Applied agentic-AI concepts

Career context file

Developers already understand CLAUDE.md and AGENTS.md: put a context file in a project and an agent can understand it before editing. AgentKit SEO applies that pattern to a career. The agent-context-file is a private Markdown source of truth that stores verified identity facts, roles, projects, links, metrics, and constraints.

As of 1.8.0 it also captures Goals and targeting: the person's ideal role, current focus, what they want to work on next, target locations (or No restriction), interests, and constraints. These are recorded as stated intent, kept strictly separate from verified facts, so downstream skills can aim output toward a goal without inventing experience to support it.

LLM Wiki

Without a knowledge layer, agents guess at platform constraints from training data: ATS parser behavior, LinkedIn field limits, GitHub Linguist rules, and other details that change or depend on context. That guessing produces confident but wrong advice.

AgentKit SEO's wiki layer follows Andrej Karpathy's LLM Wiki concept: a maintainer agent compiles the knowledge from sources and keeps it current, and runtime agents read it instead of re-deriving platform facts on every query. The project deliberately adapts his personal-second-brain idea into a shipped, versioned knowledge pack. Every skill ships per-module entries for canonical definitions, platform constraints with confidence labels, known failure modes, evidence rules, and audit output rules; these load conditionally so an agent pulls deeper context only when the current task needs it. A maintainer-only wiki refresh skill exists in the source tree for local source audits; the installed user bundle ships only the runtime skills.

Role-grounded persona and self-review

Each user-facing skill reasons as the relevant professional: hiring manager and open-source maintainer for GitHub, technical recruiter for LinkedIn, ATS screener for CV/ATS, SEO specialist for the portfolio, audience editor for X, and biographer plus fact-checker for the context file. The persona focuses the agent's perspective on what that reader actually cares about.

Before returning, every skill runs a self-review pass. It checks the draft for fabricated facts, correct evidence labels, scope and goal alignment, and impact ordering, and it flags any failure it cannot fix rather than hiding it.

Audit scorecard

The GitHub, LinkedIn, CV/ATS, and web-portfolio modules include a weighted 0-100 triage scorecard with bands and a prioritized fix-first order. It exists to decide what to fix first.

It is explicitly an internal prioritization heuristic, not a platform ranking and not a guaranteed score. A high band does not promise recruiter attention, search position, or any platform outcome; it only orders the work so the highest-impact fixes come first.

Progressive disclosure and token discipline

The orchestrator routes a request to a single module by default and asks for the smallest missing input set rather than demanding every asset upfront. An agent loads one module, then only the references and wiki entries that the current task needs. The result is token discipline: context budget is spent on the surface being worked on.

Markdown knowledge graph

The modules connect as cross-referenced Markdown files with one entrypoint and explicit edges. A broad question enters at the root self-description and narrows to a single module and a single constraint, instead of loading the whole system. See the orchestration playbook for how routing works.

Evidence and confidence labels

Every skill shares one vocabulary for how much support a claim has. Claims labeled Inference, Inaccessible, or Needs evidence are not turned into confident public copy.

Verified

Observed directly in inspected public material, local files, rendered output, supplied screenshots, extracted text, or a supplied source-of-truth file.

From context

Taken from the user's agent-context-file.

From source

Taken from supplied source material such as pasted text, exports, screenshots, public URLs, local files, or job descriptions.

Inference

Reasoned from inspected evidence, but not directly observed as a fact.

Inaccessible

Could not be inspected because the surface is private, login-gated, unavailable, blocked, or outside the task scope.

Needs evidence

A claim that should not be reused publicly until the user supplies support or it is verified in source material.

One source, many adapters

AgentKit SEO keeps one portable canonical source and generates each provider's required layout from it, so the same methodology installs into Claude Code, Codex, Gemini CLI, Antigravity, OpenCode, and portable skill folders without a second copy drifting out of sync. See the providers page for the install targets.

AI-answer-engine readiness (GEO/AEO)

Hiring discovery increasingly runs through AI answer engines, so the same properties these modules enforce, consistent facts across surfaces and verifiable proof, are what those systems can quote accurately. The project treats generative and answer engine optimization (GEO/AEO) as an evolving practice with no guaranteed ranking outcome, not a promise.

In particular, llms.txt is treated as an emerging, low-confidence convention, not a ranking lever. It can help a tool find a clean project map, but no search engine or assistant guarantees that publishing one changes visibility or citations.

Platform names, trademarks, and logos belong to their respective owners. AgentKit SEO is independent and is not affiliated with, sponsored by, or endorsed by those platforms.