agent-wiki-compare-out…
Compare successful and failed normalized agent trajectories to derive evidence-backed agent-wiki guidelines. Use when Codex has…
Consult an agent-wiki for guidelines relevant to the task at hand. The wiki itself documents how to retrieve from it (AGENTS.md). Use this skill once you know what task or sub-task you're about to do — not at session start.
Claude Code
$ npx -y skills add AgentToolkit/altk-evolve --skill agent-wiki-consult --agent claude-code
Codex
$ npx skills add AgentToolkit/altk-evolve --skill agent-wiki-consult -a codex -y
Cursor
$ npx skills add AgentToolkit/altk-evolve --skill agent-wiki-consult -a cursor -y
Gemini CLI
$ npx skills add AgentToolkit/altk-evolve --skill agent-wiki-consult -a gemini-cli -y
GitHub Copilot
$ npx skills add AgentToolkit/altk-evolve --skill agent-wiki-consult -a github-copilot -y
OpenCode
$ npx skills add AgentToolkit/altk-evolve --skill agent-wiki-consult -a opencode -y
Windsurf
$ npx skills add AgentToolkit/altk-evolve --skill agent-wiki-consult -a windsurf -y
name: agent-wiki-consult description: Consult an agent-wiki for guidelines relevant to the task at hand. The wiki itself documents how to retrieve from it (AGENTS.md). Use this skill once you know what task or sub-task you're about to do — not at session start.
This skill is a thin wrapper around the wiki's own `AGENTS.md` document. The wiki contains evidence-grounded guidelines distilled from agent trajectories; `AGENTS.md` is its agent-readable contract for navigation and retrieval. This skill tells you to:
1. Find the wiki root. 2. Read `<wiki-root>/AGENTS.md`. 3. Follow the recipe described there against the user's current task.
The retrieval logic lives in `AGENTS.md`, not in this skill. That separation is intentional: when the wiki's structure or recall heuristics change, edit `AGENTS.md` and not this skill.
Call this skill **once you know the task or sub-task you're about to do**. Concretely:
block of work.
documented.
"now I need to handle browser auth resumption").
Do **not** invoke at session start (no task to filter against), and do not invoke for trivial tasks (typo fix, single-line edit) where the wiki's overhead exceeds the work.
If the user passed a path argument (e.g. `wiki-twobatch`, `wiki-twobatch-skills`, or any other path), use it.
Otherwise auto-detect: walk up from the current working directory looking for any sibling directory matching `wiki-*` that contains an `_config.yaml` file. If multiple are found, prefer the one closest to cwd. If none are found, ask the user which wiki to consult.
Read <wiki-root>/AGENTS.md
This is the contract document. It explains the wiki's structure, the filename suffix convention, the `_index.jsonl` schema, and a recommended retrieval recipe. Read it in full — it's typically 3–5 KB.
If the file does not exist, run:
uv run python explorations/agent-wiki/skills/scripts/build_agent_wiki.py \ --wiki-root <wiki-root> catalog
Catalog's bootstrap phase will copy the bundled template into place.
Read <wiki-root>/_index.jsonl
One row per guideline / cluster / task / subtask. Schema documented in `AGENTS.md`. Rows are sorted clusters-first.
Per the recipe in `AGENTS.md` (which is advisory, not mandatory):
1. Identify topical tags + keywords from the task description. 2. Filter `_index.jsonl` rows by tag overlap or `trigger:` substring match. 3. Prefer cluster rows when a cluster and its members both match. 4. Read the top 2–5 matching pages in full. 5. State which guidelines apply (briefly) before acting on them.
Use your judgment for scoring. The wiki does not prescribe a fixed algorithm; trust the heuristics in AGENTS.md and the row content.
Report 2–5 candidate matches to the user (or to your own next-step reasoning) with:
This skill accepts:
`wiki-twobatch-skills`, or an absolute path.
context.
1. **Don't pre-load.** This skill is on-demand by design. Calling it at session start without a specific task wastes tokens and produces noise. 2. **Read AGENTS.md every time.** Wikis evolve; the contract may have changed. Caching the contract across invocations is fragile. 3. **Read clusters before atomics.** Cluster pages reference their members — you usually don't need to read the members directly. 4. **Cite when you act on a guideline.** Mention the guideline's title + link in your response so the user can audit the recommendation. 5. **Don't follow guidelines blindly.** If a guideline's `trigger:` doesn't quite match your situation, say so — note the close match and choose your own course.
Evolve helps AI agents learn from experience. It turns past trajectories into reusable guidelines that help agents complete future tasks, without changing model weights.
Compare successful and failed normalized agent trajectories to derive evidence-backed agent-wiki guidelines. Use when Codex has…
Read all atomic guidelines in wiki-twobatch/guidelines/ and propose themed clusters that group near-duplicates. Writes cluster…
Read a normalized Claude Code trajectory JSON and extract reusable guidelines into wiki-twobatch/guidelines/. Use when mining…
Ingest one or more agent trajectories (raw bob/claude traces or normalized JSON) into an agent-wiki end-to-end — convert,…
Read a normalized Claude Code trajectory JSON and write an episodic summary page to wiki-twobatch/summaries/. Use when…
Read a normalized Claude Code trajectory JSON and produce a wiki-resident SKILL.md page that future agents can invoke. Use when a…
You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or…
Ask which skill or flow fits your situation. A router over the skills in this repo.