starchild-ai-agent Why this exists
Hindsight bias rewrites memory. "I knew it was risky" feels true after the loss, but you didn't write it down. A decision journal forces you to commit your reasoning to paper before the outcome is known, then scores how well your predictions matched reality — so the user learns from prediction errors, not just outcomes.
Good outcomes from bad reasoning = luck. Bad outcomes from good reasoning = variance. You can only tell the difference if the reasoning was captured beforehand.
When to use
- User says "I'm deciding whether to..." on something consequential
- User commits to a position / pivot / hire / kill decision
- User reviews a past decision ("what did I think when I bought X?")
- User asks "what was my reasoning on..."
- A scheduled review fires (see Review loop below)
Do NOT use for: trivial preferences ("should I use tabs or spaces"), routine execution, or things the user has already decided and is just executing. Only log decisions where the reasoning itself has learning value.
The four fields that matter
Every entry captures these, no exceptions:
- The decision — what you chose, in one sentence. Not the options considered, the choice made.
- The reasoning — WHY, at the time. What you believed to be true. What would have to be true for this to be wrong. This is the whole point — capture it before the outcome.
- Expected outcome — concrete and falsifiable. "Up 20% in 6 months" not "should do well." If you can't state a falsifiable expectation, the decision isn't logged — it's too vague to learn from.
- Review date / trigger — when to come back and score. Date-based ("2026-09-01") or event-based ("next earnings call", "if price hits $X").
Optional but valuable: confidence (1-5), alternatives rejected (one-liner each, with why), what would invalidate this (the kill criterion).
Workflow
Logging a decision
from core.skill_tools import decision_journal_add
# or via bash:
# python3 skills/decision-journal/scripts/journal.py add ...
Fields: title, decision, reasoning, expected, review_on (ISO date or trigger phrase), confidence (1-5), tags (list), alternatives (dict of name→reason rejected), invalidation (string).
Returns the entry ID. Entries are stored in workspace/data/decision-journal.jsonl (append-only, one JSON object per line).
Reviewing — the part that actually teaches
Two modes:
Date-triggered review — entries with review_on ≤ today and status=open are due. Run:
python3 skills/decision-journal/scripts/journal.py due
Returns due entries. For each, the agent prompts the user: "What actually happened?" then scores it.
Manual review — user asks "review my decision on X":
python3 skills/decision-journal/scripts/journal.py find --query "X"
Scoring an outcome
python3 skills/decision-journal/scripts/journal.py score --id <entry_id> \
--actual "what happened" --outcome right|wrong|partial|unresolved \
--notes "what the gap teaches"
The score captures:
actual— what really happened (factual, not interpretive)outcome— right / wrong / partial / unresolved (did reality match expectation?)reasoning_quality— separate from outcome! Good reasoning + bad outcome = variance, not error. Bad reasoning + good outcome = luck, not skill. The agent should score reasoning_quality honestly, not let the outcome contaminate it.notes— the lesson. What did the prediction error reveal about the user's mental model?
Critical scoring discipline: never let outcome contaminate reasoning_quality. A decision to hold a stock that then crashed isn't automatically bad reasoning — only mark reasoning_quality low if the reasoning itself was flawed (ignored a known risk, over-weighted a single data point, etc.). This separation is the whole point of the journal.
Review loop (scheduled)
Optionally wire date-triggered reviews into scheduled_task so due entries surface automatically:
# Register a daily check that pushes due decisions to the user
scheduled_task(
action="register",
title="Decision journal review",
schedule="0 9 * * *", # 9 AM UTC daily
# run.py calls: python3 skills/decision-journal/scripts/journal.py due
# if any due: push summary to user via /push
)
The run.py should: call due, if empty exit silently (no push), if non-empty push a summary listing each due decision's title + expected outcome + a prompt to score it.
This is optional — the journal works fine with manual review. But the scheduled loop is what turns it from a diary into a learning system.
Reading patterns
journal.py list --status open— all unresolved decisionsjournal.py list --tag investing— filter by tagjournal.py stats— aggregate: how many right/wrong/partial, average confidence, most common prediction errors (fromnotesfield across scored entries)journal.py find --query "keyword"— full-text search across decisions + reasoning + actual outcomes
The stats command is where the meta-learning lives. After 20+ scored entries, patterns emerge: "I'm consistently overconfident on X type of decision" or "my reasoning is good but my timing is off." That's the gold.
Storage
workspace/data/decision-journal.jsonl — append-only JSONL. One object per line. Never edit historical entries in place; scoring adds a review block to the same entry (so the original reasoning is preserved exactly as written, even if it was wrong).
Schema (see references/schema.md for full):
{
"id": "dj-2026-0142",
"ts": "2026-06-23T14:30:00Z",
"title": "Hold ZG despite PE 168",
"decision": "Keep position, no add",
"reasoning": "PE high but growth story intact, sector momentum...",
"expected": "Flat to +10% over 3 months, no >15% drawdown",
"review_on": "2026-09-23",
"confidence": 3,
"tags": ["investing", "a-share"],
"alternatives": {"sell_all": "tax event + timing risk"},
"invalidation": "Q2 miss or sector break MA60",
"status": "open",
"review": null
}
After scoring, review fills in and status flips to closed.
Behavioral rules
- Capture before outcome. If the user is describing a decision already made and resolved, logging it has limited value — the reasoning is already contaminated by knowing how it went. Still log if asked, but flag it.
- Falsifiable expectations only. Refuse vague expectations ("should do well"). Push back: "What specifically would confirm or falsify this?"
- Separate outcome from reasoning quality. This is non-negotiable. See scoring discipline above.
- Don't log routine execution. "I decided to rebalance today" is not a decision worth journaling. The journal is for non-obvious calls with learning value.
- Never rewrite history. Original reasoning is immutable. Scoring appends, never edits.
- Tags are how patterns surface. Encourage consistent tags (
investing,hiring,strategy,kill-decision) sostatscan aggregate.
References
references/schema.md— full JSONL schema, field semantics, scoring rubricscripts/journal.py— CLI:add,due,find,list,score,stats