predict-future
Turn a user query into a forecast by thinking through which theories fit, then applying one or more prediction theories from references/. Present each theory's forecast separately — do not merge them into one authoritative conclusion. The value is rigor: fight overconfidence and ignoring base rates.
- Output language: respond in the same language the user used in their query.
- Theories live in references/ — think before you select; read only the files you chose; follow each file's "How to apply in this skill" section.
Workflow
Step 0: Set up the run folder
Generate a timestamp like YYYYMMDD-HHmmss and create output/{timestamp}/
(relative to the workspace root).
Step 1: Restate & clarify
State clearly what is being predicted and its current state. If critical information is missing, ask 1–3 of the most decisive questions before forecasting. For non-critical gaps, proceed using explicit labeled assumptions.
Step 2: Think through theory selection
Before reading or applying any reference, work through theory selection deliberately. Do not default to any file. Pick the minimum sufficient set that addresses the query's decisive uncertainty; add a second or third theory only when it materially improves the forecast.
Reference library
| File | Thinker | Use when the query involves… |
|---|---|---|
| general_Philip_Tetlock.md | Philip Tetlock | Open-ended probabilistic forecasts, triage, Fermi decomposition, dragonfly eye, granular probabilities, Bayesian updating, outside view |
| finance_howard_marks.md | Howard Marks | Consensus vs. price, cycles, second-level thinking, credit/macro sentiment |
| finance_robert_shiller.md | Robert Shiller | Valuation (CAPE), narratives, bubbles, sentiment contagion |
| finance_hyman_minsky.md | Hyman Minsky | Credit fragility, financing regimes, systemic breakdown risk |
| tech_paul_saffo.md | Paul Saffo | Uncertainty mapping, S-curves, weak signals, when not to forecast |
| tech_geoffrey_moore.md | Geoffrey Moore | Product adoption stage, chasm, beachhead, whole product, GTM |
| tech_clayton_christensen.md | Clayton Christensen | Disruption vs. sustaining, industry shift, incumbent response |
| intl_structural_realism.md | Waltz / Mearsheimer | Great-power behavior, balance of power, balancing vs. buck-passing, long-run geopolitical trajectory |
| intl_bruce_bueno_de_mesquita.md | Bruce Bueno de Mesquita | Specific policy/negotiation/power-struggle outcomes via stakeholder bargaining; regime behavior (selectorate) |
| intl_thomas_schelling.md | Thomas Schelling | Crisis bargaining, deterrence vs. compellence, credibility, escalation, brinkmanship |
When to include each theory (examples, not defaults)
| Include… | When the decisive uncertainty is… |
|---|---|
general_Philip_Tetlock.md |
You need triage, reference-class/base-rate reasoning, decomposition, multi-angle synthesis, or calibrated probabilities — typical for open-ended "will X happen?" questions |
finance_* |
Price, narrative, consensus, cycles, or credit structure drive the answer |
tech_* |
Adoption stage, uncertainty mapping, disruption, or incumbent dynamics drive the answer |
intl_* |
International/state behavior, geopolitics, conflict, deterrence, or policy bargaining drive the answer |
| Domain signal | Typical starting candidates (still justify each) |
|---|---|
| Equities, bonds, credit, rates, macro, housing, sentiment | finance_howard_marks.md, finance_robert_shiller.md; finance_hyman_minsky.md if leverage/credit stress matters |
| Product launch, SaaS, platforms, startups, adoption, "will X disrupt Y?" | tech_geoffrey_moore.md, tech_clayton_christensen.md, tech_paul_saffo.md — pick by which uncertainty dominates |
| Great-power rivalry, war/crisis, deterrence, sanctions, treaties, foreign policy | intl_structural_realism.md for long-run structure; intl_bruce_bueno_de_mesquita.md for specific negotiated outcomes; intl_thomas_schelling.md for escalation/deterrence — pick by which uncertainty dominates |
| Elections, geopolitics, sports, one-off events | Often general_Philip_Tetlock.md alone — but only if probabilistic forecasting is appropriate; add intl_* when state-level strategy dominates |
Write output/{timestamp}/theories-selected.md before reading any reference file:
- Per-file fit assessment (all ten files)
- Final list of selected references with rationale
- Explicit list of excluded references with rationale
Only then proceed to evidence gathering and application.
Step 3: Gather evidence
This step requires live research — do not rely on memory alone. Call at least one tool before applying theories
| Tool | Use for |
|---|---|
| Search | Discover sources, base rates, recent news, and comparable cases |
| Web extraction | Full text of a specific URL from search results |
| Control browser | Pages that need interaction or JS rendering (e.g. X/Twitter, Reddit, TikTok, Xiaohongshu)，please close web tabs after you use this tool |
Save every fetched/derived artifact as an intermediate file in the run folder. Do not fabricate facts; if something can't be verified, label it an assumption. Examples:
output/{timestamp}/research-base-rates.mdoutput/{timestamp}/research-scheduled-events.mdoutput/{timestamp}/sources.md(URLs + key extracts)
Step 4: Apply selected theories
Apply each selected reference in turn. For every theory, produce a complete, standalone forecast — reasoning plus a clear prediction conclusion (outcome, probability or qualitative call, horizon, falsifiable triggers). Do not blend theories into one answer while applying them; disagreements between theories are a feature, not a bug.
Step 5: Final output file
Write output/{timestamp}/output.md with this structure (translate headings into
the user's language). One section per applied theory, each ending with that
theory's own prediction. Omit sections for theories you did not select. No
merged "final verdict" — if theories disagree, present each view and optionally
add a short cross-theory comparison noting agreements and tensions without picking a winner.
# Forecast: [subject]
## Original input
[User's query verbatim]
## Theory selection (reasoning)
- Decisive uncertainty: …
- Selected: [filename] — rationale
- Not selected: [filename] — rationale
## Selected forecasting theories
- [filename] — selection rationale
- …
## Evidence gathered
- [Key fact] — source / or labeled as assumption
## [Theory one, e.g. Tetlock]
[Reasoning per that reference's apply steps]
### [Theory one] — prediction
- Forecast: …
- Probability / qualitative call: …
- Confidence: …
- Falsifiable conditions: …
## [Theory two, e.g. Howard Marks]
[Reasoning per that reference's apply steps]
### [Theory two] — prediction
- Forecast: …
- …
## [Theory three, e.g. Geoffrey Moore]
…
## Cross-theory comparison (optional)
- Points of agreement: …
- Points of disagreement: …
(Compare only — do not merge into a single conclusion)
Tell the user the path to output/{timestamp}/output.md when done.