Compare commits
3 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
 ejlewis
|
ef41b6b30a | ||
|
ejlewis
|
3eff77aa1a | ||
|
ejlewis
|
43d708c834 |
@@ -0,0 +1,230 @@
|
||||
# Tag Enhancement Implementation Plan
|
||||
|
||||
> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
|
||||
|
||||
**Goal:** Stop the tagger from force-fitting taxonomy tags onto notes that don't match (e.g., tagging a memoir as "productivity"), by seeding a personal-narrative cluster into the taxonomy and rewriting the prompt to permit zero-taxonomy-tag output when nothing fits.
|
||||
|
||||
**Architecture:** Single-script tool; all changes land in `tag-notes.py` (system prompt + tag cap) and `tag-taxonomy.yaml` (new cluster). No new files, no new dependencies, no test harness added. Verification is a manual re-run against a known problem note.
|
||||
|
||||
**Tech Stack:** Python, PyYAML, ruamel.yaml, local LM Studio (OpenAI-compatible) endpoint.
|
||||
|
||||
**Note on testing:** This project has no test infrastructure (per `CLAUDE.md`: "There are no tests, linter, or build step"). The tagger's correctness is judged by LLM output quality against real notes, not unit tests. Each task below ends in a manual spot-check where meaningful; end-to-end verification lives in Task 4.
|
||||
|
||||
**Spec:** `docs/superpowers/specs/2026-04-19-tag-enhancement-design.md`
|
||||
|
||||
---
|
||||
|
||||
## Task 1: Add Personal Narrative cluster to taxonomy
|
||||
|
||||
**Files:**
|
||||
- Modify: `tag-taxonomy.yaml` (append at end of file)
|
||||
|
||||
- [ ] **Step 1: Append the new cluster**
|
||||
|
||||
Append these lines to the end of `tag-taxonomy.yaml` (there is currently a `# Personal Interests` cluster ending with `- gardening`; add a blank line after `gardening`, then this block):
|
||||
|
||||
```yaml
|
||||
|
||||
# Personal Narrative & Life
|
||||
- memoir
|
||||
- personal-essay
|
||||
- reflection
|
||||
- family
|
||||
- parenting
|
||||
- recovery
|
||||
- mental-health
|
||||
- aging
|
||||
- relationships
|
||||
- childhood
|
||||
- identity
|
||||
```
|
||||
|
||||
- [ ] **Step 2: Verify the YAML still parses**
|
||||
|
||||
Run: `python3 -c "import yaml; print(len(yaml.safe_load(open('tag-taxonomy.yaml'))['tags']))"`
|
||||
|
||||
Expected: prints an integer equal to the old count + 11 (the old file had 31 tags, so expect `42`). If the number isn't old-count + 11, the YAML is malformed — fix indentation before moving on.
|
||||
|
||||
- [ ] **Step 3: Commit**
|
||||
|
||||
```bash
|
||||
git add tag-taxonomy.yaml
|
||||
git commit -m "Add personal-narrative cluster to tag taxonomy"
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Task 2: Rewrite the system prompt in `request_metadata`
|
||||
|
||||
**Files:**
|
||||
- Modify: `tag-notes.py:127-145` (the `system_prompt` f-string inside `request_metadata`)
|
||||
|
||||
**Context:** Current prompt forces 1-5 taxonomy tags and tells the LLM to be "conservative" about new suggestions. We're flipping both: allow 0 taxonomy tags when nothing fits, and let new suggestions go up to 5.
|
||||
|
||||
- [ ] **Step 1: Replace the system_prompt f-string**
|
||||
|
||||
In `tag-notes.py`, find the current `system_prompt` assignment inside `request_metadata` (starts at line 127):
|
||||
|
||||
```python
|
||||
system_prompt = f"""You analyze markdown notes and return structured metadata.
|
||||
|
||||
Return ONLY valid JSON in this exact shape:
|
||||
{{
|
||||
"tags_from_taxonomy": ["tag1", "tag2"],
|
||||
"new_tag_suggestions": ["newtag1"],
|
||||
"seo_title_suffix": "Short descriptor that will follow the note title",
|
||||
"seo_description": "Factual summary between {SEO_DESC_MIN} and {SEO_DESC_MAX} characters.",
|
||||
"seo_keywords": ["keyword1", "keyword2"]
|
||||
}}
|
||||
|
||||
Rules:
|
||||
- tags_from_taxonomy: 1-5 tags drawn from the existing taxonomy that best fit the content.
|
||||
- new_tag_suggestions: 0-2 NEW tags, only when content truly warrants it (be conservative).
|
||||
- seo_title_suffix: a short, clean, non-clickbaity descriptor of the note. Do NOT include the note title or a leading colon — only the text that would follow "<title>: ". Aim for 4-10 words.
|
||||
- seo_description: a clean factual summary, STRICTLY between {SEO_DESC_MIN} and {SEO_DESC_MAX} characters inclusive. Count characters carefully before responding.
|
||||
- seo_keywords: 10-15 relevant keywords, no duplicates.
|
||||
|
||||
Existing tag taxonomy: {taxonomy_str}"""
|
||||
```
|
||||
|
||||
Replace it with:
|
||||
|
||||
```python
|
||||
system_prompt = f"""You analyze markdown notes and return structured metadata.
|
||||
|
||||
Return ONLY valid JSON in this exact shape:
|
||||
{{
|
||||
"tags_from_taxonomy": ["tag1", "tag2"],
|
||||
"new_tag_suggestions": ["newtag1"],
|
||||
"seo_title_suffix": "Short descriptor that will follow the note title",
|
||||
"seo_description": "Factual summary between {SEO_DESC_MIN} and {SEO_DESC_MAX} characters.",
|
||||
"seo_keywords": ["keyword1", "keyword2"]
|
||||
}}
|
||||
|
||||
Rules:
|
||||
- Tags should describe what the note is substantively about, not topics it merely mentions in passing.
|
||||
- tags_from_taxonomy: 0-5 tags drawn from the existing taxonomy, ONLY when they genuinely fit. Do NOT force a taxonomy tag — return an empty list if nothing truly applies.
|
||||
- new_tag_suggestions: 0-5 NEW tags when the taxonomy doesn't adequately cover the content. Each must be a reusable category (not hyper-specific to one note). Use lowercase-hyphenated style (e.g., personal-essay).
|
||||
- seo_title_suffix: a short, clean, non-clickbaity descriptor of the note. Do NOT include the note title or a leading colon — only the text that would follow "<title>: ". Aim for 4-10 words.
|
||||
- seo_description: a clean factual summary, STRICTLY between {SEO_DESC_MIN} and {SEO_DESC_MAX} characters inclusive. Count characters carefully before responding.
|
||||
- seo_keywords: 10-15 relevant keywords, no duplicates.
|
||||
|
||||
Existing tag taxonomy: {taxonomy_str}"""
|
||||
```
|
||||
|
||||
The three substantive changes:
|
||||
1. Added philosophy line: `- Tags should describe what the note is substantively about, not topics it merely mentions in passing.`
|
||||
2. `tags_from_taxonomy: 1-5 ... best fit the content.` → `0-5 ... ONLY when they genuinely fit. Do NOT force a taxonomy tag — return an empty list if nothing truly applies.`
|
||||
3. `new_tag_suggestions: 0-2 ... be conservative).` → `0-5 NEW tags when the taxonomy doesn't adequately cover the content. Each must be a reusable category (not hyper-specific to one note). Use lowercase-hyphenated style (e.g., personal-essay).`
|
||||
|
||||
- [ ] **Step 2: Syntax check the module**
|
||||
|
||||
Run: `python3 -c "import ast; ast.parse(open('tag-notes.py').read()); print('ok')"`
|
||||
|
||||
Expected: prints `ok`. If it prints a SyntaxError, the f-string braces or quotes are wrong — fix before moving on.
|
||||
|
||||
- [ ] **Step 3: Commit**
|
||||
|
||||
```bash
|
||||
git add tag-notes.py
|
||||
git commit -m "Allow zero-taxonomy-tag output and more new tag suggestions"
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Task 3: Raise the combined tag cap from 5 to 8
|
||||
|
||||
**Files:**
|
||||
- Modify: `tag-notes.py:225` (inside `process_note`, in the `if needs_tags:` block)
|
||||
|
||||
- [ ] **Step 1: Change the slice**
|
||||
|
||||
Find this line in `tag-notes.py` (inside `process_note`, ~line 225):
|
||||
|
||||
```python
|
||||
combined = list(dict.fromkeys(list(taxonomy_tags) + list(new_suggestions)))[:5]
|
||||
```
|
||||
|
||||
Change to:
|
||||
|
||||
```python
|
||||
combined = list(dict.fromkeys(list(taxonomy_tags) + list(new_suggestions)))[:8]
|
||||
```
|
||||
|
||||
- [ ] **Step 2: Syntax check**
|
||||
|
||||
Run: `python3 -c "import ast; ast.parse(open('tag-notes.py').read()); print('ok')"`
|
||||
|
||||
Expected: prints `ok`.
|
||||
|
||||
- [ ] **Step 3: Commit**
|
||||
|
||||
```bash
|
||||
git add tag-notes.py
|
||||
git commit -m "Raise combined tag cap from 5 to 8"
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Task 4: End-to-end verification against the problem note
|
||||
|
||||
**Files:**
|
||||
- Modify (temporarily): the "Becoming a Morning Person" note in `~/Documents/ejl-zk/40 Public/41 Notes/`
|
||||
|
||||
**Context:** The tagger only touches fields that are currently empty. To force reprocessing we clear the `tags:` field on the problem note, run the script, and inspect the result.
|
||||
|
||||
- [ ] **Step 1: Locate the note**
|
||||
|
||||
Run: `find ~/Documents/ejl-zk/40\ Public/41\ Notes/ -iname "*morning*person*.md"`
|
||||
|
||||
Expected: prints one file path. Note it for the following steps (referred to below as `$NOTE`).
|
||||
|
||||
- [ ] **Step 2: Confirm LM Studio is up**
|
||||
|
||||
Run: `curl -s http://localhost:1234/v1/models | head -c 200`
|
||||
|
||||
Expected: JSON response listing at least one model, including `openai/gpt-oss-20b` (the value of `MODEL_NAME`). If the request fails, start LM Studio and load the model before continuing.
|
||||
|
||||
- [ ] **Step 3: Clear the existing tags field**
|
||||
|
||||
Open `$NOTE` in an editor and set the `tags:` frontmatter value to an empty list (`tags: []`) or delete the line entirely. Save. Do NOT clear other fields — the script won't touch already-populated ones, which is the desired isolation.
|
||||
|
||||
- [ ] **Step 4: Run the tagger**
|
||||
|
||||
Run: `cd ~/bin/note-tagger && ./tag-notes.py`
|
||||
|
||||
Expected: the script processes every note; for the Morning Person note it prints a line like ` + Added tags: memoir, personal-essay, family, recovery, aging, ...` and does NOT print `productivity` or `learning` in that line.
|
||||
|
||||
- [ ] **Step 5: Inspect the frontmatter**
|
||||
|
||||
Open `$NOTE` and inspect the `tags:` block.
|
||||
|
||||
Pass criteria:
|
||||
- Contains at least 3 of: `memoir`, `personal-essay`, `reflection`, `family`, `parenting`, `recovery`, `aging`, `relationships`, `childhood`, `identity`.
|
||||
- Does NOT contain `productivity` or `learning`.
|
||||
- Has ≤ 8 tags total (enforces the Task 3 cap).
|
||||
|
||||
Fail criteria (any triggers a rethink — do NOT patch around it):
|
||||
- Still includes `productivity` or `learning`.
|
||||
- Zero tags written.
|
||||
- More than 8 tags.
|
||||
|
||||
If it fails: the follow-up noted in the spec is to add a single few-shot example to the prompt. Stop and report the failing output; don't silently escalate to that change.
|
||||
|
||||
- [ ] **Step 6: Spot-check one other already-tagged note**
|
||||
|
||||
Run: `git status` in the notes vault (if it's a git repo) OR simply open one other note that already had tags before this run. Confirm its `tags:` field was NOT modified (the script's "only touch empty fields" invariant must still hold).
|
||||
|
||||
Expected: no change to that note's tags. If there IS a change, the empty-check logic regressed — stop and investigate.
|
||||
|
||||
- [ ] **Step 7: No commit for this task**
|
||||
|
||||
Task 4 is verification only. Any prompt-tuning follow-up is out of scope for this plan.
|
||||
|
||||
---
|
||||
|
||||
## Out of scope (from spec)
|
||||
|
||||
- Two-pass LLM classification.
|
||||
- Few-shot examples in the prompt (follow-up candidate only if Task 4 fails).
|
||||
- Changes to `seo_*` fields, `CONTENT_CHAR_LIMIT`, retry flow, slug derivation, or YAML round-tripping.
|
||||
@@ -0,0 +1,97 @@
|
||||
# Tag Enhancement Design
|
||||
|
||||
Date: 2026-04-19
|
||||
|
||||
## Problem
|
||||
|
||||
The tagger is producing wrong tags for personal-narrative content. Concrete example: the essay "Becoming a Morning Person" — a memoir about life stages, family, recovery, aging, and parenting — was tagged `productivity` and `learning`.
|
||||
|
||||
Root cause is twofold:
|
||||
|
||||
1. **Taxonomy gap.** `tag-taxonomy.yaml` has no categories that cover personal narrative, memoir, family, or reflection. The closest fits the LLM can find are productivity-adjacent tags from the "Knowledge & Learning" cluster.
|
||||
2. **Prompt bias.** The system prompt in `request_metadata` (tag-notes.py:127) requires 1-5 taxonomy tags (no zero option) and tells the LLM to be "conservative" about new tag suggestions (0-2 max). Together these force the model to pick taxonomy tags even when none genuinely apply, and discourage it from proposing the new categories that would better describe the content.
|
||||
|
||||
## Goals
|
||||
|
||||
- The "Becoming a Morning Person" essay should be tagged with memoir/personal-narrative concepts, not productivity/learning.
|
||||
- Future notes that fall outside the current taxonomy should surface new tag suggestions rather than force-fit existing ones.
|
||||
- Taxonomy can grow over time via the existing `new_tag_accumulator` → end-of-run prompt flow — no change to that mechanism.
|
||||
|
||||
## Non-Goals
|
||||
|
||||
- No two-pass LLM classification. Single call per note stays.
|
||||
- No few-shot examples in the prompt for this iteration. May be added as a follow-up if the 20B local model underperforms on the rewritten prompt.
|
||||
- No change to `seo_title_suffix`, `seo_description`, `seo_keywords`, `CONTENT_CHAR_LIMIT`, the SEO-description retry flow, YAML round-tripping, or the slug derivation.
|
||||
|
||||
## Changes
|
||||
|
||||
### 1. Taxonomy additions (`tag-taxonomy.yaml`)
|
||||
|
||||
Add a new cluster at the end of the file:
|
||||
|
||||
```yaml
|
||||
# Personal Narrative & Life
|
||||
- memoir
|
||||
- personal-essay
|
||||
- reflection
|
||||
- family
|
||||
- parenting
|
||||
- recovery
|
||||
- mental-health
|
||||
- aging
|
||||
- relationships
|
||||
- childhood
|
||||
- identity
|
||||
```
|
||||
|
||||
Rationale: these are deliberately broad and reusable. `sobriety` was considered and rejected — not expected to be a recurring theme. `recovery` is retained as a broader concept (recovery from any kind of setback, not only substance-related).
|
||||
|
||||
### 2. Prompt rewrite in `request_metadata` (tag-notes.py:127)
|
||||
|
||||
Three changes to the system prompt:
|
||||
|
||||
**a. Add a tagging philosophy sentence** at the top of the `Rules:` section:
|
||||
|
||||
> Tags should describe what the note is substantively about, not topics it merely mentions in passing.
|
||||
|
||||
**b. Allow zero taxonomy tags.** Replace:
|
||||
|
||||
> `tags_from_taxonomy: 1-5 tags drawn from the existing taxonomy that best fit the content.`
|
||||
|
||||
with:
|
||||
|
||||
> `tags_from_taxonomy: 0-5 tags drawn from the existing taxonomy, ONLY when they genuinely fit. Do NOT force a taxonomy tag — return an empty list if nothing truly applies.`
|
||||
|
||||
**c. Loosen new-tag suggestions.** Replace:
|
||||
|
||||
> `new_tag_suggestions: 0-2 NEW tags, only when content truly warrants it (be conservative).`
|
||||
|
||||
with:
|
||||
|
||||
> `new_tag_suggestions: 0-5 NEW tags when the taxonomy doesn't adequately cover the content. Each must be a reusable category (not hyper-specific to one note). Use lowercase-hyphenated style (e.g., personal-essay).`
|
||||
|
||||
### 3. Raise the combined tag cap (tag-notes.py:225)
|
||||
|
||||
Change `[:5]` to `[:8]`:
|
||||
|
||||
```python
|
||||
combined = list(dict.fromkeys(list(taxonomy_tags) + list(new_suggestions)))[:8]
|
||||
```
|
||||
|
||||
Memoir and reflection-style notes often legitimately touch 6-8 distinct themes; capping at 5 was causing otherwise-accurate tags to be dropped.
|
||||
|
||||
## Verification
|
||||
|
||||
After implementation:
|
||||
|
||||
1. Open the "Becoming a Morning Person" note in `~/Documents/ejl-zk/40 Public/41 Notes/` and clear its `tags:` frontmatter field (set to empty list).
|
||||
2. Run `./tag-notes.py`.
|
||||
3. Confirm the new `tags` value includes memoir/personal-essay-style tags (e.g., `memoir`, `personal-essay`, `family`, `recovery`, `aging`, `reflection`) and does NOT include `productivity` or `learning`.
|
||||
4. Spot-check 1-2 other notes that already have reasonable tags — confirm the rewrite didn't regress them. (All LLM-backed fields are only touched when empty, so notes with existing tags won't be reprocessed at all.)
|
||||
|
||||
If the tags still look off, the follow-up is to add a single few-shot example to the system prompt showing a memoir case with zero taxonomy tags and all new suggestions — not included in this change.
|
||||
|
||||
## Files Touched
|
||||
|
||||
- `tag-taxonomy.yaml` — append new cluster
|
||||
- `tag-notes.py` — `request_metadata` system prompt (~20 lines) and the `[:5]` → `[:8]` cap
|
||||
+13
-5
@@ -89,7 +89,7 @@ def parse_json_response(content):
|
||||
return None
|
||||
|
||||
|
||||
def call_llm_json(system_prompt, user_prompt, max_tokens=900):
|
||||
def call_llm_json(system_prompt, user_prompt, max_tokens=4000):
|
||||
payload = {
|
||||
"model": MODEL_NAME,
|
||||
"messages": [
|
||||
@@ -101,14 +101,22 @@ def call_llm_json(system_prompt, user_prompt, max_tokens=900):
|
||||
"response_format": {"type": "text"},
|
||||
}
|
||||
try:
|
||||
response = requests.post(LM_STUDIO_URL, json=payload, timeout=120)
|
||||
response = requests.post(LM_STUDIO_URL, json=payload, timeout=600)
|
||||
if not response.ok:
|
||||
print(f" ! LLM error: {response.status_code} {response.reason}")
|
||||
print(f" body: {response.text[:500]}")
|
||||
return None
|
||||
result = response.json()
|
||||
content = result['choices'][0]['message']['content']
|
||||
return parse_json_response(content)
|
||||
choice = result['choices'][0]
|
||||
content = choice['message'].get('content') or ''
|
||||
parsed = parse_json_response(content)
|
||||
if parsed is None:
|
||||
finish = choice.get('finish_reason')
|
||||
print(f" ! LLM returned no parseable JSON (finish_reason={finish})")
|
||||
if finish == 'length':
|
||||
print(" Hit max_tokens — reasoning model burned the budget before emitting output.")
|
||||
print(f" content: {content[:500]!r}")
|
||||
return parsed
|
||||
except Exception as e:
|
||||
print(f" ! LLM error: {e}")
|
||||
return None
|
||||
@@ -161,7 +169,7 @@ Your previous description was {len(previous_desc)} characters, outside the allow
|
||||
"{previous_desc}"
|
||||
|
||||
Rewrite it to fit strictly within {SEO_DESC_MIN}-{SEO_DESC_MAX} characters."""
|
||||
result = call_llm_json(system_prompt, user_prompt, max_tokens=400)
|
||||
result = call_llm_json(system_prompt, user_prompt)
|
||||
if result:
|
||||
return (result.get('seo_description') or '').strip()
|
||||
return ''
|
||||
|
||||
Reference in New Issue
Block a user