ayberkcansever Tech Radar
Goal: answer "what should I learn next?" for a principal software + AI engineer — a Thoughtworks-style radar, not a trends listicle. Every candidate lands on one of four rings in its quadrant, and the Learn ring is the answer. Every placement must survive: "why this, why now, and why for someone at this level?"
This skill finds topics; sibling learn (~/.cursor/skills/learn/) briefs them. Shared library: briefs are this skill's memory; this skill's Learn ring is that skill's input queue.
The role lens
Apply ruthlessly. Principal software + AI engineer needs:
- Decision leverage — architecture choices, build-vs-buy, platform bets, how you review others' designs. Not "another framework that renders lists".
- AI engineering depth — agents, serving/inference economics, evals, retrieval, routing, AI-adjacent infra. Signal over demos. One strand of the lens, not the scan's default topic.
- Durability — still matters in 2 years. Hot repo, no production adoption → Watch, not Learn.
- Force-multiplier — upgrades how the team works, not personal trivia.
- 101-feasibility — correct 101 in roughly a day via the learn loop. Semester-scale → recommend the entry slice.
Out of scope: junior fundamentals, certification chasing, single-vendor launch fluff, marketing-only evidence.
A scan whose pool is nearly all AI with no AI focus filter is a skew bug. Classify every candidate into a Thoughtworks quadrant:
- Techniques — ways of working (patterns, practices, approaches).
- Platforms — things you build on (clouds, runtimes, databases, inference platforms).
- Tools — software you use rather than build on.
- Languages & Frameworks — languages, major framework releases, SDKs, stdlib.
Rings (learning verbs, not Thoughtworks Adopt/Trial/Assess/Hold):
- Learn — start now. Production-grade; would start
/learnthis week. Selective; past ~8 you are not choosing. - Try — cheap pass is enough (release notes / half-day skim), or second in queue.
- Watch — real signal, not rankable yet. Every Watch blip has a concrete promote when:.
- Skip — deliberately not spending time, plus what would move it back in.
Optional focus ("AI only", "infra only") narrows subjects; level stays principal.
Architecture
Coverage is deterministic first, search second. Edit coverage in feeds.toml, not this file.
feeds.toml registry + [settings]
scripts/scan_feeds.py crawler → reviewable set + drop ledger + feed health
scripts/test_scan_feeds.py drop-rule tests (no network)
scripts/validate_triage.py ledger arithmetic
references/ loaded per step, not at activate
assets/template.html HTML renderer — fill at Step 8
Crawler pre-drops routine noise; agent judges the rest. Novelty (announcing / GA / x.0 / deprecated / open-sourced / preview) and security (CVE, vulnerability, supply-chain) never auto-drop.
Skill root is ~/.cursor/skills/tech-radar/. Agent cwd is usually a project — run scripts with that prefix (the one allowed absolute).
Storage
~/Documents/tech-briefs/radar/YYYY-MM-DD.html
~/Documents/tech-briefs/radar/intake/
~/Documents/tech-briefs/radar/feed_health.json
~/Documents/tech-briefs/radar/watchlist.json
~/Documents/tech-briefs/index.html
mkdir -p ~/Documents/tech-briefs/radar before writing. Same-day re-scan overwrites that date's file; never YYYY-MM-DD-2.html.
Inputs
None required. Optional: focus filter, window override (default 90 days), "include topics I already briefed" to disable dedupe.
Cadence: full scan every 2–4 weeks against the 90-day window. Daily/twice-weekly full scans re-triage the same items — use quick for those.
Mode: full (default) or quick ("quick scan", "update the watchlist") — crawl + triage + watchlist + short chat summary; no confirm, no rings, no HTML. Quick never places blips (would violate the two-signal rule).
Load order
Do not skip a linked file the current step names. Do not read later files early.
- [ ] Read [references/workflow.md](references/workflow.md) — Steps 1–4 always; Step 6 watchlist; Step 7 rings (full only)
- [ ] Step 1–2: memory, then crawl
- [ ] Step 3–4: manual sweep, 100% triage, validate_triage.py
- [ ] Quick → watchlist merge in workflow.md, chat summary, STOP. Do not read gates.md or output.md.
- [ ] Full Step 5: read [references/gates.md](references/gates.md), clear both gates before placing blips
- [ ] Full Step 7–9: rings, then [references/output.md](references/output.md) + [assets/template.html](assets/template.html)
Principles
- Why this, why now, why this level — or cut it.
- Learn/Try momentum = ≥2 independent dated signals. Vendor launch + vendor blog = one signal.
- Deterministic first, search second. Adding a source is a feeds.toml edit.
- Disposition everything, in a file. Crawler
drops.txt+ agenttriage.jsonaccount for the whole crawl. - Whole landscape, not the AI corner. Quadrant balance is a gate. Watch ≠ Learn.
Handoff
Full scan ends with: "Pick a blip number — I'll run the full brief (and lab) on it." On a pick, invoke the learn skill on that topic.