distill: weak request in, sharp brief out

A great brief removes ambiguity. It does not add volume. Your output should be executable by a weaker model without access to this conversation: every fact it needs is on the page, and nothing on the page is filler.

When to use

• "bro add dashboard make it good"
• "turn this rambling into a proper spec for the agent"
• "spec this typo fix, and keep it tiny, no enterprise ceremony"

When NOT to use

• The user asks whether to pursue an idea or choose between products → hotseat. An uncritiqued idea alone does not require debate.
• The request is already precise → the current owner can execute it; use masterplan only when complexity warrants planning.
• The task is a small mechanical fix ("rename this variable") → just do it.

Prerequisites

The triggering request. A quick look at the repository for context (what exists, what conventions apply): lighter than a spelunk pass; read only what the request names or obviously touches.

Evidence, authority, and assumption model

Distill clarifies intent; it does not silently become a product owner, architect, security approver, or implementation agent. Separate all brief content into these classes:

Never infer permission to delete data, break a public contract, change pricing, send external communication, access production data, spend money, or deploy. Those are explicit approval boundaries in the brief when relevant.

Brief decision framework

Distill only the ambiguity that matters to a downstream implementer. For each candidate detail, ask: does an alternative change user outcome, scope, data, permission, compatibility, risk, verification, cost, or release behavior? If not, omit it or make a narrow assumption. If yes, decide it, ask it, or record it as a deliberately unresolved decision with a safe discovery step.

Write observable requirements as behavior under conditions:

The form is a tool, not a ritual. Pair each material requirement with an oracle: visible result, returned contract, persisted state, log/audit event, or testable absence of a forbidden side effect. "Intuitive", "fast", "secure", and "production-ready" are labels until their observable condition is named.

Scale the brief to the request

A brief for a button rename that mentions "non-goals" and "rollout" is enterprise cosplay: match depth to blast radius.

Workflow

1. Extract the goal

Restate the request as one sentence: what will be true after the work that isn't true now, for whom. If you cannot write this sentence, that's a clarifying question, not a brief.

2. Ground in the repo first (lightly)

Before drafting a single question, confirm the surfaces the brief touches actually exist: the page/module named, the data involved, the commands available. Briefs referencing nonexistent files are how hallucinated work starts. Most "ambiguities" die here: what remains is genuinely missing.

3. Ambiguity detection (max 5 questions)

Actively scan the request for linguistic, semantic, and functional ambiguities. List what's genuinely blocking:

• Undefined boundaries ("make it fast" → what is the target metric?)
• Contradictory constraints
• Missing definitions of "good" or failure states

For each candidate question, apply the materiality test: name the two most plausible answers and what each would change about implementation, security, data handling, or UX. If nothing material changes, it is not a question, it becomes an explicit numbered assumption in the brief.

Ask only blocking questions, max 5, in one batch. Anything non-blocking becomes an explicit numbered assumption in the brief. Never run a back-and-forth interrogation.

4. Write the brief

Use the template in references/brief-template.md (it includes a filled example). Core sections:

• goal (one sentence) + user outcome;
• context: relevant existing pieces, with file paths;
• constraints and conventions to respect;
• scope: what's in;
• non-goals: what's explicitly out (as valuable as scope);
• agent-executable specifications: statements another agent can objectively verify without human judgment;
• edge cases: empty/loading/error/malformed states worth handling;
• non-functional requirements only when they bite (performance budget, accessibility bar, i18n, security constraint): omitted means "nothing beyond the project's existing bar";
• backward compatibility / migration notes when existing data or users are affected;
• UX expectations (only if user-facing);
• verification: how "done" gets proven: which commands must pass, what a human/browser check confirms;
• known unknowns + numbered assumptions.

5. One-pass confirmation

Show the brief with unresolved decisions distinguished from safe assumptions. If the user authorized implementation and no material decision remains, proceed within that scope; a brief is not a new approval gate. If the user requested only requirements, deliver them and stop. Ask only for a load-bearing answer.

6. Run the implementation-readiness review

Read the finished brief as an implementer with no conversation history and apply the review in references/brief-template.md (Implementation-readiness review): every item must pass before the brief ships.

If a missing item would change implementation substantially, ask the compact question batch or return a conditional brief. Do not conceal ambiguity in polished prose.

Failure handling and scope change

• Conflicting user statements: quote the conflict neutrally and ask for the decision; do not choose the more convenient interpretation.
• Repository contradicts request: preserve the request, report the observed conflict and path, then offer a revised scope or discovery task.
• Unsettled product tradeoff: frame options, consequences, and evidence needed; route to hotseat instead of pretending it is a requirement.
• Insufficient repository access: emit a provisional brief with every path/command marked assumption, then name the minimum inspection needed.
• Mid-implementation scope change: update the goal, scope, non-goals, affected acceptance criteria, and assumptions together. Do not patch only a task line while leaving the original definition of done misleading.
• Approval-bound action: stop at a precise decision/authorization request. The brief may specify preparation but not authorize execution.

Anti-bloat rules

• Every line must remove ambiguity or it gets deleted.
• No restating general knowledge ("React is a component library"): the reader is a coding agent, not a student.
• No invented requirements the user never implied. Suggestions go in a clearly marked non-binding suggestions list.
• Stop when the brief removes implementation-blocking ambiguity; brevity is a quality signal, not a fixed line target.

Tool selection / fallback

• Repository artifacts and the user's words outrank generic templates.
• Ask only when alternatives materially change behavior, data, permissions, or risk.
• Without repository access, produce a provisional brief and label path or command assumptions.

Quality gates

• A reader who never saw this chat could implement without guessing on scope: test your draft against this before showing it.
• Every agent-executable specification is objectively checkable by another agent without human intervention.
• The verification section names real commands/checks (that exist in this repo or are standard).
• Assumptions are numbered and falsifiable ("A1: single-user, no roles").
• Marker floor (applies at every scale, including compact briefs): the brief contains a literal Non-goals line and at least one numbered assumption (A1:), even if each is a single line. A brief without these markers is incomplete, not concise.
• Auth-bites rule: when the request touches login, identity, per-user data, sessions, tokens, passwords, or money, the brief names that surface explicitly (one line suffices, e.g. under Scope or NFRs). Omitting it because the brief is short is a miss, not brevity.

Stop conditions

• Brief removes blocking ambiguity → deliver for a requirements-only request; otherwise continue the already-authorized next action. Use masterplan only when the implementation needs it.
• The request reveals an unsettled product decision → stop and route back to hotseat for that decision.

Output contract

In chat: the brief. On disk (when the task will span sessions or feed planning): docs/briefs/<slug>.md with provenance header (<!-- generated by distill on YYYY-MM-DD -->).

<slug>: short kebab-case text derived from the task title.

Research basis

Read references/research.md when a decision depends on an external standard, a numerical claim, or a fast-moving practice. The ledger records what the source supports, what it does not support, and when to reverify.