Write Decision Memos Where the Trade-offs Are Real and the Conviction Is Yours

Workflow area: Delivery & Reporting · Tier: Foundation · Hands-on time: ~25 minutes

Engineering leaders make dozens of consequential calls a quarter — build vs buy, migrate vs patch, this vendor or that one — and six months later nobody can reconstruct why. The ADR discipline fixes that, but blank-page friction means most memos never get written. AI removes the friction, and introduces a new failure mode: models write recommendations with smooth, borrowed conviction, and a leader who signs one they didn't stress-test has outsourced judgment, not typing. The structure is AI's job; the call is yours.

When to use — and when not to

• Use it for one-way doorsarchitecture commitments, platform migrations, vendor lock-ins, deprecations with external users: anything expensive to reverse deserves a memo before it deserves a decision.
• Use it for build-vs-buy and tool selectionsthe options-and-trade-offs shape forces you to steelman the choice you were about to dismiss.
• Use it retroactively for decisions already causing confusiona memo written late still beats tribal memory.
• Skip it for two-way doorscheap, reversible calls should be made in minutes and revisited if wrong; a memo process on reversible decisions is bureaucracy cosplay.
• Skip it as decision launderingif you have already decided and want a memo that makes the choice look inevitable, don't. A memo whose options section is a strawman parade fools nobody and poisons the format.

The recipe

1. 1.

   Dump the raw context first. Before any prompting, write five ugly minutes of notes: what triggered the decision, constraints (budget, deadline, headcount, compliance), stakeholders, and the options actually on the table — including "do nothing." The model structures; it cannot know your constraints.

2. 2.

   Have the model build the skeleton and steelman every option. Fixed template, every time:

   Copy

   You are helping me draft a decision memo. I will make the final
   recommendation myself — do NOT recommend an option.
   
   Context notes: [PASTE NOTES]
   Options under consideration: [LIST, INCLUDING "DO NOTHING"]
   
   Produce:
   1. CONTEXT — the situation and constraints in ≤150 words.
   2. OPTIONS — for each: what it is, strongest honest case FOR,
      strongest honest case AGAINST, rough cost/effort, and key
      risks. Steelman every option equally, including do-nothing.
   3. TRADE-OFF TABLE — options vs the criteria implied by my
      constraints. Mark any cell you inferred with [ASSUMED].
   4. REVERSIBILITY — classify each option one-way or two-way
      door, and what reversal would cost.
   5. OPEN QUESTIONS — what I'd need to answer to decide well.
   
   Do not soften cons. Do not add a recommendation section.
   

3. 3.

   Interrogate the [ASSUMED] cells. Every assumed cost, risk, or capability is a place the model guessed. Confirm or correct each one — this is where hallucinated vendor capabilities and invented migration costs die. Good output looks like:

   Copy

   TRADE-OFF TABLE
   Option          | Cost to ship | Ops burden      | Exit cost
   Build in-house  | ~2 eng-qtrs  | ours forever    | low
   Buy VendorX     | ~3 wks integ | [ASSUMED] low   | high — data
                   |              |                 | export unclear
   Do nothing      | 0            | rising incidents| n/a
   
   REVERSIBILITY
   - Build: two-way for 1 quarter, one-way after teams integrate.
   - Buy: one-way door — [ASSUMED] no bulk export API. VERIFY.
   

   That VERIFY on the export API is the memo earning its keep.

4. 4.

   Write the recommendation section yourself, by hand. Three sentences minimum, in your voice: which option, why the losing trade-offs are acceptable, and what evidence would change your mind. If you cannot write the "what would change my mind" line, you are not done deciding. Never ask the model to draft this section — a recommendation you can defend under hostile questioning is the product.

5. 5.

   Circulate for dissent, then file it as an ADR. Share with the people who will live with the decision, with a named comment window ("dissent by Friday, silence is consent"). Then commit the final memo where decisions live — an adr/ directory in the repo, or your docs system — numbered, dated, with status Accepted. Update the status if it is later superseded; a contradicted ADR left marked Accepted is worse than none.

Make it yours

• Small team or startup: keep memos to one page and file them as markdown ADRs in the repo — adr/0007-buy-vendorx.md. The discipline matters more than the polish; you are writing for the you of eighteen months from now.
• Large org or GCC: use your tenant's docs AI (Copilot in Word, Duet in Docs) for the skeleton so drafts stay inside the compliance boundary, and adopt one shared template across leads — comparable memos make architecture review boards dramatically faster. Client-impacting decisions get the client-visibility check before circulation.
• Regulated environments: add a section for compliance and data-residency implications per option, and route one-way-door memos through your architecture or risk board as a standing agenda item — the memo format is exactly what auditors ask for after the fact.

Quality check

• Every option has a real case for itif one option's "for" section is limp, you strawmanned it; rewrite before circulating.
• All [ASSUMED] cells are resolveda trade-off table with unverified model guesses is fiction formatted as rigor.
• The reversibility call is explicitone-way doors decided at two-way-door speed is the classic postmortem finding; the memo exists to prevent it.
• The recommendation is yoursread it aloud; if it sounds like a model's balanced prose rather than a leader taking a position, rewrite it until it commits.
• Dissent was actually possiblea memo circulated after the purchase order is documentation, not decision-making; check the timestamps.