Build Protocol · Decision
Apply rigorous process to decisions that cost money, time, or opportunity—before committing, not after.
Inherits the 8 Iron Rules from
build-protocol. Adds 3 decision-specific rules and a 10-step workflow tuned for high-stakes choices.
When to Use
Triggers (any one):
- Decision commits >$1,000 or >1 day of effort
- Decision is hard to reverse (stocks, hardware, vendor contracts, hiring)
- Multiple alternatives with real tradeoffs exist
- Investment, trading, or portfolio questions
- Technology selection, supplier choice, buy-vs-build
- Explicit: "should I buy X", "compare options", "pick a vendor", "选哪个", "怎么选"
Don't use for:
- Reversible, low-cost choices (which coffee to order)
- Pure knowledge production (use
build-protocolinstead) - Decisions already made—this is pre-decision, not post-rationalization
Inherits from build-protocol (8 Iron Rules)
| # | Rule | Decision context |
|---|---|---|
| 1 | Independent Audit unmissable | Audit = verify methodology + data freshness, not just content |
| 2 | Plan before Execute | Map options before picking; list constraints before scoring |
| 3 | ≤2 parallel sub-agents for shared-state research (writing/editing same files); ≤4 acceptable for independent investigations | Inherits build-protocol concurrency reasoning; see trinity-harness for general-purpose limits |
| 4 | Why This Way | Every recommendation needs methodology, not just "I think" |
| 5 | Version + Errata iteration | Revisit decision log when facts change |
| 6 | 3-layer consistency | Data source ↔ analysis ↔ recommendation must align |
| 7 | Anti-Sycophancy content | Must list downsides and failure modes; all-upside = fake |
| 8 | Independent review ≠ self-audit | Recommendation author ≠ final reviewer |
Decision-specific 3 New Iron Rules
| # | Rule | Why |
|---|---|---|
| 9 | Real-time data verification — never trust document prices | Document prices are stale within 24h. A written "$X" that hasn't been verified live is fiction. |
| 10 | Methodology transparency (PE/DCF/MA/decision matrix required — no gut calls) | "I think this is a good buy" is not analysis. Every recommendation must show its math. |
| 11 | Stop-loss first, profit-target second | Asymmetric downside means exits matter more than entries. Define maximum loss before opening any position. |
The Decision Workflow (10 Steps)
Pre-decision (Steps 1–4)
Step 1 — Research: Gather options
- List all candidates (don't anchor on first option)
- Source: web search, product reviews, market data, expert opinions
- Output: options list with brief description of each
Step 2 — Set constraints
- Budget ceiling (hard limit)
- Timeline (when must decision be made / when does it take effect)
- Hard-no's (non-negotiable disqualifiers)
- Risk tolerance (for investments: max drawdown you can stomach)
Step 3 — Fetch real-time data ⚠️ non-skippable
- Stocks:
curl "https://stooq.com/q/l/?s=SYMBOL.us&f=sd2t2ohlcv&h&e=csv" - Crypto: CoinGecko API or Binance spot price
- Products: scrape current listing price (Amazon/retailer)
- Services: current vendor pricing page (not cached / not PDF)
- If real-time data is unavailable → state explicitly, do not substitute document prices
Step 4 — Compare on dimensions (build comparison matrix)
- Define 4–8 scoring dimensions relevant to this decision
- Score each option per dimension (1–5 or weighted %)
- Make tradeoffs explicit; no option should be "best on all dimensions"
Decision (Steps 5–7)
Step 5 — Apply methodology
- Investments: PE ratio (vs sector median), DCF estimate, technical levels (MA50/MA200), catalyst timeline
- Purchases: TCO (total cost of ownership over 3 years), not just sticker price
- Technology/vendor: Gartner-style magic quadrant or weighted decision matrix
- Show your work. If you used PE, show the number. If you used DCF, show the inputs.
Step 6 — Plan B and exit conditions
- Investments: hard stop-loss level, reason to exit (not just price but thesis invalidation trigger)
- Purchases: return policy window, resale value estimate
- Tech/vendor: contract termination terms, migration cost estimate
- Define these before committing, not after things go wrong
Step 7 — Make decision + document rationale
- State the chosen option clearly
- One-paragraph "Why This" (not why the others are bad—why this one fits the constraints)
- One-paragraph "Why Not the Alternatives" (must be honest, not dismissive)
- Date-stamp the decision (facts change; you need to know how old this reasoning is)
Post-decision (Steps 8–10)
Step 8 — Execute with position sizing
- Investments: size = (account risk per trade) ÷ (entry − stop-loss). Never max-leverage on first entry.
- Purchases: confirm budget compliance before checkout
- Tech/vendor: phased rollout if possible; don't migrate 100% on day one
Step 9 — Track P/L daily (not weekly, not monthly)
- Record entry price, current price, unrealized P/L in a daily log
- For non-investment decisions: track KPIs that validate the decision (e.g., tool adoption rate, cost savings vs projection)
- Daily tracking catches thesis-breaking events early; weekly/monthly is too slow
Step 10 — Review and adjust (kill bad bets fast)
- Weekly: is thesis still intact? Did the stop-loss get hit?
- Monthly: compare actual outcome vs pre-decision projection
- If thesis is broken: exit at stop-loss, do not average down hoping it recovers
- Log lessons in a decision postmortem (even for winners—understand why it worked)
The "Sycophancy of Precision" Warning ⚠️
The problem: Writing +12.85% feels more credible than "about +13%". But precision ≠ accuracy. When the underlying data is an estimate or an outdated figure, adding decimal places is misleading—it manufactures false confidence.
How it shows up:
- AI stitches together multi-source data and presents blended numbers as if they're exact
- Historical document prices get reused as if they're current
- Compound calculations give precise outputs from imprecise inputs
The rule:
- Any number derived from estimates → append "(est.)" or "(含估算)"
- Any number from a source >24h old → append the source date
- Round to 2 significant figures when inputs are uncertain; don't carry false precision
- Example:
+12.85% (est., based on 3-day average)not just+12.85%
Real-time Data Sources Required
| Asset class | Source | Command / URL |
|---|---|---|
| US stocks | Stooq | curl "https://stooq.com/q/l/?s=SYMBOL.us&f=sd2t2ohlcv&h&e=csv" |
| US stocks (alt) | Yahoo Finance | curl "https://query1.finance.yahoo.com/v8/finance/chart/SYMBOL" |
| Crypto | CoinGecko | curl "https://api.coingecko.com/api/v3/simple/price?ids=bitcoin&vs_currencies=usd" |
| Products | Retailer page | Direct web fetch; no cached PDFs |
| Forex | Stooq | curl "https://stooq.com/q/l/?s=USDJPY&f=sd2t2ohlcv&h&e=csv" |
🔴 Never use: Prices from documents, spreadsheets, or chat history. Even a price from yesterday's analysis is stale.
Anti-patterns (Decision-specific)
| ❌ Anti-pattern | Why it fails |
|---|---|
| "I think it should work" without backtesting | Gut calls without data = coin flip with extra steps |
| No stop-loss defined before entry | You will rationalize holding through any loss |
| Doubling down on a loser "to average down" | Throwing good money after bad; thesis may already be broken |
| Fake precision ("+12.85%" without methodology) | Sycophancy of precision; decimal places don't add accuracy |
| Cherry-picking favorable data | Confirmation bias in analysis = surprises at execution |
| Using document prices instead of live data | 24h-old stock price is meaningless for a decision made today |
| Daily P/L tracking skipped (tracking weekly/monthly) | Weekly is too slow to catch a thesis-breaking move |
| "Buy more to feel better" after a loss | Emotions ≠ analysis; size down when uncertain, not up |
| Recommendation with no methodology shown | "It looks good" is not analysis |
| All-upside analysis (no failure modes listed) | Real decisions have tradeoffs; omitting downside = sycophancy |
Gotchas
- Polymarket / prediction market cost basis ≠ current MTM value (look up current contract price before calculating P/L)
- Document prices expire within 24h for liquid assets; within weeks for products
- AI recommendation confidence ≠ data freshness — a confident-sounding AI answer with no live data check is high-risk
- MA/RSI are lagging indicators — they describe past price action, not future movement; use as context, not predictions
- DCF sensitivity is high — small changes to discount rate or terminal growth change the output by 30–50%; always run a range, not a single number
- Position sizing compounds errors — if your entry price is wrong by 5%, your stop-loss distance is wrong too; use real-time price in sizing calculation
References
references/decision-workflow.md— Full 10-step workflow with per-step checklistsreferences/investment-playbook-template.md— Generic investment playbook template (anonymized)references/audit-script-decision.sh— Bash script to verify methodology transparency and data freshness
Related Skills
build-protocol— Parent skill for long-form knowledge production (8 Iron Rules source)academic-research-hub— For decisions requiring deep literature reviewfeishu-bitable— For tracking decision logs and P/L records in Bitable
Distilled: 2026-04-30 · Inherits: build-protocol v1.1 Validated on: Investment decision audits, technology selection, major purchase comparisons