tq-forge-improve

/tq-forge-improve — targeted rewrite of low-scoring sections

When to use

A forged skill or agent scored <7 and you want to fix it without rewriting the whole artifact. Give it a slug (e.g. notion-daily-summary) or an absolute path; the skill identifies which dimensions are pulling the score down, rewrites only those sections, then verifies.

Procedure

export TQ_FORGE_HOME="${TQ_FORGE_HOME:-$HOME/.tq-forge}"
S="${CLAUDE_PLUGIN_ROOT:-${TQ_FORGE_HOME:-$HOME/.tq-forge}/install}/scripts"
source "$S/common.sh" && tq_ensure_home

1. Assess the artifact.

   bash "$S/improve-loop.sh" assess "<slug>"
   

   This prints per-dimension scores, highlights weak dims (<7), and gives a coaching hint for each. Capture which sections need work.

2. Read the current content. Read the SKILL.md (or AGENT.md + system-prompt.md) at the path assess printed. Understand what each weak section currently says.

3. Rewrite only the weak sections:

   Dimension	Section to rewrite
   clarity	## When to use — <=3 sentences, concrete trigger
   actionability	## Procedure — add copy-pasteable commands
   completeness	Add missing section(s): Pitfalls / Verification / Tags
   specificity	Replace vague language with real paths / tool names
   consistency	Define the output-format schema once, up top (agents)
   handoff_clarity	hand_off_to in tools.json — add a when: per entry

   Use Edit for targeted swaps; use Write only if the file is malformed. Preserve every section you are NOT rewriting, verbatim.

4. Verify the improvement.

   bash "$S/improve-loop.sh" verify "<slug>"
   

   This re-scores, prints before/after, and updates skill-log.json.

5. Branch on the result.

   • avg >=7: print ✅ <slug> now scores X/10. Ready to promote. Suggest /tq-forge-promote <slug> if it's still in the sandbox.
   • avg <7: print the remaining weak dims and flag it:

     bash "$S/forge-checkpoint.sh" review "<slug>" "<weak-dim>" "<score>"
     

Pitfalls

• Only rewrite weak sections. Editing a strong section risks dropping its score while raising the weak one. Target precisely.
• assess auto-detects sandbox and production paths. If the slug exists in both, it prefers the sandbox copy — improve there, then re-promote.
• Agents have 6 scoring dimensions vs 4 for skills; handoff_clarity and consistency apply only to agents.
• Adding copy-pasteable bash commands often raises actionability 2-3 points on its own — try that before broader rewrites.

Verification

• improve-loop.sh verify <slug> reports passed: true in --json mode.
• skill-log.json shows the updated score and a last_tested timestamp.
• The before score < 7 and the after score >= 7 (delta printed).

Tags

tq-forge improve quality rewrite single-slug