anttitevanlinna
UserEvidence-tested knowledge base for the agentic transformation. Answers the CTO question with practitioner evidence, not vendor stories.
Categories
Indexed Skills (18)
research-review
Run the four research review personas (source-type-auditor, zombie-stat-detector, freshness-checker, evidence-ladder-classifier) in parallel against a new or modified research file, synthesis doc, or article draft. Merges findings, dedupes, and presents a unified punch list with severity ordering. Use before publishing any research synthesis or article that depends on research claims.
lecture-ooda
Run a research OODA cycle scoped to ONE lecture or exercise, driven by the standing question in that file's `<!-- backing -->` block. Re-verifies the file's source stamps, hunts the named roster for stance movement, and writes findings to the block's Flagged list for discussion. Use when a lecture's subject may have moved, before a cohort, or when the backing block's `last-run` is stale. Does NOT edit student-facing body prose.
eval-fire
Run a single eval class (writing | story | technical | behavior | pedagogy | strategy | cross_module | slides | voice_panel) against one or more curriculum files. Dispatches a class-judge subagent with the relevant compendiums (filtered by `eval_classes:` frontmatter) and the matching judge prompt template. Returns a structured per-rule verdict. Mirrors `/research-review`'s parallel-launch pattern but scoped to one class per invocation. The full per-file audit lives in `/curriculum-pre-ship-audit` (seven per-file classes + cross_module at module-set scope); this skill is the single-class on-demand fire.
compound
Turn a correction, pattern, or decision into a durable schema-validated memory entry that fires at generation time. Invoke at end of session, or after any significant correction. For content/pedagogy/sales corrections, also applies a one-line amendment to the relevant check_*.md compendium and reports the landed diff.
goalcheck
Iterate on content until it passes eval — driven by CTO conversion, not score-chasing
research
Use when a new user joins the research system, starts their first session, says hello, introduces themselves, asks what this is, or wants to query the knowledge base. Also triggered by /research command.
session-historian
Surface prior investigations from past Claude Code sessions on the current topic. Greps ~/.claude/projects/*/memory/ and recent transcripts for keyword matches. Use when starting work that might repeat a past attempt ("did we already try X?", "have we made this decision before?", "what was the outcome of the Y investigation?"). Prevents re-walking dead ends.
wind-down
Run the end-of-session protocol — enumerate deferred checks, run them, capture corrections via /compound, append taste calls to curriculum/taste-notes.md, then clear sentinels and push. Invoke when Antti signals the session is wrapping ("let's wind down", "we're wrapping up", "ship it", or `/wind-down`). Do NOT invoke mid-session — this is the wrap ritual, not a mid-stream gate.
access-control-analysis
Map the access surface of a feature a software engineer is about to ship. Use when the engineer names a specific feature or PR and wants a structured read of who can reach what, under which trust boundary, through which checkpoint, before a staff engineer or CISO reviews. Read-only analysis, not a refactor.
stride
Run a STRIDE threat-modelling pass against an access-surface map a software engineer has already produced for a feature they're about to ship. Use when the engineer has a surface map in hand and wants threats walked across the six STRIDE categories before picking one to harden against. Read-only analysis, not a patch.
content-creation
Create or revise curriculum content for Agents 102 (modules, exercises, lectures). Use when writing a new module/exercise/lecture, reshaping existing content, creating a new training variant, or reviewing whether content follows project rules. Enforces the three-pass build, the one-exercise-per-module principle, include-link conventions, and content-strategy ↔ module file alignment.
mermaid-diagrams
Render Mermaid .mmd diagram sources to committed .svg for curriculum docs. Use when creating or updating a diagram embedded in a module/supplementary/reference page, or when a .mmd source changed and its .svg must be regenerated. Produces vector (crisp at any width, tiny), not raster.
self-study
Facilitator for a student doing Agents 101 alone — no in-person trainer. Invoke on first session to set up the working directory and orient the student; invoke subsequent sessions with "continue" to pick up where they left off. Runs as the Teacher Claude alongside a separate Builder Claude where exercises execute. Manages progress tracking, the 4 Cs cadence per module, and Builder working-directory checks. Do NOT invoke for curriculum authoring work — that belongs in the agents-102 repo.
agentic-nerd
Optional running companion for a software engineer doing Agents 102 Agentic Engineering 101 alone — no in-person trainer. Invoke on first self-study session to orient; invoke later sessions with "continue" to pick up where they left off. Runs as the Teacher Claude alongside a separate Builder Claude in the student's real repo where exercises execute. Manages progress, the 4 Cs cadence per module, and the module-by-module push-backs the in-room trainer would deliver. Classroom/cohort delivery must not depend on or expose this skill. Do NOT invoke for curriculum authoring — that belongs in the agents-102 repo.
agent-security
Run agent-adapted STRIDE + access-control analysis on an agent system. Produces a ranked risk list with agentic mitigations (scope / split / filter / gate / review).
security-tools
Quick local-environment verification for security analysis work. Runs the bundled `check.sh` script and reports whether the shell environment has the basics the security skills depend on. Invoke at the start of a security-focused session as a pre-flight.
company-ai-policy
Audit an agent system against Kaleva Retail Group's AI and data-usage policies. Produces a row-per-rule compliance report with verdicts (compliant / violating / "I can't tell") and evidence from the target system's files.
grill-me
Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.