Synthesis

Turn discovery evidence into durable artifacts the team—and agents—can load before locking a feeling north star. Synthesis is the gate: listening without writing changes nothing; writing without evidence invents customers.

What it is

Synthesis clusters episodes, promotes or kills proto-personas, and writes Ideal Customer Profile (ICP), need states, workflow, and voice of the customer (VoC) into project docs—then leaves ephemeral notes behind. Until that pack lands in a DocSlime-shaped docs/ tree, strategies and Tools, Techniques, or Practices (TTPs) are guessing with confidence.

This is not a second product narrative in .productfeeling/. Reviews and raw sessions stay local; truth lives in docs/strategy/, docs/experience/, and when appropriate docs/PRODUCT.md / DESIGN.md. A slide deck outside the repo is not a source of truth for agents.

flowchart LR
  Sessions[Session notes] --> Cluster[Cluster episodes]
  Cluster --> Decide{Evidence?}
  Decide -->|yes| Docs[docs/ synthesis pack]
  Decide -->|no| Kill[Kill or keep proto]
  Docs --> North[Feeling north star]
  Docs --> Agents[Agent loadable context]

Why it works

Scattered interview notes do not change onboarding copy or paywall trust. Named files with owners, evidence links, and a last-validated date do. Humans forget; agents invent. Both failure modes shrink when “who is this for, what do they fear, what is the current alternative?” is answerable from the repo without a chat reconstruction.

Synthesis also prevents “we talked to users” theatre. If a discovery week ends with no durable file changed, discovery did not finish—regardless of how many coffees were booked. Hypothesis-driven work (Blank) ends in a decision: promote the claim, revise it, or kill it. Compliments and future-tense enthusiasm do not graduate.

Going deeper

How to synthesise without turning transcripts into a second novel:

1. Cluster before you write. Group episodes by need state, segment, and job—not by interviewee name. Patterns that survive three independent stories become candidates for personas and north-star ownership; one-offs stay footnotes.
2. Promote with citations; proto everything else. A claim in docs/ should point at interviews or behavioural sources and carry a last validated date. Assumptions keep an explicit proto- label—or get deleted. Designing on unlabeled guesses is how fiction becomes roadmap.
3. Separate ephemeral from durable. Session notes, audio, and scratch capture cards may live in .productfeeling/sessions/. ICP, personas, journey, workflow, and language graduate to docs/. Do not dump full transcripts into the product narrative.
4. Write the minimum pack that agents can load. Typically: ICP + buyer/user split in strategy; need states and workflow/journey in experience; language and voice cues in PRODUCT/DESIGN. Enough that /productfeeling persona and jobs do not invent Maya.
5. Gate the north star. Lock Feeling North Star only when an owner persona and their fear/competence stakes are evidenced—not when a workshop produced a vibe word. Optionally stress-test ICP assumptions with RedTeam before strategy lock.

For builders and agents

Synthesis is an integration point for the skill. Run /productfeeling sequence (Customer discovery) ending in handoff, or /productfeeling init then /productfeeling handoff when refreshing a thin docs tree. Prefer DocSlime fill/kiss when scaffolds still contain prompts; do not duplicate truth into skill-local files when docs/ exists.

For agents summarising discovery: refuse to invent segments; quote or cite; label uncertainty. For humans: put owners on files so freshness is someone’s job. Instrument the meta-metric—days since ICP last validated—and treat stale personas as a product risk equal to stale code dependencies.

Impeccable craft comes after synthesis. Polishing surfaces for an unwritten customer is expensive fiction.

When to use it

• After a round of interviews (or continuous discovery cadence)
• Before locking Feeling North Star or running a strategy playbook
• When proto-personas have enough evidence to promote or delete
• When agents keep inventing customers because nothing is in docs/

Do

• Promote claims with evidence; keep proto- labels on the rest—or delete them
• Write a short synthesis pack to agreed paths (see Durable targets below)
• Link claims to interviews or behavioural sources; date last validated
• Separate: session notes (ephemeral) vs product/experience truth (durable)
• Run /productfeeling handoff (DocSlime) when the docs tree is scaffold-thin—fill/kiss as needed
• Optionally stress-test ICP assumptions with RedTeam before locking strategy

Don’t

• Dump interview transcripts into docs/ as the narrative
• Duplicate ICP/journey truth into .productfeeling/ when docs/ exists
• Lock a north star on proto-personas alone
• Skip synthesis and jump straight to Impeccable craft
• Treat a slide deck outside the repo as the SoT for agents

Founder Tip

If an agent cannot answer “who is this for and what do they fear?” from docs/, synthesis is not done.

Make It Yours

1. Kill list — proto claims to delete this week.
2. File map — which discovery outputs land in which docs/ paths.
3. Evidence links — one citation per major persona and need-state claim.
4. Gate check — north star and next strategy blocked until synthesis pack exists.

Insights & Metrics

1. Graduation rate — Validated claims written to docs/ ÷ Claims still only in sessions/chat.
2. Freshness — Days since ICP / primary persona last validated (set a team threshold).
3. Agent loadability — Can context.mjs / selective docs load name owner + fear + current alternative without chat invention?

Behind the Data

• What is still proto that you are already designing as fact?
• Where does truth live today—repo, Notion, or someone’s head?
• Did the last “discovery week” change a single durable file?

Related concepts

• Feeling North Star, Jobs-to-be-Done
• Inputs: 01–05; outputs feed all Strategies

Further reading

• Customer Development is Not a Focus Group (Steve Blank) — Hypotheses in, decisions out.
• The Mom Test (Rob Fitzpatrick) — What counts as evidence worth writing down.
• DocSlime — Opinionated docs/ tree for durable product context.

Agent skill

• Primary command / sequence: /productfeeling sequence — Customer discovery playbook ending in handoff; or /productfeeling init then /productfeeling handoff when refreshing docs
• Related commands: /productfeeling persona, /productfeeling jobs, /productfeeling brief (only after north star is grounded)
• When the agent should load this page: “synthesis”, “write up discovery”, “update personas in docs”, “graduate findings”, “before north star”
• Companion handoff: DocSlime — primary: write/update the synthesis pack in docs/; fill/kiss if scaffold-thin. Impeccable — only after north star / emotion brief. RedTeam — /redteam assumptions on ICP before locking. No external discovery skill.
• Feeling north star this practice serves: grounded clarity that survives the next session and the next agent
• Anti-goals: transcript dumps as docs, dual SoT in .productfeeling/, craft before synthesis
• Reference path: skill/reference/sequence.md
• Durable DocSlime targets:
  • docs/strategy/ — ICP, personas, market/buy notes
  • docs/experience/ — need states, workflow, journeys
  • docs/PRODUCT.md / DESIGN.md — voice, language, constraints
  • Ephemeral: .productfeeling/sessions/, .productfeeling/reviews/ — not a second narrative