lyndonkl/synthesis-application

synthesis-application

Critical-evaluation gate that runs after component extraction and before final artifact construction. Asks Adler's third-level questions: "Is it true? What of it?" Catches logical gaps, missing pieces, contradictions, and practical-applicability issues before they propagate into a downstream artifact.

The skill is invoked autonomously by an agent on a structured set of extracted components. It does not host a dialogue with the operator.

Workflow

- [ ] Step 1: Completeness check — are all major component types present?
- [ ] Step 2: Logic check — do the parts cohere? Any contradictions?
- [ ] Step 3: Applicability check — can this actually be applied?
- [ ] Step 4: Gap-fill recommendations — what would the calling agent need to fill?
- [ ] Step 5: Output structured findings + a single GO / GO-WITH-GAPS / NO-GO verdict

Inputs

The calling agent passes:

• extracted_components: a structured payload of what was extracted (terms, propositions, arguments, solutions, hypotheses, etc.). Format depends on caller.
• purpose_context: what this is being extracted for. The completeness criteria depend on it. Examples:
  • purpose=skill_construction — caller is about to build a SKILL.md from these components
  • purpose=paper_pass_3_input — caller is about to do a Pass 3 deep read on these components
  • purpose=evidence_synthesis — caller is about to write a report from these components
• domain_hint: optional, the field the source is in.

Output structure

## Synthesis-and-Application Output

### Completeness check
For each major component type expected for purpose={purpose_context}:
- {Component type}: {present | partial | missing} — {one-line rationale}
- ...

### Logic check
- Gaps: {list of "the artifact jumps from A to C without explaining B" issues, or "none found"}
- Contradictions: {list of "section X contradicts section Y" issues, or "none found"}
- Hidden assumptions: {assumptions stated implicitly that should be made explicit, or "none flagged"}
- Unsupported claims: {claims without evidence or reasoning, or "none flagged"}

### Applicability check
- Concrete enough to act on: {yes | partial | no} — {rationale}
- Decision criteria specified where needed: {yes | partial | no} — {rationale}
- Edge cases covered: {yes | partial | no} — {which edges are unaddressed}

### Gap-fill recommendations
- {Specific item the caller should add or seek before downstream commitment}
- ...

### Verdict
{GO | GO_WITH_GAPS | NO_GO}

Rationale: {2-3 sentences}

Component types by purpose_context

Different downstream artifacts need different inventories of components. The completeness check uses the relevant inventory.

purpose = skill_construction

A SKILL.md needs:

• Terms: key concepts defined unambiguously
• Propositions: claims supported by evidence or reasoning
• Arguments: complete logical sequences from premise to conclusion
• Solutions: concrete examples or templates that demonstrate application
• Decision criteria: how to choose between alternatives where the methodology branches
• Triggers: when this skill should be invoked

purpose = paper_pass_3_input

A deep-read needs:

• Main argument: the through-line, in 3-5 sentences
• Big Question: the field-level problem
• Hypotheses: specific claims being tested
• Evidence inventory: figures, tables, references that carry the argument
• Confusions: what didn't land in Pass 2 — the agenda for Pass 3

purpose = evidence_synthesis

A report needs:

• Claims: structured propositions
• Sources: provenance for each claim
• Source quality: assessment per source
• Conflicting evidence: where sources disagree
• Confidence ratings: per-claim

For other purpose_context values, the calling agent should specify expected component types in the input — this skill does not silently extend the inventory.

Logic checks — what to look for

Gaps — premises missing between conclusions:

• "The document jumps from A to C without explaining B."
• "Step 5 references something Step 4 doesn't establish."

Contradictions — internal conflicts:

• "Proposition 2 says X, but Proposition 7 says not-X."
• "The framework's Dimension 1 and Dimension 3 overlap in unstated ways."

Hidden assumptions — load-bearing premises that aren't surfaced:

• "The methodology assumes the user has access to PHI without saying so."
• "The framework only works at scale > 10K users — never stated."

Unsupported claims — assertions without backing:

• "Claims X works 80% of the time but cites no source."

The output is the list. The skill does not fix gaps — it surfaces them and recommends action.

Applicability checks — what to look for

• Concrete enough to act on: an extracted "principle" like "always be data-driven" is too vague. Concrete would be "before each sprint, pull last month's metric trend and identify one delta to investigate."
• Decision criteria specified: when the methodology branches ("if X, do A; if Y, do B"), is X-vs-Y specified clearly enough that a different agent could follow it?
• Edge cases covered: what happens at the boundaries — empty input, conflicting input, error states? If silent on edges, flag.

Common patterns

Pattern A — Skill creation gate

purpose=skill_construction. Run after Step 3 (component extraction) and before Step 5 (skill construction) in a skill-creation workflow. The verdict gates whether the agent proceeds to build SKILL.md or returns to the source for more extraction.

Pattern B — Paper Pass 3 readiness

purpose=paper_pass_3_input. Run after Pass 2 (content grasp) on a paper extraction. The verdict gates whether Pass 3 is worth running or whether the paper needs re-reading at Pass 2 first.

Pattern C — Pre-synthesis evidence audit

purpose=evidence_synthesis. Run on a research-claim-map output before writing the synthesis report. Catches "the gathered evidence doesn't actually support the conclusion you're about to write."

Guardrails

1. Don't fix gaps; surface them. This skill produces a structured list of issues and a verdict. It does not modify the extracted components. The calling agent decides what to do.
2. Don't lower the bar on completeness when components are sparse. GO_WITH_GAPS means "you can proceed but here's what's missing" — not "this is fine because we don't have more."
3. Verdict matters — don't soften it. GO / GO_WITH_GAPS / NO_GO is a discrete signal. If you mean NO_GO, write NO_GO. The calling agent's pipeline depends on it.
4. Don't import information from outside the source. Completeness is judged against what was extracted, not what the model knows about the topic.

Related

• inspectional-reading — the first reading level, run before extraction begins.
• structural-analysis — the second level, runs between inspectional and component extraction.
• research-claim-map — for the evidence-synthesis purpose, the upstream skill that produces the structured claim-source-quality payload this skill evaluates.
• negative-contrastive-framing — pairs naturally with the "what is not said" portion of the logic check.
• The skill-creator skill at skills/skill-creator/SKILL.md invokes this skill as its Step 4.
• paper-three-pass-extraction invokes this skill before Pass 3 to gate whether a deep read is worth it.