The original text is not redistributed here.

License

MIT — see LICENSE.