Synthesis Quick Answers

Synthesis Quick Answers — Lookup Companion Pattern

The Problem

Focused project sessions accumulate context on purpose — that's what makes them good at the work they're for. But not every question belongs there. "When is a colleague back from vacation" and "has a release shipped yet" are real, frequent, and usually urgent-feeling questions that have nothing to do with whatever a focused session is mid-task on. Answering them inline does two kinds of damage:

• Context pollution. The focused session's window fills with unrelated lookups, and its summarized history gets noisier every time it compacts.
• Cost mismatch. A one-line factual lookup doesn't need the reasoning depth or the model tier a hard architecture or strategy session runs at. Paying Max-effort-tier prices for "when is X back" is waste, repeated daily.

The fix is not "be disciplined about not asking." The questions are legitimate and often time-sensitive. The fix is a separate, cheap, low-ceremony surface built for exactly this shape of question — one that exists specifically so the other sessions can stay clean.

The Pattern

A quick-answers companion: one ongoing-status project inside the user's own personal knowledge workspace (per synthesis-project-management, at ~/workspaces/{workspace}/ai-knowledge-{workspace}/ — the same location synthesis-onboarding scaffolds), used from a session on a routine-tier model, whose entire mandate is answering lookups by reading everything else in the workspace and writing almost nothing back.

It is deliberately not a "seat" in the operations sense (compare an <org>-operations-style project, which owns rituals, syncs, and triage). It owns no workflow. It is a read path across every other project, plus the workspace's knowledge base — and its value depends on staying that narrow. The moment it starts drafting, deciding, or sending, it has become a second copy of the work it exists to protect other sessions from absorbing.

"Automatic" is a file, not a habit. The routing line lives in the workspace's Git-tracked .agents/workspace-AGENTS.md. The onboarding engine exposes that one source to both clients through the workspace-root AGENTS.md and CLAUDE.md entry points. A companion without the routing line still has to be invoked by hand, which is the friction this pattern exists to remove.

Configuration

Setup

This pattern needs a personal knowledge workspace — the same ai-knowledge-{workspace} repo synthesis-project-management and synthesis-onboarding already use, at ~/workspaces/{workspace}/ai-knowledge-{workspace}/. Don't invent a substitute location (a loose folder somewhere else, a name that doesn't match): a companion that lives outside the convention every other project already follows is a second, incompatible system, not a lighter version of the same one.

1. Ensure the personal knowledge workspace exists. Run synthesis workspace ensure --name {name}. It safely scaffolds ~/workspaces/{name}/ai-knowledge-{name}/, including a seeded projects/index.yaml, a Git-tracked .agents/workspace-AGENTS.md, a tracked .agents/knowledge-base.yaml declaring the source/ bundle, and collision-safe workspace entry points for both clients. It is idempotent when the workspace already exists, and it leaves every seeded file alone once you have edited it. If the synthesis command is not installed yet (a plugin-only installation has no launcher), install it first with the stable bootstrap: curl -fsSL https://raw.githubusercontent.com/synthesisengineering/synthesis-skills/stable/onboard.sh | sh -s -- setup --profile skills-only. A shared organization workspace is a sibling, not a substitute; this pattern's project files belong in the personal repository.
2. Create the project the normal way (synthesis-project-management): status: ongoing, bounded: false, noun-first id (e.g. {workspace}-quick-answers). Give it a thin CONTEXT.md and, if the workspace has enough standing routing knowledge to be worth writing down (which sources answer which question shapes), a short REFERENCE.md.
3. Register it in projects/index.yaml under its own initiative if the workspace doesn't already have a natural home for "standing non-ops infrastructure" — don't force it under an operations initiative that implies it owns rituals.
4. Add one routing line to the tracked source at ~/workspaces/{name}/ai-knowledge-{name}/.agents/workspace-AGENTS.md. Never edit the workspace-root entry points: the onboarding engine owns those and both clients resolve the tracked source through them. Commit the source change in the personal knowledge repository so a new machine receives it.
5. Tell the user which model tier to select for the session (see Configuration), and that this is a one-time-per-session setting they make, not something this skill can do for them.
6. Seed resources/FAQ.md with a header; leave it empty otherwise. It fills from use.

Operating Protocol

Run this per question, every time — it's the whole point of the pattern:

1. Classify before searching. What kind of fact is this?

   • A person's status/availability/role → team directory, calendar, recent Slack/chat, the KB's people/org docs.
   • A team's charter/roster/current work → the KB's org docs, that team's tracked projects.
   • A project's status/history/decision → that project's own index.yaml entry and CONTEXT.md/REFERENCE.md — not a full re-read of its session history.
   • A release/ship fact → git tags, changelog, release notes, deploy records — not the KB's prose summary of it. Then query only the source(s) that answer that shape of question. Loading another project's entire context to answer one fact defeats the purpose of a cheap companion.

2. Verify anything volatile before asserting it. This is not optional politeness — it is the pattern's entire value proposition. A quick-answers session that confidently repeats a stale cached fact is worse than no session at all, because it's trusted precisely because it's fast. Follow synthesis-grounding-discipline's cache-vs-truth rule: CONTEXT/REFERENCE files and prior session summaries are caches, not truth; run the verifying command for anything that could have changed (a person's schedule, a project's status, whether something shipped, a date). This workspace-management pattern exists because a stale "is being refreshed and moved to" sentence sat in a cache for two months before an agent repeated it as current — the exact failure mode this skill's speed advantage would otherwise make more likely, not less. This step's outcome — verified live, or only found in a cache — is what step 3's trailer reports; there is no separate step where confidence gets guessed after the fact.

3. Answer tersely, and carry a grounding trailer on every answer, without exception. Per synthesis-concise-messaging: the fact first, one sentence of context only if genuinely load-bearing, then one closing line naming the source and a confidence tier from synthesis-grounding-discipline's vocabulary:

   Tier	Means	Trailer example
   Verified	Confirmed via a live verifying command / tool call this turn — file re-read, live query, git log, an API or calendar read	Source: git tag -l 'v0.11*' (verified live) — Confidence: Verified
   Cached	Read from a context file, KB doc, or prior session log without re-verifying live — name the cache's own as-of/last-updated date when it carries one	Source: csa-2026-q3/CONTEXT.md, as of 2026-08-31, not re-verified this session — Confidence: Cached
   Uncertain	No direct source found; this is inference or a best guess, not an observed fact	Confidence: Uncertain — no source found; try <person/team/doc>

   A trailer is one line, not a paragraph — it names what was checked, nothing more. Never omit it: a fast answer with no stated confidence is indistinguishable from a guess, which defeats the entire pattern. Cached is not a downgrade to apologize for — plenty of quick answers are legitimately answered from a stable reference and that's fine to say plainly. What's not fine is a volatile fact (step 2's list: schedules, status, ship state, dates) answered as Cached when it should have been verified — that is a defect, not a shortcut: go verify, or say "Uncertain" and stop.

4. Log it. Append one line to resources/FAQ.md: date, question, the answer as given, the source(s) checked, and the confidence tier. This is a side effect, not extra work — it turns repeat-question friction into a growing, skimmable artifact, and it's what actually earns the name "FAQ" over time. Don't log ephemeral asks that will be false by tomorrow (exact meeting times, in-flight numbers) unless the pattern of asking is itself worth recording.

5. Route durable facts onward, don't hoard them here. If an answer surfaces something that belongs in the workspace's knowledge base (a role change, a team's charter, a standing fact about a product), that goes through the workspace's normal synthesis-knowledge-capture path — not into this project's own files as a second copy. This project's writes stay limited to its own CONTEXT.md/FAQ.md and, when the user says so, a KB capture.

Scope Boundary — What This Is Not For

Keep the mandate narrow on purpose:

• No decisions, no drafting, no sending messages, no calendar changes, no autopilot delegation. Every one of those belongs in a session with the corresponding skill loaded and the corresponding scrutiny applied.
• If a question turns out to need real investigation — multiple sessions, a plan, a deliverable — say so and hand it to its own project rather than absorbing the work here. The companion's job is triage-speed answers, not the work the answer points toward.
• Don't let this become a second inbox. It answers what's asked; it doesn't proactively surface items (that's synthesis-chief-of-staff territory) or own any cadence (that's an operations seat's job).

Relationship to Other Skills

• synthesis-project-management supplies the project itself (index entry, tiered CONTEXT.md/REFERENCE.md, cross-agent handoff so the same project works from Claude Code and Codex identically).
• synthesis-context-lifecycle governs the tiered-memory mechanics once the project exists.
• synthesis-grounding-discipline is why step 2 of the protocol is not skippable, and its cache-vs-truth vocabulary (verified vs. cached, name the layer) is exactly what step 3's confidence trailer reuses rather than inventing a parallel scheme.
• synthesis-concise-messaging shapes the answer format.
• synthesis-model-tiers supplies the routine tier recommendation and the vocabulary for stating it without attempting to switch it.
• synthesis-knowledge-capture is where durable facts actually get saved, not this skill.
• synthesis-onboarding provides the stable synthesis workspace ensure command used in Setup step 1 — this skill never scaffolds a substitute of its own.