| name | decision |
| description | Fast path for recording a routine decision as a light ADR (200-400 words, 6-section DDR shape). Use when you just chose one approach over another and want to capture the why before it slips out of memory. NOT for big architecture — that's `adr-architect` agent territory.
Triggers: "record decision", "/decision", "запиши решение", "quick adr", "light adr", "ddr"
|
| origin | forgeplan |
/decision — quick decision record
You are helping the user capture a routine decision they JUST made (lib choice, config tweak, small refactor approach) as a light ADR (200-400 words, 6 sections including Revisit Trigger).
When to use this skill (vs adr-architect agent)
Use /decision (this skill) | Use adr-architect agent |
|---|
| Decision touches 1-2 modules | Decision touches ≥3 modules |
| User says "let's record this", "quick ADR" | User says "full ADR", "major architectural decision" |
| Routine choice (lib, config, small refactor) | System-level decision (supersedes ADR, multi-sprint scope) |
| ADI surfaces ≤2 trade-offs | ADI surfaces ≥3 unresolved trade-offs |
| You're mid-session and don't want to break flow | You can spend 30+ min on the decision |
If in doubt — start with /decision (light). Easier to escalate from light to full than to compress full into light.
Procedure
Step 1 — Read the template
Read ${CLAUDE_PLUGIN_ROOT}/templates/adr-light.md
Use it as your structure — 6 named sections.
Step 2 — Gather facts from the live session
Don't ask the user for things you can already see. From the current chat context:
- Context: what was discussed in the last 1-3 turns leading to the choice.
- Hypotheses considered: which options came up in conversation. If only one was discussed, prompt the user for two more (DDR rule: minimum 3).
- Decision: which option the user picked.
Ask the user ONLY for what's missing:
- Revisit Trigger (the key MUST field — usually not in chat history)
- Reasons for rejecting alternatives (if not already stated)
Step 3 — Score F+G+R per hypothesis
For each hypothesis, score by the rubric in the template:
- F (Formality): 0 if "gut feel", 9 if "spec/proof"
- G (Granularity): 0 if "slow", 9 if "p99=47ms @ 10k RPS"
- R (Reliability): 0 if "Slack anecdote", 9 if "peer-reviewed or our prod benchmark"
If the chosen hypothesis's F+G+R sum <12, OFFER to dispatch the evidence-gatherer agent for rigorous evidence collection. Do NOT force this — it's user's choice whether the rigour is worth the time.
Task(
subagent_type = "agents-pro:evidence-gatherer",
prompt = "Gather evidence for hypothesis '<H>'. Current F+G+R sum is <N> (below 12). Search 20-30 sources across 5+ classes, score R per source, ask user for production data if applicable, return per-source breakdown."
)
When the agent returns, take its synthesis EVID's aggregate F+G+R and re-score the hypothesis. Update the light ADR's Evidence table with the new numbers + cite the evidence-gatherer EVID by ID. If still <12 after rigorous gathering — be honest about the weak foundation in the Decision section ("chose this despite weak evidence because ").
Step 4 — Create the ADR
adr = forgeplan_new(kind="adr", title="<one-sentence decision>")
forgeplan_update(id=adr.id, body=<filled adr-light.md template>)
Step 5 — Create EVID + activate
The EVID for a light ADR can be inline (1-2 paragraphs verifying the decision was actually applied — e.g., "config X is now in production as of commit Y"):
evid = forgeplan_new(kind="evidence", title="Verification — decision <NNN> applied", parent_id=adr.id)
forgeplan_update(id=evid.id, body=<bold-pattern with Verdict: APPLIED, Congruence level, Evidence type>)
forgeplan_activate(id=evid.id)
forgeplan_activate(id=adr.id)
Step 6 — Confirm to user
Tell the user:
- ADR-NNN created and active
- Revisit Trigger summary (1 line — when this will need re-opening)
- If F+G+R was low: remind about
evidence-gatherer as a follow-up if rigour matters
Hard rules
- Never skip Revisit Trigger. This is the DDR distinction from old-style ADR. A decision without a trigger is a memorial.
- Never artificially inflate F+G+R. If it's a gut feel, write F=2 G=2 R=2. Self-honest weak score is more valuable than a fake strong one.
- If you exceed 400 lines, STOP and switch to
adr-architect full template. This is a light record; if it needs more, the decision is bigger than light scope.
- Never wait on the user for >2 follow-up questions. Make reasonable defaults, note what's assumed, let the user correct after the ADR is created.
- Default Depth = Tactical. Light ADR is by definition tactical scope.
References
- Template:
plugins/fpl-skills/templates/adr-light.md
- Full ADR alternative:
plugins/fpl-skills/templates/adr-full.md (via adr-architect agent)
- DDR methodology (user-provided 2026-05-24): 6 sections, 200-400 words, mandatory Revisit Trigger
- Sprint Z2 (PRD-053): Revisit Trigger enforcement via guardian + decay-watch
- Sprint Z4 (PRD-055): evidence-gatherer agent for rigorous F+G+R scoring
