| name | ctx-inspiration |
| description | Process new files in the global inspiration library. Updates INDEX.md, appends to or rewrites synthesis files, runs quality checks, logs outcomes by job type. Run whenever new content is added to context-library/inspiration/raw/. |
/ctx-inspiration — Inspiration Library Enrichment
Processes new files in context-library/inspiration/raw/. Keeps INDEX.md current, builds and maintains synthesis files by theme, evaluates synthesis quality, and logs how the library is used so it improves over time.
When to run
- You've added new files to
context-library/inspiration/raw/ (transcripts, articles, social posts)
- You want to check the library is up to date before starting a campaign or idea session
- You want to trigger a full synthesis for a theme after several appends
Step 1 — Find unprocessed files
Scan all subfolders of context-library/inspiration/raw/ for files where enriched: is blank or missing.
Show:
Found [N] unprocessed file(s):
transcripts/
— [filename] — [title from frontmatter or "untitled"]
articles/
— [filename]
[M] already enriched (skipping).
Process all [N], or specific ones?
Wait for confirmation. Process files one at a time — do not read all files upfront.
Step 2 — Complete frontmatter
For each file, read it and check the full schema is populated:
title:
channel:
host:
guest:
summary:
themes:
topics:
tone:
format:
audience:
project-relevance:
enriched:
For any missing required field: read the file content and fill it in. If genuinely unknowable (e.g. host of an article), use unknown. Do not ask the user for fields Claude can infer from the content.
Tag taxonomy for reference:
- themes:
AI-agents, automation, productivity, content-strategy, personal-OS, personal-knowledge-management, OpenClaw, agent-architecture, no-code-AI, startup-ideas, AI-tools
- tone:
practical, educational, inspirational, technical, conversational
- format:
interview, tutorial, explainer, essay, case-study, list
- audience:
beginners, practitioners, technical, non-technical, executives
- project-relevance: standalone project slug (
pm-os), portfolio root (ideas), or portfolio item (ideas/autonomous-agents)
Step 3 — Update INDEX.md
Read context-library/inspiration/INDEX.md.
- If the file already has a row (match on filename): update it with any frontmatter changes
- If not present: add a new row using this format:
| [filename-link] | [title] | [themes, comma-separated] | [tone] | [project-relevance] | [summary — one line] |
File in the correct section by type (## Transcripts, ## Articles, ## Social). If the section doesn't exist yet, create it.
Update total-files: in frontmatter.
Step 4 — Synthesis: evaluate before writing
For each theme in the file's themes frontmatter:
If no synthesis file exists for this theme:
This is a first-source situation. Go directly to Full Synthesis (Step 6). No evaluation needed.
If synthesis file exists:
Read context-library/inspiration/synthesis/[theme].md. Run three evaluations:
Eval 1 — Source count threshold (frontmatter read only)
Read appends-since-synthesis: from synthesis frontmatter.
- If ≥ 5: flag as
THRESHOLD_HIT
Eval 2 — Gap closure check (section read only)
Read the ## Gaps section of the synthesis.
- Check if the new file's
topics or themes frontmatter matches any listed gap.
- If yes: flag as
GAP_CLOSED (the synthesis body doesn't reflect this yet — structural drift)
Eval 3 — Contradiction signal (reasoning task)
Read the existing ## Key Patterns section and the new file's summary field.
- Ask: does any new insight make an opposing claim about the same topic as an existing pattern?
- Look for: same subject + negating language ("without", "no longer", "contrary", "unlike", "doesn't require")
- If found: flag as
CONTRADICTION_DETECTED and note which pattern(s) are in tension
Show evaluation results:
Synthesis evaluation — [theme]
Appends since last full synthesis: [N] [THRESHOLD_HIT if ≥5]
Gap closure detected: [yes/no — which gap]
Contradiction signal: [none / "New source contradicts Pattern 3: ..."]
Recommendation: [Append / Full synthesis recommended]
If any flag is raised: recommend full synthesis. Let the user decide:
One or more signals suggest a full synthesis may be needed.
Run full synthesis, or append anyway? (full / append)
Default to append unless user says full.
Step 5 — Append mode
Extract from the new file (read the full content):
- New Key Pattern(s): 1-2 distinct patterns not already present in the synthesis. Skip if the insight is already captured.
- Notable Quotes: any quotes worth preserving verbatim
- Content Angles: 1-2 specific angles this source enables (format:
Angle — Format — Relevant Campaign)
- Gap updates: remove any gaps that this source now fills; add new gaps it reveals
Append under each relevant section, with a source attribution line:
*Source: [title](../raw/transcripts/filename.md) — added [date]*
Update synthesis frontmatter:
appends-since-synthesis: [increment by 1]
sources: [add this file to the list]
Step 6 — Full synthesis mode
Read ALL source files listed in the synthesis frontmatter, plus the new file.
Rewrite the synthesis completely using this structure:
---
theme: [theme]
last-full-synthesis: [today]
appends-since-synthesis: 0
sources:
- [all source file paths]
relevant_to: [union of all project-relevance values across sources]
---
# Synthesis: [Theme]
Cross-source patterns, key ideas, and content angles. [N] sources as of [date].
## Key Patterns
[Numbered list — each pattern is a distinct, named insight. No duplicates.]
## Notable Quotes
[Verbatim quotes worth preserving, with attribution]
## Content Angles
| Angle | Format | Relevant Campaign/Project |
|-------|--------|--------------------------|
## Gaps / Open Questions
[What's still missing. What would strengthen this synthesis.]
## What to Add Next
[Types of sources that would add most value]
After rewriting: run the Quality Sub-Agent (Step 7).
Step 7 — Quality sub-agent (full synthesis only)
After a full synthesis, evaluate it against five criteria. Do not show working — show results only.
| Check | Pass/Fail | Notes |
|---|
| Coverage | ✅/⚠ | Are all listed sources represented in Key Patterns? |
| Distinctness | ✅/⚠ | Are patterns genuinely separate, or do any overlap significantly? |
| Contradictions | ✅/⚠ | Any unresolved tensions between patterns? |
| Actionability | ✅/⚠ | Are Content Angles specific enough to act on without more context? |
| Gap accuracy | ✅/⚠ | Have previously listed gaps been removed if sources now fill them? |
If any check fails: show what's wrong and offer to fix before moving on.
Step 8 — Update project links
For each entry in the file's project-relevance: frontmatter:
Portfolio item (e.g. ideas/autonomous-agents):
Portfolio root (e.g. ideas, content-engine):
- No automatic file update — the INDEX.md serves as the reference point.
Standalone project (e.g. pm-os):
- No automatic file update unless the project has a
context/inspiration-sources.md file. If it does, add the file there.
Step 9 — Mark enriched
Update the raw file's frontmatter:
enriched: [today's date]
Step 10 — Outcome log
Ask:
What job did you process this synthesis for? (or skip)
1. content-creation — writing a post, campaign, blog
2. idea-development — developing an idea, writing a hypothesis
3. technical-implementation — building a skill, automating a process
4. decision-making — choosing a direction or approach
5. research — deepening understanding of a topic
6. skip
If not skipped, ask two more:
What did the synthesis give you that helped? What was missing or harder than it needed to be?
(type "none" to skip)
Did you need to load individual source files to supplement the synthesis? (y/n)
This last question is the behavioral signal: if yes, the synthesis didn't fully serve the job — it needed the raw files to fill gaps. Log went-deeper: yes in the eval-log. Over time, patterns in went-deeper by job type reveal which synthesis sections need more depth for which use cases.
Log to context-library/inspiration/synthesis/eval-log.md:
| [date] | [synthesis theme] | [job type] | [project from .current-session] | [what helped] | [what was missing] | [went-deeper: y/n] | [eval flags triggered] |
Step 11 — Calibration check
Read eval-log.md. Count total rows.
If row count is a multiple of 10:
[N] outcomes logged. Patterns emerging:
Most used synthesis: [theme]
Most common job type: [type]
Most common "missing" complaint: [if any pattern]
Want to review evaluation thresholds? (y/n)
If yes: show current thresholds (source-count ≥5, gap closure, contradiction signal) and invite the user to adjust.
Step 12 — Summary
Inspiration library updated
File: [filename]
INDEX.md: [added / updated]
Synthesis: [theme] — [appended / full rewrite]
Quality check: [pass / N warnings] (full synthesis only)
Project links: [N updated]
Outcome logged: [job type / skipped]
Next: [most logical action — e.g. "Run /ctx-inspiration on [next unprocessed file]" or "Start content session with updated synthesis"]
Eval-log file format
context-library/inspiration/synthesis/eval-log.md
Create if missing:
---
last-calibration-check: —
total-entries: 0
---
# Inspiration Library — Outcome Log
Tracks how the library is used and whether synthesis files served the job.
Review every 10 entries to calibrate evaluation thresholds.
| Date | Synthesis | Job Type | Project | What Helped | What Was Missing | Flags Triggered |
|------|-----------|----------|---------|-------------|-----------------|-----------------|
Notes
On synthesis theme matching:
A file with themes: [AI-agents, automation] should update both synthesis/AI-agents.md and synthesis/automation.md if both exist. If only one exists, update that one and note the other is not yet initialised.
On full synthesis cost:
Reading 6 × 30k-word transcripts for a full synthesis is expensive. Prefer reading the first ~200 lines of each (the highest-signal section) unless the quality check reveals a coverage problem. Only read the full file if coverage fails.
On contradiction signal:
In v1, the contradiction check is a reasoning task — Claude reads the summary and existing patterns and asks itself if there's a direct conflict. This is imprecise but cheap. Future: use embeddings or structured claim extraction.
