Deep Decide

Deep-Decide

Use this when someone needs a better decision, not just more information. The workflow is: independent perspectives, then forced dissent, then one synthesized verdict. It is portable and self-contained: a bundled Python script (scripts/deep_decide.py, standard library only) makes the model calls using keys you supply in your own environment.

What it does NOT do

• It is not a research tool. For gathering facts, market data, or sources, run a

deep-research skill first and paste the findings in as context. Deep-decide reasons over what it is given; it does not browse.

• It does not execute the decision. It recommends; it never sends, deploys,

buys, or commits anything.

• Consensus is not proof. The verdict preserves dissent and TBU assumptions on

purpose; it does not launder uncertainty into false confidence.

Bring your own keys

The script reads credentials from the environment. Set at least one; set more than one to get genuine cross-model dissent (perspectives are rotated across the providers you configured):

Each default is that provider's current flagship, and the calls run at high reasoning effort with adaptive thinking on where the provider supports it. Override any model with its env var, or the depth with DEEP_DECIDE_EFFORT (low | medium | high | xhigh | max, default high).

| Env var | Provider | Model env override | Default model | |---|---|---|---| | ANTHROPIC_API_KEY | Claude | DEEP_DECIDE_ANTHROPIC_MODEL | claude-opus-4-8 | | OPENAI_API_KEY | OpenAI | DEEP_DECIDE_OPENAI_MODEL | gpt-5 | | GEMINI_API_KEY | Gemini | DEEP_DECIDE_GEMINI_MODEL | gemini-2.5-pro | | OPENROUTER_API_KEY | OpenRouter (any model) | DEEP_DECIDE_OPENROUTER_MODEL | openai/gpt-5 | | CURSOR_API_KEY (+ CURSOR_API_BASE) | Cursor | DEEP_DECIDE_CURSOR_MODEL | gpt-5 |

A subscription (ChatGPT, Claude.ai, etc.) is not an API key. These calls bill your API account. Check what is configured before running:

python3 scripts/deep_decide.py --providers

User-facing contract

When someone says deep-decide or a nearby trigger, treat the skill as the interface. Do not lead with a raw terminal command.

1. Restate the decision, options, and criteria if they are clear.
2. Run the no-cost plan and show the provider/perspective/call-count breakdown.
3. State the paid-call count, then run live only after the person approves paid

calls in the current turn (or explicitly asked for a live run).

1. Return the verdict in plain language: recommendation, the dissent that

matters, assumptions to verify, and next actions.

If options or criteria are missing but can be reasonably inferred, infer them and mark them as inferred. Ask a question only when the decision would be materially distorted without the missing information.

Running it

No-cost plan (default: prints which providers it found and what it would do, and makes no API calls):

python3 scripts/deep_decide.py "Should we do X or Y?" \
  --option "Do X" --option "Do Y"

Live run (paid):

python3 scripts/deep_decide.py "Should we do X or Y?" \
  --option "Do X" --option "Do Y" \
  --criterion "12-month cash impact" \
  --criterion "team adoption" \
  --execute

Run with no decision argument to be prompted for it interactively. Override the four default perspectives with --perspective "label: what this lens focuses on" (repeatable). Output bundles are written under ./deep-decide-runs// (verdict.md, combined.md, manifest.json, one file per perspective); override the location with --out-dir.

Status tokens (first stdout word; exit 0 only on OK_*): OK_PLAN, OK_DECIDE, OK_PROVIDERS, ERR_NO_PROVIDERS, ERR_NO_DECISION, ERR_RUN.

Default shape

The default bundle is five paid calls: four perspective passes plus one final synthesis. The default perspectives are:

• operator: execution reality, dependencies, operational drag, what breaks.
• skeptic: red-team critique, hidden assumptions, failure modes.
• finance: ROI, cash/time cost, opportunity cost, operating leverage.
• stakeholder: adoption, incentives, trust, communication burden.

With several keys set, those four are spread across providers (e.g. operator on Claude, skeptic on OpenAI) so the dissent is cross-model, not one model arguing with itself. The synthesis "chair" defaults to Claude when available.

The research arsenal (built in)

This skill ships its own evidence layer, scripts/research_stack.py: a catalog of 17 research providers across three tiers, with live bring-your-own-keys adapters for the search-capable ones. See the whole thing and what is configured:

python3 scripts/research_stack.py --stack    # or: python3 scripts/deep_decide.py --stack

Tiers:

• Deep web research: Exa, Firecrawl, Parallel.ai, Gemini Deep Research, Browserbase.
• Scholarly / academic: arXiv, OpenAlex, Crossref, PubMed, Semantic Scholar, CORE,

Google Scholar (via SearchAPI/SerpAPI). arXiv, OpenAlex, Crossref, and PubMed work with no key at all.

• GitHub / code: GitHub code search, GitHub repo search, grep.app, Exa GitHub

scan, Sourcegraph. grep.app needs no key.

Gather live, cited evidence on its own:

python3 scripts/research_stack.py "your question" -n 5 --tier academic

Or seed it straight into a decision so the perspectives reason over real sources instead of guesses:

python3 scripts/deep_decide.py "Should we adopt approach X?" \
  --option "X" --option "status quo" \
  --research "evidence for and against approach X" --execute

Each finding is tagged with its source and URL and is injected as "treat as leads to verify, not proven truth." Providers marked catalog in --stack (Parallel, Gemini Deep Research, Browserbase, CORE, Google Scholar, Sourcegraph) are part of the stack and shown with their env vars, but use async/job or key-gated APIs out of scope for the synchronous evidence pass; the live providers are queried directly. Always verify load-bearing claims against current primary sources before acting on the verdict.

Gotchas

• Pure standard library, no pip install. If `python3 scripts/deep_decide.py

--providers` runs, you are ready.

• --execute is the only thing that spends money. Everything else is free and

offline.

• One key is fine: the run becomes single-provider, multi-perspective. The

cross-model dissent advantage only appears with two or more keys.

• A perspective that fails (bad key, rate limit) is recorded in manifest.json

and the run continues with the rest; the run only fails if every perspective fails.

• Do not use this as a publishing or outbound-action gate. It produces a

recommendation, nothing else.

Source & license

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

• Author: bogdanbaciu21
• Source: bogdanbaciu21/skills
• License: MIT
• Homepage: https://www.bogdanbaciu.com

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