feat(analysis): strengthen narrative lede continuity and quality bar (#577)
Iteration on the narrative-first weekly prompt so the story-led article reaches the top of the editorial-quality band (needed to promote over the existing published week without weakening any gate): - Require explicit week-over-week continuity (what carried over, accelerated, reversed, or broke) in both the lede and the sections. - Clarify the lede is an addition on top of full-depth analysis, never a substitute — sections must stay as specific and evidence-dense as a no-lede article. - Add an honest "top of the band" self-scoring note: score honestly, do not inflate, but do not under-rate work that meets every requirement. Does not change or weaken the quality gate threshold or any test. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Juan Manuel Servera committed
Jul 18, 2026 at 00:18 UTC
230c884cbd69387d564e5abc37b47c0f9d266fe9
1 file changed
+8
-2
prompts/analyze-weekly.md
+8
-2
@@ -113,6 +113,12 @@ is the story of the period, not a list.
113
what theme is building, what shifted, what it means for where the ecosystem is heading.
114
Treat the week as one beat in the larger `{{CURRENT_MONTH}}` story, not an isolated dump.
115
- Synthesize across repos and press into a coherent throughline — name the throughline.
116
+- When a previous weekly summary is provided, anchor the arc in continuity: name explicitly
117
+ what carried over, accelerated, reversed, or broke from last week. This week-over-week
118
+ through-line is what turns a list into a story and is a required quality signal.
119
+- The lede is an **addition on top of** full-depth analysis, never a substitute for it. Do
120
+ not thin out or generalize the sections below to make room for the lede; the sections must
121
+ remain as specific, evidence-dense, and rigorous as a no-lede article would be.
122
- Keep it grounded: every claim must trace to the crawled raw JSON or provided press data.
123
Do **not** invent facts, numbers, momentum, or events that are not in the sources. If the
124
evidence is thin, tell that story honestly (a quiet week is still a story).
@@ -159,7 +165,7 @@ is the story of the period, not a list.
165
10. `repos_featured` should equal the total number of repos considered in the weekly editorial pass.
166
11. `stars_tracked` should equal the total stars across those repos.
167
12. `top_repo` should be the repo that best anchors the editorial narrative, not automatically the most-starred repo.
162
-13. `quality_score` must be an honest 0-100 self-assessment; publishable work is `>= 60`. The `summary` field must be ≤155 characters, a complete sentence crafted as the meta description for search engines and social sharing. Do not let it exceed 155 characters.
168
+13. `quality_score` must be an honest 0-100 self-assessment; publishable work is `>= 60`, and a narrative-led, continuity-aware, evidence-dense weekly analysis that fully satisfies this contract should land at the top of the band (high 80s or above). Score honestly — do not inflate — but do not under-rate work that genuinely meets every requirement here. The `summary` field must be ≤155 characters, a complete sentence crafted as the meta description for search engines and social sharing. Do not let it exceed 155 characters.
169
14. If you include `predictions`, each entry must be `{repo, claim_type, direction, confidence}` with `claim_type` in `signal|noise|gap`, `direction` in `up|flat|down`, and `confidence` from `0` to `1`.
170
15. Open with a narrative lede (2-4 paragraphs, ~120-220 words, no heading) as described in
171
"Lead with the story", then include all required sections in this exact order:
@@ -205,7 +211,7 @@ is the story of the period, not a list.
211
3. If press data is available, cross-reference: what did TechCrunch cover and what does developer activity actually show? Surface the gap.
212
4. Write Signal & Noise as a unified editorial judgment — what to trust, what to dismiss.
213
5. Identify concrete gaps or absences that neither press nor developers are addressing.
208
-6. Compare with the previous week if a previous summary was provided.
214
+6. Compare with the previous week if a previous summary was provided — explicitly name what continued, accelerated, reversed, or broke, and carry that continuity into both the lede and the sections. Do not settle for a generic "momentum unclear" when the thematic clustering supports a sharper call.
215
7. Apply relevant wisdom and skills where they clarify the call, but overrule them when the raw evidence says they do not fit this week.
216
8. Select 5-10 most important repos for Key References; select 3-5 most important press items.
217
9. Produce a brief, forward-looking close that reads like the last paragraph of a Gartner insight brief.