| name | synthesize |
| description | Use when distilling the through-line gist of one or more sources — the spine, argument, tension, or recurring frame running through a set of documents, notes, research, or transcripts, OR across the ideas within a single rich piece — into a few concise paragraphs. Triggers: "synthesize", "what's the through-line/gist", "extract the insight", "pull these together". Not for faithful summary or condensation that covers what a source says, nor for comparisons or catalogs where enumeration is the deliverable. |
Synthesize: Distill a Spine-Led Gist from One or More Sources
Distill the through-line of one or more sources — a set of documents, or the ideas within a single rich piece — into a short, spine-led gist: a few tight paragraphs built around a spine drawn from the material, where support is insight on its own and the reader is pointed to the source, not served a copy.
Inform the user when invoking this skill by name: synthesize.
Two modal failure modes pull against this and both reassert at draft time, even after a spine is chosen — defeating them is the whole job, not a one-time setup step:
- Listicle — an enumerated march through themes; the model's reflex for "synthesize."
- Fluff — padded coverage that reproduces sources instead of distilling them.
A spine is an internal organizing frame — a constraint, a claim, a recurring metaphor, or a tension — that the material carries: one source among several states it and the rest support it, or one thread runs through the ideas of a single piece.
It must come from the material, not be imposed over it (stages of grief, a narrative arc).
See references/finding-the-spine.md.
When Not to Use
- Faithful summary or condensation — covering what a source says, proportionally, with no through-line.
That is summary, not synthesis, whether the input is one source or many.
- Comparisons or catalogs where enumeration is the deliverable (vendor capability matrix, feature comparison).
The list is correct; do not force prose.
This is the escape hatch — name it and step aside.
- Line-editing or polishing prose — use
good-prose, electively, after the gist exists.
Core Discipline
- Gist, don't cover.
The output is at most a few paragraphs — the more concise the better.
Comprehensiveness is the enemy; leave out everything the spine doesn't need.
The gist is what survives ruthless trimming.
- Support is insight, not reproduction.
Every sentence of support must carry insight on its own; point the reader to the source rather than reproducing it.
But surface the gem: a quote, table, or figure that is itself distilled insight — high-signal, stands alone, serves the spine — is the support, not coverage creep; reproduce it.
The test: does reproducing it carry the insight, or merely relay content the reader could skim for?
- The spine carries the argument; sources serve the spine, not the reverse.
- Each source earns its place under the spine; those that don't fit are named as exclusions, not crammed in.
- Self-interested sources don't get the last word.
When a source is promoting its own result, product, or position (vendor PR, a lab announcing its own work, marketing), offer a grounding pass before committing the spine — external context to restore the caveats it downplays.
Surface the offer; fold outside context in only when flagged and grounded, never silently.
External context enters only through this gated pass — not a license to pad with background.
See
references/finding-the-spine.md.
- Re-grounding old wisdom in new rationale is legitimate synthesis — don't prune it.
Naming the rationale shift is the synthesis; the shift is load-bearing, not the practice's novelty.
- Strip affect that doesn't serve the argument; keep affect that does.
Workflow
- Intent gate — narrated checkpoint.
Gist/through-line, faithful summary, or enumeration?
Faithful summary → not this skill.
Enumeration is the deliverable → step aside (escape hatch).
User supplied a spine → short-circuit to 1b.
- Spine scan — narrated.
- 1a DISTILL (no spine given) → 1–3 candidate spines from the material (across sources, or the ideas within one).
- 1b VALIDATE (spine given) → test it; verdict may be holds / holds-but-sharpen / amend / doesn't-hold + alternative.
- Self-interest check → source promoting its own result/product/position?
Flag it; offer a grounding pass before step 3 (see reference).
- Material gate — GATE.
No spine emerged (material doesn't rhyme)?
Do not force prose.
Offer: synthesizer-POV spine (flagged as a different move) OR hand back.
- Spine selection — GATE (always).
Confirm the one ("spine = X — good?") · pick among several · re-pick after a validate-reject (user's spine vs the alternative).
- Build — narrated, reversible.
A few tight paragraphs on the spine.
Support is insight, not reproduction; point to sources.
Name exclusions; re-ground old wisdom; cut fluff.
- Audit & repair — narrated, reversible.
Check the draft against the failure catalog; repair in place.
Steps 1b (validation with teeth) and 2 (the no-spine branch) are where this skill earns its keep.
Read references/finding-the-spine.md before running steps 1–3 — the validation verdicts and the no-spine branch live there; running them from memory loses the teeth.
Steering Contract
The user always drives — they can override at any fork, guaranteed by transparency and reversibility, not by the agent stopping for sign-off at every step.
The kind tag on each workflow step marks which is which: GATE steps stop for the user before acting; the rest narrate and proceed, reversible by the user's next turn.
Surface every consequential choice (the spine, the no-spine fallback, the grounding-pass offer) before acting on it; never commit to a spine silently.
The spine is the single most expensive thing to get wrong — everything downstream hangs off it — so step 3 gates even when there is one obvious candidate.
Audit Catalog (Step 5)
Read references/audit-failure-modes.md for detection cues and fixes; repair each in place.
The list below is an index, not the procedure — the cues that catch each mode live in the reference.
- Listicle relapse — drifted back to enumerated themes/practices.
- Coverage creep / fluff — sprawled into comprehensive summary, or padded; support reproduces sources instead of distilling insight.
- Abandoned framing — an organizing image in the intro that doesn't carry through.
- Quote-as-spine — a source's quote doing the argumentative work the synthesizer should do.
- Significance inflation — "striking," "the strongest convergence," superlatives standing in for argument.
- Source taken at face value — a self-interested source's spin survived into the gist; caveats it downplays are missing.
- Forced-fit spine — material crammed under a spine it doesn't support (watch this after validate mode).
- Over-pruned wisdom — an old practice dropped instead of re-grounded in its new rationale.
Composition With Other Skills
Standalone — no built-in gather or polish steps, and no automatic handoffs.
The user or agent may electively run mcp-research (gather sources) before, or good-prose (polish prose) after.
References
references/finding-the-spine.md — distill vs. validate modes, validation with teeth, the no-spine branch, source-earns-its-place.
references/audit-failure-modes.md — each failure mode with detection cues and repairs.