Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Creates new Architecture Decision Record (ADR) documents for significant architectural changes using a consistent template and repository-aware naming and storage guidance. Use when a user or agent decides on an architectural change, needs to document technical rationale, or wants to add a new ADR to the project history.
.claude/skills/giuseppe-trisciuoglio-adr-drafting/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 875% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 105% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 43% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 343% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 76% | 0% |
Creates new Architecture Decision Record (ADR) documents for major architectural choices so teams can keep a clear history of why important technical decisions were made.
This skill helps create a new ADR from discovery to final markdown file. It confirms the decision details, inspects the repository for any existing ADR conventions, and drafts a new ADR with the standard sections Title, Status, Context, Decision, and Consequences.
When the repository does not already have an ADR convention, default to storing ADRs in docs/architecture/adr and use a zero-padded filename such as 0001-use-postgresql-for-primary-database.md.
See references/template.md for the default ADR template and references/examples.md for example ADRs and naming patterns.
Use this skill when a user or agent has decided on a meaningful architectural change and needs to document the rationale, chosen direction, and trade-offs in a new Architecture Decision Record. It fits requests such as creating an ADR, documenting an architecture decision, writing a decision record, or preserving the project history behind an important technical choice.
Ask the user for the minimum information needed to draft a new ADR:
Proposed by default if not yet finalized)If the user actually wants to update an existing ADR, change statuses in older ADRs, or manage supersession links, explain that this skill only drafts new ADR documents and ask whether they want to proceed with a new record instead.
Inspect the repository before drafting:
docs/architecture/adrdocs/adradrarchitecture/adrdocs/architecture/adr0001Preferred default naming when no convention exists:
docs/architecture/adrNNNN-short-kebab-title.md# ADR-NNNN: <Decision Title>Create a draft using the standard structure:
markdown# ADR-NNNN: Decision Title ## Status Proposed ## Context What problem, constraints, or trade-offs led to this decision? ## Decision What architectural choice was made? ## Consequences What becomes easier, harder, riskier, or more expensive because of this decision?
Drafting rules:
Before writing files, present:
Ask for approval before creating the file. If the user wants adjustments, revise the draft first.
After approval:
User request: "Create an ADR for moving from SQLite to PostgreSQL"
Expected flow:
docs/architecture/adr/0001-move-to-postgresql.mdUser request: "Document the decision to split billing into a dedicated service"
Expected flow:
See references/examples.md for longer ADR examples.
Proposed unless the user confirms another statedocs/architecture/adr if the repository already uses another ADR location| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 3,597 | 4,561 | +27% | 1 | 1 | 0% | 304 | 1,565 | +415% | 0 | 0 | — |
case-02 | fail→pass | 2,266 | 8,569 | +278% | 1 | 1 | 0% | 275 | 2,680 | +875% | 0 | 0 | — |
case-03 | fail→fail | 5,297 | 6,147 | +16% | 1 | 1 | 0% | 337 | 1,645 | +388% | 0 | 0 | — |
case-04 | fail→pass | 7,043 | 4,291 | -39% | 1 | 1 | 0% | 989 | 2,023 | +105% | 0 | 0 | — |
case-05 | fail→pass | 7,022 | 2,561 | -64% | 1 | 1 | 0% | 1,217 | 1,741 | +43% | 0 | 0 | — |
case-06 | fail→pass | 3,319 | 2,448 | -26% | 1 | 1 | 0% | 372 | 1,649 | +343% | 0 | 0 | — |
case-07 | pass→pass | 3,936 | 3,364 | -15% | 1 | 1 | 0% | 558 | 1,793 | +221% | 0 | 0 | — |
case-08 | fail→pass | 5,999 | 2,068 | -66% | 1 | 1 | 0% | 915 | 1,610 | +76% | 0 | 0 | — |
case-09 | pass→pass | 6,814 | 3,008 | -56% | 1 | 1 | 0% | 1,039 | 1,819 | +75% | 0 | 0 | — |
case-10 | fail→fail | 8,335 | 15,094 | +81% | 1 | 1 | 0% | 766 | 2,676 | +249% | 0 | 0 | — |
case-11 | fail→pass | 12,378 | 6,588 | -47% | 1 | 1 | 0% | 1,783 | 2,398 | +34% | 0 | 0 | — |
case-12 | pass→pass | 8,342 | 3,252 | -61% | 1 | 1 | 0% | 1,230 | 1,865 | +52% | 0 | 0 | — |
case-13 | pass→pass | 1,993 | 2,469 | +24% | 1 | 1 | 0% | 341 | 1,713 | +402% | 0 | 0 | — |
case-14 | fail→pass | 5,767 | 3,108 | -46% | 1 | 1 | 0% | 842 | 1,776 | +111% | 0 | 0 | — |
case-15 | pass→pass | 22,995 | 7,395 | -68% | 1 | 1 | 0% | 1,831 | 2,337 | +28% | 0 | 0 | — |
case-16 | fail→pass | 14,656 | 2,530 | -83% | 1 | 1 | 0% | 1,393 | 1,700 | +22% | 0 | 0 | — |
case-17 | pass→pass | 8,196 | 2,914 | -64% | 1 | 1 | 0% | 1,264 | 1,824 | +44% | 0 | 0 | — |
case-18 | pass→pass | 9,043 | 2,706 | -70% | 1 | 1 | 0% | 1,338 | 1,729 | +29% | 0 | 0 | — |
case-19 | fail→pass | 4,618 | 3,134 | -32% | 1 | 1 | 0% | 544 | 1,778 | +227% | 0 | 0 | — |
case-20 | pass→pass | 13,959 | 6,544 | -53% | 1 | 1 | 0% | 2,141 | 2,220 | +4% | 0 | 0 | — |
case-21 | fail→pass | 10,380 | 2,530 | -76% | 1 | 1 | 0% | 1,615 | 1,732 | +7% | 0 | 0 | — |
case-22 | pass→pass | 13,715 | 3,397 | -75% | 1 | 1 | 0% | 1,706 | 1,761 | +3% | 0 | 0 | — |
case-23 | fail→pass | 8,032 | 2,446 | -70% | 1 | 1 | 0% | 1,237 | 1,680 | +36% | 0 | 0 | — |
case-24 | fail→pass | 12,627 | 2,948 | -77% | 1 | 1 | 0% | 1,791 | 1,756 | -2% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 24 cases were attempted, and 22 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +50 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.