Skill/tool usage drops off as a project matures — lc/astra skills stop being auto-invoked later in a project

[calumhrmurray]

What happened

In the early stages of a project the lc/astra skills and tools are invoked
consistently and effectively. As the project matures, Claude increasingly stops
reaching for them on its own — work that should route through a skill gets done
ad-hoc instead.

Concrete example from a live session: the user asked Claude to review a paper,
and Claude did the work without invoking the relevant ASTRA skill
(paper-extraction / the astra paper workflow). It hand-rolled the steps
(WebFetch the abstract, manual quote extraction, astra paper add --pdf) rather
than going through the skill path. The end result was fine, but the skill that
exists for exactly this was bypassed.

Impact

The skills encode the correct, provenance-tracked workflow. Skipping them later in
a project means inconsistent process (manual vs skill-driven), and the value of
having the skills degrades precisely when the project is most complex and most
needs the guardrails.

Possible direction (not a prescription)

• Skill-selection seems to weaken as session/project context grows — worth checking
  whether long-context or accumulated CLAUDE.md state is crowding out the
  skill-trigger reminders.
• A periodic "did a skill match this request?" check, or stronger trigger phrasing
  for paper/spec/recipe operations later in a project, might help.

Environment

• ASTRA: 0.2.9
• lightcone-cli: 0.3.6
• Python: 3.12.5
• OS: Linux 4.18.0-553 (EL8)

This was written by Claude.

[calumhrmurray]

Folding in #151 as a concrete instance of this same root cause (skills not auto-invoked in a mature project).

Instance: in an existing ASTRA project, the user asked Claude to "read some papers to check we are not missing anything." Claude ran ad-hoc web search instead of routing through the ASTRA literature flow, and wrote nothing back into the spec.

Specific expected routing this adds (not in the original report): a literature-review request in an existing project should mirror lc-new behavior —

• invoke the astra skill for the prior_insights/evidence schema,
• use the lc-extractor agent to pull verified quotes from the papers,
• land them as structured prior_insights in astra.yaml, each attached to the decision it bears on.

Note the project's astra.yaml had no prior_insights section at all, so there was no existing literature layer to extend — the ad-hoc path produced unverified paraphrases (one WebFetch even failed to extract the real content) with no provenance.

Closing #151 as a duplicate of this umbrella issue.

This comment was written by Claude on behalf of the user.