Decision Log

Capture decisions in a consistent, ADR-style (Architecture Decision Record) format so that three months from now, the user — or the next PM in the seat — can reconstruct why, not just what.

When to apply this

Use this skill when the decision has any of these properties:

• Costly to reverse. Model choice, vendor commitment, schema migration, public API shape.
• Likely to be re-litigated. Anything where someone will ask "wait, why didn't we do X?" later.
• Bound to a model or version. AI decisions tied to a specific model (Sonnet 4.6, Opus 4.7, etc.) — these will be revisited at the next model migration, and the rationale must survive.
• Made under uncertainty. When you're picking with incomplete information; future-you needs to know what was unknown at the time.

Skip for: trivially reversible choices, daily execution decisions, anything captured well enough by a commit message or PR description.

The format

Use this structure. Keep each section short — the whole thing should fit on one screen.

# ADR-NNN: [Short, declarative title]

**Date:** YYYY-MM-DD
**Status:** Proposed | Accepted | Superseded by ADR-MMM | Deprecated
**Deciders:** [names/roles]
**Reversibility:** Easy | Moderate | Hard | One-way door

## Context
What is the situation that requires a decision? What constraints apply?
Two to four sentences. State only facts and constraints, not the choice yet.

## Options considered
- **Option A:** [name]. Pros: … Cons: …
- **Option B:** [name]. Pros: … Cons: …
- **Option C:** [name]. Pros: … Cons: …

(Always at least two options. "We had no choice" is rarely true and worth interrogating.)

## Decision
We chose **[Option X]**.

## Rationale
Why X, given the trade-offs. Be specific about which constraint or value tipped the call.
This is the section future-you will care about most.

## Consequences
- What this enables
- What this forecloses
- What we'll have to live with that isn't great

## Revisit triggers
Concrete events or thresholds that should prompt revisiting this decision.
Examples: "next model family release", "if eval drops below 0.85", "if request volume 10x's", "Q4 planning".

AI-specific decisions: extra fields

For decisions about models, prompts, evals, or agent design, add:

• Model & version: Pin the exact model ID at decision time (e.g., claude-sonnet-4-6).
• Eval state: Score on the current golden set, and the eval version.
• Cost/latency at decision time: Per-request cost, p50/p95 latency.
• Fallback path: What happens when this choice fails? Named explicitly.

These fields are what make the decision survive a model upgrade. Without them, the next person re-derives the trade-off from scratch.

Where to keep it

Recommend the user store ADRs in the repo, in a top-level docs/decisions/ folder, numbered sequentially (ADR-001-foo.md, ADR-002-bar.md). They live next to the code, get reviewed in PRs, and survive re-orgs. A wiki tab is a graveyard.

How to help the user

1. Diagnose first. Confirm the decision is worth logging (use the criteria above). If it's trivially reversible, say so and offer a one-liner instead.
2. Probe for the missing options. If the user presents only one option, push: "What were the alternatives? What did you reject and why?" Decisions with only one stated option usually have hidden ones.
3. Push for revisit triggers. Most users skip this section. It's the highest-leverage one.
4. Write the ADR yourself in the structure above, marked clearly as a draft. The user can sharpen it; don't make them assemble from a template.
5. Suggest the filename. Following the ADR-NNN-short-title.md convention. Look for existing ADRs in the repo to get the next number right.

Anti-patterns to flag

• "Decision: we'll figure it out later" — this isn't a decision; convert to an open question with an owner and a deadline.
• Rationale that's actually a restatement — "We chose X because X is better" → push for the underlying value or constraint.
• Missing reversibility classification — affects how much rigor the decision deserves; don't let it slide.
• Vague revisit triggers — "when relevant" means never. Replace with a concrete event or metric.