Research Synthesis

research-synthesis

Consolidate mixed discovery evidence into one durable, discovery-ready synthesis: surface converging themes, contradictions, and gaps — then, when jobs exist, judge what the evidence says about each one, without overstating certainty. This is the PM's digest-the-mess machine: real client input is spectacular but chaotic, and downstream agents can only be as good as how well it is organised for them.

This skill consumes job knowledge; it never mints or rewrites it. jobs-to-be-done is the single path for creating, refining, or superseding jobs — even when the evidence on the table screams for a rewrite.

Framing: jobs when they exist, greenfield when they don't

• Jobs exist (docs/product/jtbd/ has active jobs): they are the required framing. Read them; preserve their wording verbatim. The synthesis renders a job verdict per active job in scope.
• Greenfield (no jobs yet — day-one discovery): synthesize anyway. The primary output becomes converging themes and raw new-job signals, and the synthesis routes straight to jobs-to-be-done to mint the first jobs from them. Guard the drift risk explicitly: a greenfield synthesis organises evidence for job creation — it does not quietly become the product's list of needs.

Where it writes

• docs/product/research/<YYYY-MM-DD>-<slug>.md — one append-only file per synthesis run, in the target repo (repo-first-and-always; a Cowork folder is a valid target, never a blocker). Never edit or delete a previous run — a superseding synthesis says so in its own header and points back. This is the stable citation target: PRDs, jobs, and plans cite research/2026-07-24-onboarding forever.
• docs/product/context.md Digest — a 2–4 line summary of the run plus a pointer to the research file, so the one file every agent reads stays current without bloating.
• Evidence not yet in docs/product/input/ that you pulled from a connector (Notion, Granola, Slack, ops API) is staged verbatim first as input/<YYYY-MM-DD>-<source>-pulled/, then synthesized from there — per the contract, skills never edit or delete batches. Pasted evidence works the same; never block on a connector.
• No target repo/folder at all → deliver the synthesis conversationally and offer to save; no durable claims.

Inputs

• The research question, opportunity area, or decision the synthesis supports.
• docs/product/jtbd/ (framing, when present) and context.md (glossary — use its settled terms; flag term drift you notice, don't silently coin new ones).
• Evidence: un-ingested input/ batches, pasted notes, staged connector pulls.
• Optional: hypotheses or segment assumptions worth testing. (No personas as framing — the A-Team works situation-and-struggle, not demographics.)

Outputs — the synthesis file

Use references/research_synthesis_input_template.md (same section order): research question · input jobs (ids + verbatim headlines) · source inventory · evidence-strength notes · converging themes · contradictions · gaps · frequency × impact matrix · job verdicts · new-job signals · discovery questions · handoff. Plus, A-Team-specific:

• sources: frontmatter tracing every batch/file the run consumed, so a reviewer can audit what you were told versus what you inferred.
• Job verdict per in-scope active job — supports / challenges / suggests refinement — citing the specific evidence. The job files stay untouched: verdicts are routed to jobs-to-be-done (REVIEW), which owns any confidence change or superseding rewrite.
• New-job signals: recurring struggles, workarounds, switching behavior that fit no input job. Recorded raw, with source and why they smell like a different job — explicitly not job statements.

Workflow

Use references/research-synthesis-framework.md for the full method. The shape:

1. Frame. Confirm the research question. Load active jobs (or declare greenfield mode out loud). Load the glossary.
2. Gather & stage. Collect un-ingested input/ batches and pasted evidence; stage any connector pulls verbatim into input/ first.
3. Normalize. Record source type, segment, timing, and caveats per source before combining, so strength labels stay honest.
4. Theme. Group repeated findings into converging themes, keeping the supporting evidence visible instead of flattening it away.
5. Contradict. Call out contradictions, segment differences, and outliers explicitly; never smooth them into false agreement — they become discovery questions and ledger entries.
6. Label strength qualitatively (Strong / Moderate / Directional / Unknown) and explain each label. No numeric scores; if numbers are absent, do not fabricate them.
7. Prioritize on the frequency × impact matrix — qualitatively. High-frequency + high-impact leads; low-frequency + high-impact stays attached to its segment; high-frequency + low-impact is quality-of-life; low-frequency + low-impact is noted and deprioritized.
8. Judge (jobs mode): render each job verdict with its evidence. Keep job wording untouched.
9. Signal (both modes): collect raw new-job signals.
10. Draft the synthesis file in template order, ending with the discovery questions that follow from contradictions and gaps.
11. Read-back — mandatory before writing. The research file and the context.md digest note are durable: present both for correction in this session, then write and commit (one commit, message naming what the run found — e.g. docs(research): onboarding synthesis — 03 challenged, 2 new signals).
12. Route. Challenged/refinement verdicts and new-job signals → jobs-to-be-done. All-supporting verdicts → discovery-plan (unknowns into a plan) or prd-writer (scope is concrete). Close with Next likely skill(s) · What to pass forward · Suggested next prompts.

No human present

Synthesis itself needs no grilling, but the read-back gate does need a human. With nobody present: produce the full synthesis in your report only, write nothing durable, and state that the run awaits review — an escalation is a defined output, not a failure. Never commit an unreviewed synthesis: it would become audit-grade evidence nobody audited.

Examples

• Input: examples/example-input.md · Output: examples/example-output.md