synthesis-guard

/vibe-thesis:guard — Synthesis Guard

The third moment in the writing loop, after scaffold (orchestrator) and voice (voice-synthesis): catching when the prose drifts toward self-review tone.

Prerequisites

Current directory must contain a VIBE_THESIS_MARKER stanza in CLAUDE.md.

Why this exists

Academic prose has a characteristic failure mode: the work praising itself. Symptoms include:

• Inflationary adjectives (groundbreaking, comprehensive, rigorous, definitive, novel, state-of-the-art).
• Self-praise framings (we make N contributions, to the best of our knowledge no prior work, our approach uniquely, this is the first).
• Defensive over-qualification (despite limitations, our findings demonstrate, while the methodology is bounded, the results clearly show).
• Conclusions that re-assert importance instead of letting findings speak (these results have profound implications, this work opens new directions).

The third ThesisStudio pillar default is Honest Limits. Self-review tone is the inverse — and it's the dominant failure mode of LLM-assisted academic writing because the model is trained to defend its outputs.

This skill catches the pattern as a lint pass. Advisory — the user revises.

Behavior

1. Verify marker. Refuse politely if missing.

2. Parse argument.

   • standard (default): inflationary language + self-praise framings only (the highest-signal patterns, lowest false-positive rate).
   • strict: full pattern set including defensive over-qualification + conclusions-re-asserting-importance.

3. Determine body file set. Read THESIS_MODE from CLAUDE.md.

   • dissertation / masters → 03_BODY/<chapter>/*.md recursively.
   • article → 03_BODY/*.md.
   • If mode is missing or unparseable, fall back to 03_BODY/**/*.md.

4. Scan each file. For each .md file in the body set, scan line-by-line for the patterns below. Skip code blocks (lines between fenced triple-backtick blocks).

5. Report. For each finding, emit one line:

   <file>:<line>  <pattern-category>  "<quoted offending phrase>"  → <one-line suggestion>
   

6. Summary. End with: <N> findings across <M> files. Strict pattern set: <yes/no>. Pillar reference: Honest Limits (ThesisStudio default pillar #3).

7. Exit cleanly always. This is advisory; not a gate. Pre-render integration in /vibe-thesis:vibe-render --guard is what blocks renders if findings are present, not this skill itself.

Pattern table

Standard (always-scanned)

Category	Match	Suggestion
Inflationary	\b(groundbreaking|comprehensive|rigorous|definitive|novel|state-of-the-art|seminal|paradigm-shifting|cutting-edge)\b	Name what's covered or claimed instead. "Comprehensive" → list the dimensions. "Novel" → name what's new and what came before.
Self-praise framings	\b(we make \w+ contributions?|to the best of our knowledge|no prior work has|our approach uniquely|this is the first|we are the first to|this work demonstrates that)\b	Let the work demonstrate; don't announce the demonstration. Cite the comparison point if claiming first-ness.

Strict-only (additional patterns)

Category	Match	Suggestion
Defensive over-qualification	\b(despite limitations,?|while \w+ is bounded|notwithstanding|albeit|while not exhaustive)\b.{0,80}\b(demonstrate|show|prove|establish|evidence)	Either the limitation matters (then own it without demonstrating around it) or it doesn't (then drop the qualifier).
Conclusion re-assertion	(in conclusion|to summarize|in summary).{0,200}(profound|important|significant|crucial|critical|opens new|paves the way)	Conclusions should restate findings, not their importance. Let the reader weigh significance.
Hedging-with-importance	\b(arguably|perhaps|it could be said).{0,80}(important|significant|profound|notable)	If you have to argue it's significant, the work hasn't earned the claim. Strengthen the evidence or drop the importance claim.

Output examples

Clean run:

0 findings across 8 files. Strict pattern set: no.
Pillar reference: Honest Limits (ThesisStudio default pillar #3).

Findings run:

03_BODY/01-introduction.md:14   inflationary       "We propose a comprehensive framework"   → "Comprehensive" → list the dimensions covered.
03_BODY/01-introduction.md:23   self-praise        "to the best of our knowledge, no prior work has addressed"   → Either cite the comparison point or rephrase as a positive claim ("This work addresses X by ...").
03_BODY/03-methods.md:8         inflationary       "a novel approach to"   → "Novel" → name what's new and what came before.
03_BODY/05-conclusion.md:42     conclusion-reassertion (strict)   "These results have profound implications for"   → Restate the findings instead. Significance is the reader's call.

4 findings across 2 files. Strict pattern set: yes.
Pillar reference: Honest Limits (ThesisStudio default pillar #3).

Edge cases

• Body content is empty (just-scaffolded project): report 0 findings across 0 files. Body is empty — guard pass deferred until you write content.
• A pattern lives inside a code block or a quoted citation: skip it. Code blocks are detected by triple-backtick fences; quoted citations are blockquote-prefixed (>).
• A file has 50+ findings: print the first 10, then ... <N more findings in this file. Re-run with --strict for full output.
• The user has explicitly approved an inflationary phrase via a per-file pragma comment (<!-- guard: ignore inflationary -->): respect the pragma for that file. (v0.2 feature — for v0.1, all findings surface.)

Integration with /vibe-thesis:vibe-render --guard

When invoked from /vibe-thesis:vibe-render with the --guard flag, the render skill dispatches /vibe-thesis:guard standard, surfaces findings, and asks the builder to acknowledge or fix before rendering. The guard skill itself doesn't know about that integration — it just runs the lint and reports.