Getting it into your agent
It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.
git clone --depth 1 https://github.com/romainsimon/skills-for-decision-makingnpx agentmods add skills/romainsimon/skills-for-decision-making/learning-from-outcomesWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/romainsimon/skills-for-decision-making/learning-from-outcomes)<a href="https://agentmods.dev/skills/romainsimon/skills-for-decision-making/learning-from-outcomes"><img src="https://agentmods.dev/badge/skills/romainsimon/skills-for-decision-making/learning-from-outcomes/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/romainsimon/skills-for-decision-making/learning-from-outcomes"><img src="https://agentmods.dev/badge/skills/romainsimon/skills-for-decision-making/learning-from-outcomes.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00079 | $0.02043 |
| Opus 5 | $0.00039 | $0.01022 |
| Sonnet 5 | $0.00016 | $0.00409 |
| Haiku 4.5 | $0.00008 | $0.00204 |
Grade A, and why
learning-from-outcomes scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 11d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 208 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Learning from outcomes
Experience does not automatically produce judgment. Without a written prediction and a scoring step, a review produces a story in which the good outcomes were foreseen and the bad ones were unforeseeable. That story feels like learning and teaches nothing.
Three things make a review compound:
- Score the predictions, not the outcomes.
- Assign credit across time, because results arrive long after the decisions that caused them.
- Replay old decisions deliberately, because recent experience otherwise overwrites older lessons entirely.
Workflow
- [ ] 1. Collect the predictions that came due
- [ ] 2. Score them
- [ ] 3. Separate decision quality from outcome quality
- [ ] 4. Assign credit for delayed results
- [ ] 5. Replay two old decisions that are not due
- [ ] 6. Update the priors that were wrong
1. Collect what came due
Pull every prediction from the decision log whose date has passed. Record the outcome as a plain binary before discussing anything, and before anyone explains why it happened.
If there is nothing to collect, the problem is upstream: decisions are being recorded without falsifiable predictions. Fix that in framing-decisions first; nothing here works without it.
2. Score them
Run the calculator. Do not compute these by hand.
node scripts/calc.js calibrate predictions.json
Brier decomposition, bin assignment and the base-rate benchmark are exactly the kind of arithmetic that comes out subtly wrong when done in prose, and a miscalibration verdict derived from a wrong number is worse than no verdict. Paste the real output into the review.
Input shape: examples/calibrate.json. Four numbers matter:
- Brier score - mean squared error of your probabilities. Lower is better.
- Always-base-rate - what you would have scored by ignoring specifics and always quoting the base rate. If your Brier score is not beating this, your judgment is adding nothing and the honest response is to quote base rates until it does.
- Verdict - overconfident, underconfident, or well calibrated.
- The bin table - of the things you called 80%, how many happened? Overconfidence concentrates in the high bins, which is where it does the most damage.
What ships with it
5 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 11d ago First seen · 208 lines · 79 tokens per session scan A ebeb0ebcbfee
learning-from-outcomes is a skill published in the GitHub repository romainsimon/skills-for-decision-making (11 stars, last pushed 1mo ago), licensed MIT. It adds 79 tokens to every session and 2,043 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
thinking-model-router
When unsure which thinking skill fits, map domain and problem type, then return NONE or one primary skill by default (at most three complementary).
thinking-systems
When behavior is emergent across components—fixes elsewhere break, loops/delays dominate—map boundary, stocks/flows, feedback, archetypes, then rank leverage.
architecture-aware-init
Selects architecture paradigm via research before scaffolding. Use when architecture is undecided and the choice needs justification and documentation.
thinking-five-whys-plus
When a fault is localized and the proximate cause is known but the systemic root is not, chain evidence-linked whys with a counterfactual stop and a countermeasure.
thinking-map-territory
When a claim, doc, test, metric, or assumption conflicts with observed behavior, stop theorizing from the map and verify the live code or data; let territory overrule.
thinking-theory-of-constraints
When throughput or latency is pipeline-limited, identify the single binding constraint and exploit, subordinate, elevate, then recheck—ignore non-constraints.