aims

Aims / Objectives Refinement

You are generating and iteratively refining the core research objectives (or specific aims) for a grant proposal. This is the intellectual heart of the proposal — objectives must be ambitious yet feasible, clearly articulated, and aligned with the agency's review criteria.

Arguments

• --proposal-dir <path>: Proposal directory (required)
• --max-rounds <N>: Max refinement rounds (default: 5)
• --adversarial-rounds <N>: Max adversarial review rounds during aims (default: 2)
• --no-multi-model: Skip multi-model adversarial review (external CLIs) even if providers are installed
• --multi-draft: Generate multiple candidate drafts via gw-multi-draft and pick the strongest (if config enables it)

Parse from the user's message.

Procedure

1. Load Context

Read the following inputs from the proposal directory:

• foa_requirements.json — funding opportunity requirements and review criteria
• landscape/competitive_brief.md — what competitors are doing, gaps, opportunities
• landscape/literature.md — if available from a prior literature phase
• config.yaml — agency, mechanism, scoring thresholds, aims.multi_draft flag

Load agency review criteria:

uv run gw-agency review-criteria <agency> <mechanism>

2. Generate Initial Objectives

Draft 3-5 research objectives. Claude (this agent) does the drafting directly based on:

• The research question and PI's expertise
• FOA requirements and scope
• Competitive landscape gaps (what differentiates this proposal)
• Preliminary data (if available)

For each objective, include:

• Clear statement of the aim
• Rationale (why this objective matters)
• Approach (how it will be achieved)
• Expected outcome and deliverable
• Success criteria (measurable)

Multi-draft mode (optional): If --multi-draft is passed OR config.yaml has multi_draft.enabled: true, generate N parallel candidate drafts and then converge. gw-multi-draft is a flat CLI that invokes a user-supplied command template with each variant id. For aims, invoke it like this:

uv run gw-multi-draft \
  --command "claude --print 'Draft aims variant {variant_id} using the context in <proposal_dir> and save to {output}'" \
  --variants v1 v2 v3 \
  --output-dir "<proposal_dir>/state/checkpoints/aims/specific_aims/drafts" \
  --log-dir "<proposal_dir>/logs" \
  --select-best \
  --output-file "<proposal_dir>/logs/aims_multi_draft.json"

gw-multi-draft spawns the N invocations in parallel, writes each draft to its own file, scores them (word count + citation count), and picks the best one. Claude then reads the best draft and synthesizes it into sections/objectives.md.

3. Score Against Agency Criteria

Evaluate the objectives against the agency's review criteria:

For EU agencies (Horizon Europe, ERC, MSCA):

• Excellence: Scientific quality, novelty, ambition, methodology soundness
• Impact: Expected outcomes, dissemination, exploitation potential
• Implementation: Feasibility, work plan realism, resource allocation

For Romanian agencies (UEFISCDI):

• Scientific quality: Novelty, contribution to the field
• Methodology: Appropriateness, rigor
• Feasibility: Timeline, resources, PI capability
• Expected impact: Publications, training, societal impact

Score each criterion on a 1-5 scale. Identify the weakest criterion.

4. Iterative Refinement Loop

For each round (up to --max-rounds):

1. Identify weakness: Which criterion scored lowest? Why?

2. Targeted revision: Revise objectives to strengthen the weakest dimension without degrading others

3. Re-score: Re-evaluate all criteria after revision

4. Adversarial review (up to --adversarial-rounds times):

   Check which provider CLIs are available (mm-detect returns ≥1 installed provider):

   MM_AVAILABLE=0
   if command -v mm-detect >/dev/null 2>&1; then
     if uv run mm-detect --json 2>/dev/null | python3 -c "
   

import json, sys d = json.load(sys.stdin) installed = [k for k, v in d.get('providers', {}).items() if v.get('installed')] sys.exit(0 if installed else 1) "; then MM_AVAILABLE=1 fi fi

**When multi-model is available** (and `--no-multi-model` is not set), run
adversarial critique across 3 external models in parallel:

```bash
cat > /tmp/gw_aims_critique_prompt.txt <<EOF
Critique these specific aims adversarially for a $agency $mechanism proposal.
Attack from three angles:
1. Scientific rigor and novelty
2. Feasibility given timeline and budget
3. Alignment with the agency review criteria (pasted below)

Reference the aims text and the criteria exactly. End with a prioritized
list of concrete revisions (5-7 items) ranked by expected score impact.

Output your critique as markdown. Be harsh — real reviewers are.

--- AIMS ---
$(cat "$proposal_dir/sections/objectives.md")

--- AGENCY CRITERIA ---
$(uv run gw-agency review-criteria "$agency" "$mechanism")
EOF

uv run mm-ask \
  --models gpt-5.4-high,gemini-3.1-pro,grok-4-20-thinking \
  --prompt-file /tmp/gw_aims_critique_prompt.txt \
  --output "$proposal_dir/review/aims_critique.json" \
  --timeout 300 \
  --verbose

Read the 3 critiques from the JSON output. Synthesize by: (a) listing every weakness mentioned by ≥2 reviewers (consensus), (b) listing weaknesses unique to one reviewer (edge catches), (c) ranking proposed revisions by how many reviewers suggested them. Incorporate the prioritized revisions into the next refinement pass.

Fallback (single-Claude path): If no provider CLIs are installed (mm-detect shows zero), perform the adversarial critique directly — role-play a skeptical reviewer from the agency panel, list weaknesses, and revise. Less robust than multi-model consensus but still catches the most common flaws.

5. Human checkpoint: Present the current objectives to the PI with scores per criterion. Ask:
   • Are the objectives aligned with your research vision?
   • Are there objectives missing or ones that should be removed?
   • Is the scope realistic for the proposed timeline and budget?
6. If PI approves, exit loop. If PI requests changes, incorporate and continue.

5. Save Output

Save the refined objectives to sections/objectives.md.

Also update state with the number of rounds used:

uv run gw-state update <proposal_dir> --phase aims --status complete

EU-Specific Notes

• Horizon Europe: Objectives map to Work Packages. Each objective should correspond to a WP with clear tasks, deliverables, and milestones. Structure the objectives document to make this mapping explicit.
• ERC: Objectives are the scientific aims of the PI's research program. They should reflect the PI's personal intellectual vision and long-term research trajectory, not a consortium's shared goals.
• MSCA: Objectives should emphasize the training and career development dimension alongside scientific goals. Two-way knowledge transfer between host and researcher.

Notes

• If the refinement loop stalls (max rounds reached, PI still not satisfied), save the current progress to sections/objectives.md and allow the PI to edit the file manually. The pipeline can resume after manual edits.
• Objectives are the foundation for all subsequent writing — the proposal, risk analysis, budget, and work plan all reference them.

Error Handling

• No provider CLIs installed: Silently fall back to single-Claude adversarial critique. Do not error.
• gw-multi-draft missing: Warn and fall back to single-draft generation.
• Scoring stalemate: If scores do not improve across 2 consecutive rounds, halt the loop and escalate to the human checkpoint with a "stuck" note.
• Max rounds hit: Save current draft and warn the PI; do not loop indefinitely.
• PI rejects all drafts: Capture the PI's rejection reasons as explicit constraints and restart from step 2 with those constraints injected.