SKILL.md
Tech Sensemaking
Analyze a technology announcement through 4 strategic questions using Verbalized Sampling to produce diverse, non-obvious insights grounded in the user's specific context — whether that's a business, a product, a feature branch, a codebase, a personal project, or anything else.
Required Input
- Announcement: pasted text, URL, or description of the technology change.
- Context subject: what to analyze the announcement against (asked in Phase 2 if not obvious).
Workflow
Phase 1 — Intake
- Accept the announcement. If a URL, scrape it (Firecrawl, WebFetch, or browser tools).
- Write a neutral announcement summary (200-400 words):
- What was announced (capabilities, features, changes) - What constraints or limitations were mentioned - Availability and timeline - What was NOT said (notable omissions)
- Present the summary to the user. Ask: "Does this capture the announcement accurately, or should I adjust anything before analysis?"
Phase 2 — Context Loading
Determine the context type and load accordingly. If the user hasn't specified what to analyze the announcement against, ask:
"What should I analyze this announcement against? For example:
- A business (prodmgmt.world, your startup, etc.)
- A product or feature (your SaaS, an open-source project, etc.)
- A codebase or feature branch (a repo you're building)
- A personal goal or project
- Something else?"
Context sources by type
| Context type | Where to look | What to extract |
|---|---|---|
| Business | Vault notes (qmd search), Context/ files if they exist, user description |
What it does, revenue model, competitive position, goals, constraints, team size |
| Product | README, product docs, vault notes, user description | What it does, target users, current capabilities, roadmap, tech stack |
| Codebase / feature branch | Source code, README, CLAUDE.md, recent commits/PRs | Architecture, dependencies, current problems, what's being built |
| Personal project | Vault notes, user description | Goals, constraints, timeline, what's been tried |
| Role / career | Vault notes, user description | Current role, skills, goals, industry, constraints |
| Generic / exploratory | User description, web research | Domain, key players, known constraints, relevant trends |
Compress into a context block (~300 words) covering:
- What the subject is and what it does
- Current state, capabilities, or position
- Goals or direction
- Key constraints (time, resources, technical, strategic)
- What's already been tried or is in progress (if known)
This block is injected into every VS prompt. Rich context is the primary defense against generic outputs — see the /verbalized-sampling skill's FM-2 failure mode. If context is thin, say so and ask the user to enrich it before proceeding.
Phase 3 — VS Generation
This phase uses the /verbalized-sampling skill for generation mechanics.
Read references/vs-prompts.md for the 4 prompt templates. Substitute {announcementsummary} and {contextblock} into each template.
VS configuration for this skill:
- Variant: VS-CoT (each output includes a
reasoningfield) - Distribution: tail sampling, p < 0.10
- k = 5 per question (20 total outputs)
- FM-1 mitigation: exclusion constraints baked into each prompt template
- FM-2 mitigation: rich context block injected into every prompt
- FM-3 mitigation: 4 separate problem frames + coverage requirements within each
The 4 questions:
| # | Question | Key | What it surfaces |
|---|---|---|---|
| 1 | New Outcomes | new_outcomes |
What's now possible — including second/third-order effects |
| 2 | New Affordances | new_affordances |
What actions can now be taken — concrete, effort-rated |
| 3 | Relative Value | relative_value |
Which possibilities matter most for THIS context vs. alternatives |
| 4 | Secret Levers | secret_levers |
Non-obvious causal chains for durable advantage |
Phase 4 — Critique
Apply the critique-and-improve loop from the /verbalized-sampling skill (see its references/critique-framework.md) to all 20 outputs.
For each item, check the 6 dimensions:
- Naive? Would this appear in a tech journalist's hot take?
- Under-specified? Could execution start this week without further research?
- Magical thinking? Does this assume steps will "just work"?
- Ignores constraints? Does this violate real constraints from the context block?
- Base rate blind? Is this approach known to fail in similar contexts?
- Over-engineered? Is this a cannon to kill a fly given the subject's stage/scale?
Triage:
failure_count 0-1: Keep as-isfailure_count 2-3: Generate 2 improved variants, pick the bestfailure_count 4+: Drop entirely
Phase 5 — Present
Output in readable format, grouped by question. Within each question, order by probability ascending (most novel first).
Format:
## Q1: New Outcomes
### High diversity (p < 0.05)
1. [text] (p=0.03, time_horizon: medium, order: second)
**Reasoning:** [condensed reasoning]
### Moderate diversity (p 0.05–0.10)
2. [text] (p=0.07, ...)
...After all 4 questions, add:
## Synthesis: Top 3 Moves
[Identify the 3 highest-leverage actions across all 4 questions. For each:]
1. **Action**: What to do
2. **Why now**: Why this announcement creates the window
3. **First step**: The concrete next actionEnd with:
## What I excluded (and why)
[List 2-3 items that were dropped by the critique pass, with the failure reasons.
This shows the user what "obvious" looks like so they calibrate their own thinking.]Usage Notes
- The skill works best when the announcement is fresh (within days). Stale announcements produce outputs closer to what's already been discussed publicly.
- If context is thin, the outputs will drift generic. Enrich context before running on high-stakes announcements.
- For major platform shifts (not just feature announcements), consider running this skill twice: once on the announcement itself, once on the second-order market reactions 1-2 weeks later.
- The "Secret Levers" question (Q4) is the highest-value output. If time-constrained, run Q4 alone.
- This skill delegates VS mechanics to
/verbalized-sampling. Consult that skill for distribution thresholds, failure modes, output formatting, and the meta-prompt for generating custom VS prompts.