Agent Skills: brief-best-practices

What a brief is

A BRIEF.md is the verifier's spec — the codified taste that says what "good" means for a surface, so an agent can verify its own iterations and run a loop without guessing or interrupting. Where SPEC.md is the contract (what to build: REQ-*, invariants), the brief is the bar (what "good" is, and who judges).

It is the BRIEF rung in the stack VISION → SPEC → BRIEF → HARNESS → LOOP → BOUNDARY: the brief is what makes the autonomous loop trustworthy, because the harness runs the brief's floors and the oracle judges against them.

When to author one

Author or update a brief when work will loop (you'll iterate against it more than once) or when the cost of being wrong is high. Trivial, one-shot changes need no brief — do not manufacture ceremony. Build cheap verifiers freely; propose an expensive brief+harness before investing in it.

Naming & placement

Always BRIEF.md, colocated with the surface it governs: root for project scope, apps/foo/BRIEF.md, packages/bar/BRIEF.md, or a docs subtree (e.g. docs/toys/BRIEF.md). Dated working memory lives beside the brief, never inside it — e.g. DELTA.md (per-round gaps, ranked) and DEVIATIONS.md (infeasible → nearest-feasible, logged). The brief is present-tense law; git is the changelog.

The seven slots (required concerns, adaptable shape)

The contract is that all seven concerns are present, in this order — not a fixed ##-count. The content adapts to the domain (a payments flow, an indexer, a CLI, a 3D toy); the concerns never change. A domain may split a concern across sections, or close with a culminating "Final acceptance" coda: one whole-surface test that restates the Boundary as a gate ("then the human ships it; the real gate is a user who…"). That coda is a framing of Boundary + Oracle, not an illegal eighth slot. The rule is don't drop a concern — not don't add a heading; still, keep sections as few as the content allows. Absent a house dialect (see Authoring rules), these seven are the skeleton.

1. Bar — one sentence: what "shippable" means for this surface. The north-star "done."
2. Dimensions — the few axes "good" decomposes into (correctness, idempotency, auditability, security, latency, recognizability…). Keep it short; these are the quality factors, not a feature list.
3. Floors — the minimum on each dimension with how it's measured (a floor without a measurement method is useless: "p95 < 200ms, measured via X"). The gate, not the ceiling — passing the floor licenses ship, not perfection.
4. Oracle — the independent verifier: what runs, who judges, and why it can't be gamed (maker ≠ judge). Pick the pattern that fits the surface: property tests or a staging run against forked state (objective surfaces); a deterministic simulator over fixed golden/archetypal inputs whose emitted trajectory a domain expert reads (pure engines — the inputs are fixed, so tuning to flatter one case visibly shifts the others); a blind human-judge quorum — fresh-context judges, ideally across vendors, naming the artifact with no context (subjective/taste surfaces, where no automated check can decide). For live systems the oracle extends past ship into telemetry — the prod signals that confirm it stays good.
5. Never — outcomes that are always a fail regardless of everything else (the safety invariants / "never events"). Concrete and absolute.
6. Decisions — calls already made, the tradeoff/priority policy ("security > latency; security can force a redesign, latency cannot"), and assumptions, so the agent never re-asks. This section grows: every answered question becomes a permanent entry. This is where mid-loop questions go to die.
7. Boundary — what requires the human: publish, biometric, live secrets, and genuine unknowns. Naming it tells the agent exactly what it may and may not do unattended.

Show, don't just tell. Any slot that is ambiguous earns a concrete instance — a golden example and/or an anti-example. Agents ground on exemplars; the Never list and the Oracle especially benefit.

Governance preamble

Open every brief with a one-line law statement, e.g.:

Law doc for <surface>, present-tense, no narrated history — git is the changelog. Amend Decisions and Boundary only with human confirmation; log the rationale. Dated working memory lives in DELTA.md / DEVIATIONS.md beside this file.

Authoring rules

• Match the house first. If the repo already has ratified briefs, copy their shape — section names, voice, any closing coda — over this skeleton. Consistency across the brief set beats the generic template; the seven concerns are the fallback when no house dialect exists yet. A reviewer judges a brief against the house dialect, and must not reject it for matching the repo's own law.
• Evidence-based. Ground Dimensions and Floors in the real surface; cite reference exemplars. Do not invent thresholds, signals, or behaviors.
• The oracle must be independent. Maker ≠ judge for any subjective dimension. Name why it can't be gamed — without independence the gate is theater.
• Floors are gates, not ceilings. A passing artifact may still owe refinement; say so. Never weaken a floor to pass a gate — an infeasible item gets the nearest-feasible alternative plus a DEVIATIONS.md entry, and the gap stays on record.
• Calibrate claims to enforcement. Match absolutist words ("never", "cannot") to what the oracle actually proves. Overclaiming invites reject cycles.
• Parsimony. Few, well-crafted floors that cover the cases beat a long brittle list. The brief is read every loop; every line earns its place.
• Mutation policy. Do not edit Decisions or Boundary without explicit human confirmation. When brief/implementation drift is found, surface it — the human decides.

Lifecycle

• Creation. When work begins to loop or the cost of being wrong is high. Draft the seven slots; the human ratifies. The harness is built to run the Floors; the Oracle is wired before iteration starts (harness-first).
• Maintenance. The Decisions section grows as questions are answered — every AskUserQuestion that should never recur becomes an entry. Floors tighten as the bar rises (rewrite as if always true). Cross-check the brief against the implementation whenever both are in context; surface drift.
• Retirement. When a surface is removed, remove or archive its BRIEF.md. Do not leave a stale law describing deleted behavior.

How the brief drives the loop

The brief is inert until it runs:

1. The harness runs the Floors and emits pass/fail with evidence (cheap, fast, fail-closed).
2. The oracle judges the subjective Dimensions independently.
3. The loop iterates act → verify → orient → decide until every Floor passes, or terminates as a bounded, honest blocked (with evidence and a proposed path).
4. Presence axis: attended, the human may opt out of ceremony for trivial work; unattended, the brief + harness are the only backstop — rigor is maximal and the Boundary (publish, secrets, biometric) is never crossed by the loop.

References

• template.md — the blank seven-concern skeleton, copy-paste ready when no house dialect exists yet; if the repo already has briefs, mirror those instead.
• example-payments.md — a filled brief for a money-transfer flow: the objective archetype, where floors are machine-checkable and the oracle is property tests + staging. The subjective/taste archetype — a visual or design surface whose bar is "elegant, calm, recognizable" and whose oracle is a blind human-judge quorum (maker ≠ judge) closing on a culminating "Final acceptance" frame — is the harder, more common product case; build its oracle from the blind-judge pattern in the Oracle slot.