All skills

Assess documentation for gaps, staleness, quality, and files outside the defined doc set. Preserve useful knowledge in canonical docs, then remove superseded or unnecessary files. Check changed code when the tree is dirty, otherwise assess the whole repo. Explicit modes: --review, --update, --generate, or --session to capture durable knowledge from a conversation or transcript. Use for doc creation, cleanup, freshness, quality, or saving occasional manual test procedures and results whose numbers a later run will compare against; routine suite runs and ordinary QA passes do not need test records.

Use this Skill: https://skilld.dev/gh/nielsmadan/agentic-coding/doc

This session only. Nothing lands on disk.

referencesmode-session.md

≈871 tokens on demand. Your agent reads this file only when SKILL.md points to it.

Session Mode (--session)

Invocation: doc --session [--md <conversation.md>] [--report <file>]

Integrate the durable knowledge from one work session into the project's docs. Without --md, use the current conversation and its available tool results. When supplied, --md selects a rendered conversation (plain markdown of USER/ASSISTANT turns); do not silently mix in the current session's environment or results. You do NOT read Claude Code .jsonl files here. Capture does not depend on a code diff or rerun completed tests.

Extract and place five kinds of knowledge, using the SAME layout and conventions as --update/--generate (see references/principles.md):

  1. Gotchas — pitfalls, surprises, "things we ran into" → the ## Gotchas section of the most relevant module/feature doc.
  2. Decision rationale — why a product/tech decision was made → the ## Why section.
  3. Behavior changes — what the session actually changed → refresh the affected ## How it works (current state) sections.
  4. External findings — what we learned about a dependency, platform, harness or external API by probing it: a behavior that contradicts its docs, a version-specific quirk, a limit we hit → docs/reference/<subject>.md. Sessions are the main source of these, and the reason the bucket exists: the finding cost an experiment here and will otherwise be re-established from scratch next time. Carry the date and the version probed across from the session, plus how it was probed; if the session only read upstream's docs, that is a documented claim with a link, not a verified one. Correct-usage conventions for one library go to library-docs / library-use instead.
  5. Occasional manual tests — performance measurements, restore drills, compatibility experiments, and other operations whose results a later run will want to compare against → docs/tests/<name>/. Read manual-tests.md and preserve both the reusable procedure and a dated run record. Most testing does not qualify: routine suite/CI runs do not become records because someone launched them by hand, and an ordinary feature QA or exploratory pass belongs in the conversation, not in a file. Keep missing historical details explicitly unrecorded rather than reconstructing them from today's checkout.

Rules

  • Find the right target docs via docs/overview.md and the existing tree. Prefer augmenting existing docs over creating new ones.
  • Never touch docs/product/ (owned by review-product).
  • Only record durable, project-specific knowledge — skip transient chatter and anything already documented. It is fine to conclude there is nothing worth adding.
  • Preview the proposed edits to the user and apply on confirmation. Leave changes uncommitted (the user commits).
  • Fallback: if the repo has no docs/, qualifying manual tests still go directly into docs/tests/<name>/. For other knowledge without a suitable home, write a single docs/session-harvests/<name>.md; derive the name from the transcript filename or the current session's topic. Link test records rather than duplicating them in the harvest.
  • Report: if --report <file> is given, write the list of doc files you created or edited to it, one path per line (repo-relative). If nothing was integrated, write an empty file. This is how the caller records what was harvested.

Source: SKILL.md on GitHub

1 alert6mo3 checks · Risk SAFE
  • Gen Agent Trust Hub6mo

    The skill manages documentation review and generation by processing project files. While the core functionality is safe, the skill is susceptible to indirect prompt injection because it reads and analyzes untrusted documentation and project instructions (CLAUDE.md), which could contain malicious directions. No evidence of data exfiltration or remote code execution was found.

  • Socket6mo

    No alerts

  • Snyk6mo

    Risk: HIGH · 1 issue

Signed by skilld at 6e73347. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated 2 weeks ago
effort
high
Other metadata
argument-hint
[ (no args = context-aware assess) | --review | --update | --generate <target> | --session [--md <file>]] [--all | --staged | --unpushed]

README badge

README badge for nielsmadan/agentic-coding/doc